Mutltimodal MRI Brain Tumor Segmentation using 3D Attention UNet with Dense Encoder Blocks and Residual Decoder Blocks

Mutltimodal MRI Brain Tumor Segmentation using 3D Attention UNet with Dense Encoder Blocks and Residual Decoder Blocks | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Method Article Mutltimodal MRI Brain Tumor Segmentation using 3D Attention UNet with Dense Encoder Blocks and Residual Decoder Blocks Tewodros Megabiaw Tassew, Betelihem Asfaw Ashamo, Xuan Nie This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-2717573/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Medical image segmentation is essential for disease diagnosis and for support- ing medical decision systems. Automatic segmentation of brain tumors from Magnetic Resonance Imaging (MRI) is crucial for treatment planning and timely diagnosis. Due to the enormous amount of data that MRI provides as well as the variability in the location and size of the tumor, automatic seg- mentation is a difficult process. Consequently, a current outstanding problem in the field of deep learning-based medical image analysis is the development of an accurate and trustworthy way to separate the tumorous region from healthy tissues. In this paper, we propose a novel 3D Attention U-Net with dense encoder blocks and residual decoder blocks, which combines the bene- fits of both DenseNet and ResNet. Dense blocks with transition layers help to strengthen feature propagation, reduce vanishing gradient, and increase the receptive field. Because each layer receives feature maps from all previous layers, the network can be made thinner and more compact. To make predic- tions, it considers both low-level and high-level features at the same time. In addition, shortcut connections between the residual network are used to pre- serve low-level features at each level. As part of the proposed architecture, skip connections between dense and residual blocks are utilized along with an attention layer to speed up the training process. The proposed architecture was trained and validated using BraTS 2020 dataset, it showed promising results with dice scores of 0.866, 0.889, and 0.828 for the tumor core (TC), whole tumor (WT), and enhancing tumor (ET), respectively. In compar- ison to the original 3D U-Net, our approach performs better. According to the findings of our experiment, our approach is a competitive automatic brain tumor segmentation method when compared to some state-of-the-art techniques. Artificial Intelligence and Machine Learning Nuclear Medicine & Medical Imaging Brain tumor segmentation MRI images Attention U-Net Dense block Residual block Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1. Introduction Brain tumors are a serious medical condition that can have devastating effects on the nervous system and surrounding healthy brain tissue. They are divided into two main categories: primary and metastatic. Primary brain tumors originate from the cells of the brain or its immediate vicinity, while metastatic ones form elsewhere in the body then spread to the brain through circulation. Primary tumors can be classified further as glial, composed of glial cells, or non-glial, formed on or in other structures such as nerves, blood vessels, and glands; these may be benign (non-cancerous)or malignant (can- cerous). Treatment for both types varies depending on how advanced they are but typically involves surgery followed by radiation therapy or chemother- apy depending upon individual circumstances. In the United States alone, 700,000 people had primary brain tumors, with over 88,000 diagnosed by 2022. According to the World Health Organization, cancer causes 9 million deaths globally each year, accounting for 22 percent of all chronic diseases and ranking second only to cardiovascular disease, which causes 17.9 million deaths each year. Primary malignant brain tumors, for example, account for 1.4 percent of all new cancer diagnoses in the United States each year. Primary malignant brain tumors have a global mortality rate of 2.0/100,000 in women and 2.8/100,000 in men (Ellison, 2020). Medical imaging is an essential tool in the diagnosis and treatment of tumors. It allows for accurate tumor segmentation, assessment of tumor size and shape, and evaluation of response to treatments such as radiation or chemotherapy. The most commonly used medical imaging techniques are computed tomography (CT) scans, ultrasound examinations, X-ray exami- nations, and magnetic resonance imaging (MRI). These technologies allow healthcare professionals to observe the structure of tissues within a patient’s body in order to accurately diagnose tumors or assess how well a treatment is working. Medical imaging plays an important role in providing early de- tection of cancerous lesions which can lead to better prognosis outcomes for patients with tumors (Raza et al., 2023). In clinical medicine, MRI has be- come the most widely used imaging examination for brain tumor diagnosis, because of its sharpness and high tissue resolution, ability to produce images of the brain’s tissues in clear contrast, and versatility in allowing different parameters to be set in order to obtain specific anatomical information (Raza et al., 2023; Alagarsamy et al., 2020). It is a safe and advanced imaging pro- cedure that makes use of nuclear magnetic resonance to produce a reliable depiction of internal body structure and tissues without causing harm to the organ being imaged or causing high levels of ionization or radiation effect. It creates an accurate portrayal of internal body structure and tissues by combining strong magnetic waves ranging from 1.5 T to 3 T with radiofre- quency waves (Azhari et al., 2014; Tanneedi et al., 2017). The use of MRI for glioma diagnoses allows physicians access to detailed information about the cells within a patient’s brain by offering up sub-regions such as peritumoral edema, necrotic core, enhancing tumor core, and non-enhancing tumor core depending on factors such as degree of invasion or prognosis. This technology also helps healthcare professionals determine if any additional treatments are necessary after the initial diagnosis has been made since it can provide fur- ther insight into any changes that may have occurred over time regarding the tumor’s size or shape prior to beginning treatment (Alagarsamy et al., 2020; Crimi et al., 2019; Ghaffari et al., 2020). Multimodal MRI scans are a powerful tool for diagnosing and monitoring diseases of the brain. By combining four different modalities, T1 weighted (T1), T1 weighted with post-contrast (T1Gd), T2 weighted (T2), and Fluid Attenuated Inversion Recovery (FLAIR) scans, it is possible to gain an ac- curate representation of the various histological sub-regions within the brain (Ghaffari et al., 2020; Zhang et al., 2011). The purpose of each modality can be broadly understood based on their distinct intensity distributions (Zhang et al., 2020). For example, T1w is mainly used to measure healthy tissues. While T2w has a bright tumor region visible on scan results, the FLAIR scan helps distinguish edema from cerebrospinal fluid (CSF) (Bauer et al., 2013). Meanwhile, both the whole tumor area as well as its core can be seen clearly when viewing axial slices through both a FLAIR and T2 scan, respectively. Additionally, a post-contrast enhanced view via a T1Gd allows for the vi- sualization of enhancing tumors. This combination provides clinicians with valuable insight into disease states that would otherwise remain undetected or overlooked. An example demonstrating these four MRI modalities is shown in Figure 1, where ground truth segmentation was applied to patient data from the BraTS2020 training dataset. This demonstrates how multimodal MRIs pro- vide invaluable information regarding tissue structure, which helps build an understanding of normal physiology versus pathological conditions. In con- clusion, multimodal MRIs are essential tools used by healthcare professionals across many disciplines due to their ability accurately reflect differences be- tween healthy tissues and diseased states without resorting to invasive pro- cedures such as biopsy or surgery. Automation plays an important role when it comes to accurately iden- tifying areas affected by cancerous cells without damaging healthy parts of the body during surgery. Manual segmentation of brain tumors from MRI images is a time-consuming process that can be easily influenced by the subjective judgement of experts. Therefore, automated and semi-automated approaches have been developed over the last two decades to improve accu- racy and reduce the time needed for this task. The challenge with automated segmentation lies in its complexity as tumorous cells may appear anywhere inside the brain tissues with varying size, shape or appearance which makes it difficult to distinguish them from normal tissue boundaries due to poor contrast (Al-qazzaz, 2020; Baid et al., 2020). Moreover, different imaging modalities such as CT scans or Magnetic Resonance Imaging (MRI) pro- vide different types of information about a tumor’s location within the brain which requires specialized algorithms tailored specifically for each type. This has led researchers towards developing more sophisticated techniques such as deep learning models that incorporate prior knowledge into their predictions, allowing radiologists to speed up their workflow while achieving consistent results across multiple cases (Liu et al., 2022). 2. Related Works The early detection and treatment of a brain tumor are essential because it is a potentially fatal condition. In the past, traditional machine learning methods that relied on hand-crafted feature engineering were used to acquire representations from brain images. However, these methods are not sufficient for accurate brain tumor segmentation (BTS) due to their limited ability in capturing local information of the tumor regions (Raza et al., 2023 ). Recent studies show that there are currently three different types of brain tumor segmentation algorithms: classical approaches, machine learning approaches (excluding deep learning approaches), and deep learning approaches (Tie et al., 2021 ). The most popular traditional approaches are the edge detection method (Caselles et al., 1993 ; Muthukrishnan and Radha, 2012 ), threshold method (Prastawa et al., 2004 ; Stadlbauer et al., 2004 ), region growth method (Gibbs et al., 1996 ; Weglinski and Fabijanska, 2011 ), watershed approach (Lin et al., 2006 ; Maiti and Chakraborty, 2013 ), and level set approach (Menze et al., 2010; Prastawa et al., 2003 ). These approaches are rapid and inexpensive to deploy, but they have limits due to their inability to generate good results in difficult cases, such as large or irregularly shaped tumors. Edge detection works by extracting boundary information from images while threshold in- volves setting a certain intensity value that divides objects from background pixels; both have limited accuracy when dealing with noisy or blurry images. Region growing begins at a point within the object and expands outward until it reaches the boundaries defined by user input parameters; however, this technique does not always guarantee accurate results because it may miss small regions within an image or include parts outside its borders due to noise levels in the image data. Watersheds use gradient magnitude maps in combination with markers manually placed at the center of each object. Level sets employ partial differential equations that evolve contours over time according to how they interact with edges detected using gradient oper- ators. Both require more computational power than simpler techniques like edge/thresholding yet still lack robustness when faced against challenging scenarios where multiple objects exist close together sharing similar intensi- ties across their boundaries making them difficult distinguishable from each other. Machine learning methods are increasingly being utilized for the purpose of object classification and detection. These machine learning algorithms use training data to autonomously learn the features of a given dataset, allowing them to accurately classify objects within it. This has been especially useful in medical imaging applications such as brain tumor segmentation, where re- searchers have used machine learning methods like support vector machines (SVM) (Bauer et al., 2011 ; Ruan et al., 2007 ), clustering-based approaches (Stupp et al., 2017 ; Narayanan et al., 2019 ), conditional random fields (CRF) (Meier et al., 2014 , 2015 ) and random forest techniques (RF) (Tustison et al., 2015 ; Breiman, 2001 ). Bauer et al (Bauer et al., 2011 ), for example, devel- oped a classification method based on hierarchical SVM in conjunction with a CRF-based technique to segment a brain tumor from 3D MRI images. And Ruan et al. (Ruan et al., 2007 ) employed support vector machines (SVMs) to detect the abnormal region and then used multi-scales to extract the brain tumor field. Spatially Constrained Fish School Optimization (SCFSO) and Interval type-II Fuzzy Logic System (IT2FLS) algorithms were proposed by Stupp et al. (Stupp et al., 2017 ) for segmenting tumor and non-tumor re- gions in MR brain images. The segmentation results obtained by applying IT2FLS to the subsequent clustering assignment following the use of SCFOS were impressive, taking less processing time while delivering good results. For segmenting brain tumors in multi-modal MR images, Narayanan et al. (Narayanan et al., 2019 ) combined particle swarm optimization (PSO) and bacterial foraging optimization (BFO) with an upgraded clustering method based on fuzzy c means. Whereas Tustison et al. (Tustison et al., 2015 ) suggested a supervised brain tumor segmentation approach based on ran- dom forest (RF) and Markov random field that produced good performance when applied to multimodal MRI scans, Meier et al. (Meier et al., 2015 ) em- ployed decision forest for classification and conditional random field (CRF) for spatial regularization in brain tumor segmentation. The main benefit of using these types of machine learning algorithms is that they require mini- mal human intervention; once trained on a given dataset they can be applied quickly and efficiently without requiring additional input from operators or clinicians. Furthermore, their accuracy tends to improve over time as more training data becomes an important factor when dealing with rare diseases or conditions which may not have sufficient datasets readily available for anal- ysis at first instance. In recent years, deep learning which is a particular kind of machine learn- ing has made remarkable achievements in many fields, such as pattern recog- nition, image segmentation, image detection and image classification (Zhang et al., 2019 , 2021 ; Xiang et al., 2021 ). It has reentered the spotlight ever since AlexNet won the ImageNet competition in 2012 (Krizhevsky et al., 2012 ; Deng et al., 2009 ). Many researchers have applied deep learning meth- ods for medical imaging applications such as the automatic segmentation of brain tumors from MRI scans. Deep Learning models have been shown to outperform traditional machine learning algorithms when it comes to ac- curacy rate for automated BTS segmentation tasks with minimal human intervention required during the training phase itself. The early application of deep learning in brain tumor segmentation is mainly to modify the convolution network. It is able to extract features from volumetric scans using convolutional neural networks (CNNs). CNNs can capture both global and local information about the scanned image, allow- ing them to accurately detect subtle changes in tissue structure. Convolu- tional neural networks (CNNs) have revolutionized the field of deep learning and made tremendous progress in areas such as computer vision, natural language processing, and robotics. CNNs are a type of artificial neural net- work for automatically learning high-dimensional hierarchical features that use convolution operations to extract features from an input image or video. This has enabled researchers to create several specialized deep learning mod- els such as ResNet (He et al., 2016 ), VGG-Net (Szegedy et al., 2015 ), and Inception networks (Szegedy et al., 2017 ) for various tasks. These architec- tures can be used together with open-source frameworks like TensorFlow and PyTorch which makes development easier in academic fields like computer vision where semantic segmentation is used to classify objects within im- ages or videos accurately. Fully convolutional networks (FCN) enable CNNs to label each pixel using a straightforward upsampling technique which has proven invaluable when it comes to recognizing patterns at multiple resolu- tions without needing additional training data for different scales or sizes of inputs (Long et al., 2015 ). Because of the rapid progress of machine and deep learning techniques, we are now able to generate 3D volumetric data from 2D brain imaging slices in a timely, accurate, and cost-effective manner. This enables the adoption of more complicated models like U-Net, VNet, and other 3D models that can extract a greater number of features than typical 2D models. U-Net is one of the most widely used deep learning architectures for biomedical image seg- mentation. It was first introduced by Ranneberger et al. (Ronneberger et al., 2015 ) and has since gained popularity because of its ability to deliver good segmentation results despite limited training data. The U-Net architecture is a powerful tool for segmentation tasks, particularly in the medical field. Its left contracting (encoder) part is responsible for feature extraction while its right expanding (decoder) part produces an output and maps it back into raw image pixels. Skip connections are essential to this model, as they allow the extracted features from the encoder block to be passed on to the decoder block so that tumors can be detected accurately. In recent years, different studies have used combinations of skip connections with models such as U- Net++ (Zongwei et al., 2018 )d Net3+ (Huimin et al., 2020 ). Furthermore, attention models (Prajit et al., 2019 ) have also been ac- tively used in vision tasks due to their potential accuracy increase when compared with traditional methods. Attention models filter out irrelevant features from input images which save computational power and improve performance; thus making them a popular choice among researchers using U-Nets in medical applications. Several studies have been undertaken in the medical industry, particularly with U-Net, to boost the accuracy of the model with attention models filtering away unimportant elements from the input images to conserve computational power and accelerate the model. For instance, the authors of (Oktay et al., 2018 ) developed a unique attention gating approach based on the U-Net (Olaf et al., 2015 ) architecture. In their research effort, the experts found that the attention module suppresses the unnecessary areas of the input image, while accentuating the salient features. They added that an attention module-enhanced model can operate with lit- tle additional computational burden while improving the model’s sensitivity and predictive power. The use of 2D or 3D convolutions during the training of a deep CNN model is essential for accurate and effective segmentation. U-Net, Res-Net, and DenseNets (Simon et al., 2017 ) were previously proposed, but they only utilized two-dimensional input images thus missing out on additional con- textual information between slices. 2D convolutional networks are limited in their ability to capture spatial information from medical images, while 3D convolutions require more memory and computational resources. To address this issue, W. Chen et al (Chen et al., 2019 ) proposed a separable 3D U-Net model that uses separate 2D and 1D layers to learn both spatial features (2D) as well as temporal features (1D). This approach allows for full utiliza- tion of all three dimensions within the brain volume without overburdening memory requirements or taking up too much computing power. In order to train their proposed approach for each orthogonal view (axial, sagittal, and coronal) individually, the authors added separable temporal convolution to the residual inception model. For improved performance, they also used a multi-view fusion approach to merge all of the convolutional outputs. In terms of efficiency, the suggested model performed well on the BraTS 2018 test dataset. For patients with brain cancer or other illnesses that generate tumor growths inside the skull cavity, accurate brain image segmentation models are crucial for early tumor detection and enhancing survival rates. Local and global characteristics are vital for making judgments regarding tu- mor identification; however, gradients of low-level features approach zero as we move deeper into this process, making it impossible to identify borders or edges effectively without utilizing more sophisticated methods such as deep learning techniques like those utilized by these researchers in their study. The residual ResNet with a residual block was proposed by He et al. (He et al., 2016 ) to simplify neural network optimization and address the degra- dation problem in deep learning architectures. This type of network has been widely applied for image classification, recognition and segmentation tasks. Myronenko et al. (Myronenko, 2019 ) further improved this architec- ture by adding a variational auto-encoder (VAE) branch based on 3D U-Net while replacing the original U-Net convolutional blocks with residual blocks. Huang et al. (Huang et al., 2017 ) proposed Dense Convolutional Networks (DenseNets), which are designed such that all layers are directly connected while ensuring maximum information transmission between them; thereby allowing more effective feature propagation, alleviating vanishing gradients issues, and reducing parameter numbers at the same time. Subsequently, many researchers have combined DenseNets with U-nets for medical image segmentation tasks as well due to their effectiveness in these areas of appli- cation. MDU-Net (Zhang et al., 2019 ) is an improved version of U-net that uses three different multi-scale dense connections in the encoder, decoder and con- nection between them. This type of network was first proposed by Wang et al. (Wang et al., 2019 ) to segment retinal vessels and achieved good results. Building on this success, Ziang et al.(Ziang et al., 2020 ) combined Inception Residuals with dense connections to further improve the performance of U- net for medical image segmentation tasks such as tumor detection or brain tissue classification. By combining the advantages of DenseNet and ResNet, Tie et al. (Tie et al., 2021 ) proposed a new 3D U-Net with dense encoder blocks and resid- ual decoder blocks for segmenting the brain tumor multi-class 3D MR im- ages and achieved better results. Raza et al. (Raza et al., 2023 ) proposed a combined model of deep residual network and U-Net model (dResU-Net) to make predictions using both low-level and high-level information at the same time. Moreover, shortcut connections between residual networks are used to maintain low-level characteristics at each level and to enhance the training in the brain tumor segmentation process. According to results that were obtained by comparing the works, dResU-Net has greatly enhanced the segmentation performance of brain tumor sub-regions. Due to the enormous variability of MRI brain tumor images and the small proportion of brain tumor, segmentation is an extremely tough task. Although deep learning has produced significant improvement and impres- sive results in segmentation with top results of dice scores 0.8660, 0.8357, and 0.8004 for whole tumor, tumor core, and enhancing tumor core respectively on the BraTS2020 validation data set, there is no doubt that this segmen- tation accuracy could not be used for clinical diagnosis. MRI brain tumor segmentation still has to improve a lot until completely addressed. This pa- per proposes a new network architecture based on 3D Attention U-Net to solve the accuracy problem of MRI brain tumor segmentation. 3. Methodology This section discusses the dataset’s details, preprocessing procedures, and the proposed method’s implementation in detail. The proposed 3D Attention U-Net architecture is also covered in this part, along with detail on the loss functions that were used during the training stage. 3.1. Dataset The dataset that we used to train our model is taken from “Medical Im- age Computing and Computer-Assisted Intervention (MICCAI) Multimodal Brain Tumor Segmentation Challenge (BraTS) 2020” which is publicly avail- able in BraTS 2020 challenge dataset. It was gathered by medical experts from the ”University of Pennsylvania and UPenn’s Center for Biomedical Image Computing and Analysis (CBICA)” (Menze, 2015 ). It is divided into training and validation datasets which are manually annotated by neuroradiologists. It includes 3D MRI brain scans from 369 patients with gliomas, 76 of them had LGG, while the rest images are from patients with HGG. Each 3D scan consists of 155 slices, with each image measuring 240 x 240 dimensions. There are four main types of labels used to describe the different areas of a brain tumor including Label 0, Label 1, Label 2, and Label 4. Label 0 is reserved for non-tumor areas such as background or normal tissue area while Labels 1 and 4 refer to neurotic (NCR) or non- enhancing core (NET), edema (ED), respectively as well as enhancing tumor cores (ET). The regions labeled greater than 0 are referred to collectively as “whole tumor” regions with those labeled specifically 1 and 4 is called “tumor core” regions containing the ET region. 3.2. Pre-Processing The BraTS 2020 dataset was preprocessed by the competition’s organiz- ers before it was made publicly available. The images were all co-registered to the same anatomical template, interpolated to the same resolution (1 mm 3 ), and have isotropic resolution. Despite the fact that all MRI images have been pre-processed, the brightness of the images vary substantially. As a result, we normalized each 3D MR modality image using a Min-Max normalizer to re- move the impact of image intensity change on prediction accuracy. By doing so, all features will be converted into a range [0,1], with 0 and 1 serving as the minimum and maximum values for each feature/variable. It is calculated by the following formula: $${x}_{\text{scaled }}=\frac{x-\text{m}\text{i}\text{n}\left(x\right)}{\text{m}\text{a}\text{x}\left(x\right)-\text{m}\text{i}\text{n}\left(x\right)}$$ 1 where x scaled is the new value of each entry in the data and x is the old value of each entry in data. max(x) and min(x) are the maximum and mini- mum values of the features respectively. Due to memory constraints, all MR images and segmentation masks were reduced to 120 x 120 x 96 dimensions from 240 x 240 x 155. Nearest-neighbor interpolation was used to resize both the images and the masks. When resizing segmentation masks, it is necessary to apply ’nearest neighbor’ interpolation. This is because other techniques, such as Bilinear interpolation, consider the closest 2x2 neighborhood of known pixel values surrounding a pixel. It then takes a weighted average of those pixels to get the final interpolated value, which can be a float value. This creates a smoother image than nearest-neighbor interpolation, but it results in full black images for some slices having segmentation masks. Nearest-neighbor does not really interpolate the pixels. Instead, it looks for the closest one in the source image that matches the location of the target image. It’s the fastest interpolation method, and requires the least amount of processing time, and allows for sharper detail. The biggest drawback, how- ever, is that the resulting image may contain jagged edges. But it is suitable for this task as it wont affect the segmentation masks. All four modalities were stacked together to take advantage of the entire information given in four sequences. At the training time, training examples with dimensions of 120x120x96x4 (where 4 denotes the four modalities, i.e., T1, T1ce, T2, FLAIR) were provided to the model as input. 3.3. Network Architecture Due to its local and global feature extraction technique, U-Net is a promi- nent convolutional neural network used for biomedical semantic image seg- mentation. It is made up of two paths: an encoder path that extracts features from the input image and a decoder path that combines these extracted fea- tures to form the output mask. It employs skip connections between each level of encoders and their associated decoders, allowing the classifier to con- sider both low-level and high-level features when determining the segmen- tation mask. However, various constraints impede U-Net training, such as a lack of data augmentation techniques or a restricted number of available layers. It is also challenging for real-time application scenarios where speed is crucial because it needs a substantial amount of computer resources for training. The gradient vanishing problem is a limitation of a U-Net network that can lead to inefficient learning at early layers. Additionally, a large num- ber of parameters present in the model can lead to overfitting if not trained properly. Low-level features are essential for segmentation masks, as they contain information about boundaries, edges, and location of tumorous re- gions. As the network goes deeper during down-sampling operations, richer higher-level details become available than lower-level ones, but local detail and location data are lost due to convolutional and nonlinearity operations. This study introduced a dense-encoder and residual-decoder based atten- tion 3D U-Net model for the task of BTS to address the problem of vanishing gradient in the encoder section of the U-Net model during deep network train- ing. We used 3D U-Net as a baseline in our proposed architecture to develop a novel attention 3D U-Net with encoder-decoder blocks to segment brain tumors from MRI images. Our network architecture is depicted in Figure 3. The encoder section gathers context information from the input image, while the decoder section performs accurate segmentation. In Fig. 3 , the network encoder is presented on the left, and the network decoder is shown on the right. We fed four modalities 3D MR images (T1,T1Gd ,T2 and FLAIR) with sizes of 120 × 120 × 96. Thus, the input data is a 4D patch with a size of 120 × 120 × 96 × 4. The first convolutional block consists of a 7x7 3D convolution followed by a group norm and ReLu; this produces an output feature map with size 120×120×96×24. The feature map is the fed to a max pool layer followed by a double convolution layer with a kernel size of 3, which produces an output with a size of 60 × 60 × 48 × 24. Finally, the feature map goes through four dense and transition layers. Because each dense layer in a dense block has access to the feature maps of all dense layers preceding it, it has global information about the entire block, which is useful for accurate segmentation. The contribution of the output feature maps of each dense layer to the global information is con- trolled by the growth rate of the dense blocks. A dense layer is made up of one 1 × 1 × 1 convolution layer and one 3 × 3 × 3 convolution layer. Following each convolution layer, group normalization and the ReLU activation function are utilized. We first perform a 1x1 convolution with 128 filters to reduce the size of the feature maps, and then perform a more expensive 3x3 convolution with a growth rate of 32 feature maps which uses padding to ensure the dimensions remain constant. Concatenation is then used to com- bine the output from each layer in order to preserve the information from earlier layers. The size of the feature vectors is reduced by using a transition block. As opposed to a pooling layer, Transition Block performs a 1x1 convolution with 32 filters, followed by another 1x1 convolution with a stride of 2, which reduces the size of the volume and the number of feature maps by half. In order to reduce the information loss associated with pooling, convolution should be performed with a stride of 2. Each convolution is then followed by a group norm and ReLU layers. A trilinear interpolation layer is used for upsampling, which is subse- quently concatenated with low-level features from the corresponding dense block. The output is then passed through a spatial and channel attention block first. Our proposed attention module combines channel, spatial, and skip connection in parallel. However, combining excited features at the same time runs the danger of making feature learning inconsistent. As demon- strated in Fig. 5 , we integrate skip connections to reduce network redun- dancy and sparsity. Similar to (Islam et al., 2020 ) by utilizing 3D inter-spatial and inter- channel feature relationships, we introduce 3D attention units to generate 3D spatial and channel attention. We initially carry out a 1x1xC convo- lution to group all spatial feature correlations into the HxWx1 dimension, after which we carry out average pooling and feed it to the neural network to produce the 3D attention map. We combine skip-connection to reduce the singularity and sparsity brought on by these parallel excitations, improving segmentation prediction and allowing for more generalized learning. The structure of the residual block is shown in Fig. 4 , which is com- prised of two 3 × 3 × 3 convolution layers with stride of 1. Group normaliza- tion and ReLU activation functions are used after to each convolution layer. The output of the residual block is the addition of the output of the last convolution layer and the input of the residual block. The shortcut connec- tion is first passed through an additional 1 × 1 × 1 convolution layer with stride of 1. From the top to the bottom of the encoder, the number of feature map increases and the size of feature map reduces progressively. Specifically, Group Norm layer (GN) and activation layer were applied after each convo- lution layer. The Group Norm aids in creating deeper and wider networks. It first groups the channels and normalize inside each group. It is a trade-off between LayerNorm and InstanceNorm. When batch sizes are small, group normalization outperforms batch normalization. The batch size in our work is one. As a result, we applied group normalization across the entire net- work. However, when the batch size is significantly large, GN does not scale as well as BN and might not be able to match the performance of BN. The activation layer implements a rectified linear unit (ReLU) layer. The ReLU is employed to optimize the proposed architecture and consequently enhance the performance. 3.4. Losses and Evaluation metrics The loss and evaluation metrics utilized to evaluate the performance of the proposed model are explored in detail in this section. By minimizing the loss function, the proposed model learns parameters/features from MR images. In the BraTS datasets, about 98 percent of total voxels are non- tumor (Label 0, commonly known as background), with only a few voxels belonging to tumor (Labels 2, 3, 4). We used a BCE-Dice loss function-based region to overcome the class imbalance problem. This loss combines the Dice loss with the standard binary cross-entropy (BCE) loss, which is commonly used in segmentation models. Mixing the two approaches allows for some loss variability while benefiting from BCE’s stability. The equation for multi-class BCE by itself will be familiar to anyone who has studied logistic regression: As evaluation metrics, we used the IoU metric, also known as the Jaccard Index, and the Dice coefficient, also known as the Dice-Sørensen coefficient. The Dice score, which calculates the discrepancies between the segmentation result and the ground truth, is a widely used metric for pixel segmentation. The dice score can be calculated as follows for a given predicted and ground truth segmentation X and Y, where each tumor voxel is labeled as 1, and non-tumor voxels are labeled as 0. $$DSC(X,Y)=\frac{2|X\cap Y|}{\left|X\right|+\left|Y\right|}$$ 3 The Jaccard metric is similar to the Dice metric in that it calculates the ratio between the overlap of positive instances between two sets by evaluating the performance of pixel segmentation models and their mutual combined values: $$J(A,B)=\frac{|A\cap B|}{|A\cup B|}=\frac{|A\cap B|}{\left|A\right|+\left|B\right|-|A\cap B|}$$ 4 4. Experimental Results And Analysis 4.1 Implementation details The Python programming language and PyTorch as its backend were used to implement the suggested model. Due to constrained memory and compu- tational resources, the model was trained for 100 epochs with a batch size of 1. Lion (EvoLved Sign Momentum) optimizer with a learning rate of 0.0001 and weight decay of 0.01 was used to replace Adam. It is a new optimizer discovered by Google Brain that is purportedly to be better than Adam(w), in Pytorch. Each layer of the network was normalized and made more sta- ble by using the activation function ReLU with group normalization. The studies were carried out using the BraTS 2020 benchmark dataset, with 201 images used for training, 34 images for validation, and 134 images for testing purposes. The detailed information about the hyperparameter values that were provided during the training of the model is given in Table 1. Table 1: List of hyperparameters used for 3D Attention UNet 3D Attention U-Net clearly outperforms other approaches when it comes to model complexity. It took around 11.8 hours to train the proposed model through 100 epochs, with an execution time of 7.12 minutes per epoch. It can be deployed in a real-world clinical setting because our suggested model’s online test time for one person was less than one second. By leveraging more effective computational resources, the model’s efficiency may be improved even more. Detailed information about the hardware requirement and com- putational efficiency of the proposed model is given in Table 2. Table 2: Hardware requirement and computational efficiency of the proposed model 4.2 Results and discussion Figure 6 shows the bar plot of dice and jaccard score obtained by our network on BraTS2020 validation dataset. Table 3 lists the value of dice and jaccard of ET, WT, and TC in the three regions of our segmentation results. The average value, standard deviation, median value, 25 percent quantile value, and 75 percent quantile value are given for each indicator value. It can be seen from Table 3 that the dice mean of WT, TC, and ET are 0.889, 0.866, and 0.828, respectively, and the Jaccard mean of ET, WT, and TC are 0.718, 0.809, and 0.779, respectively. The WT region has the highest prediction accuracy in our method, whereas the ET region has the lowest. This is due to the fact that the WT region is the largest and the ET region is the smallest. The greater the region, the easier it is to predict. And the smaller the region, the more difficult it is to predict. Table 3: The performance of our proposed method on BraTS 2020 validation dataset. The performance of our method is compared with the original 3D U-Net method. Figure 7 and Table 4 list the dice and jaccard score obtained by 3D U-Net on the BraTS2020 validation dataset for each class. The proposed technique achieved a higher dice score for each sub-region of the brain tumor than the baseline work because of the use of the dense block with a tran- sition at the decoder part and residual convolutional blocks at the encoder part. The proposed Attention 3D U-Net model significantly improves the segmentation accuracy of the ET class which was the most difficult class to segment. Table 4: The performance of original 3D U-Net on BraTS 2020 validation dataset. Dice Jaccard WT TC ET WT TC ET Mean 0.885 0.793 0.761 0.8 0.682 0.646 StdDev 0.69 0.169 0.206 0.107 0.189 0.202 Median 0.895 0.849 0.815 0.81 0.738 0.687 25quantile 0.847 0.762 0.763 0.735 0.616 0.616 75quantile 0.94 0.896 0.872 0.887 0.811 0.772 Our method was qualitatively compared with original 3D U-Net. We randomly selected a patient and used both our method and original 3D U- Net to segment the brain tumor. Figure 8 shows slices of the four modalities of ground truth of the patient and the slices of the predicted segmentation image using our method and the ground truth. 4.2 Comparison with state-of-the-art studies In comparison to state of art models for brain tumor semantic segmenta- tion, the proposed method is able to segment these individual tumor regions (WT, TC, and ET) close to the ground truth, as shown in the findings. The most challenging areas to segment were the enhancing tumor and its dis- persion with necrosis. When it comes to segmenting augmenting enhancing tumor and tumor core, many existing segmentation models performed badly. The proposed Attention 3D U-Net model, on the other hand, successfully seg- mented these regions. Table 5 also illustrates the quantitative performance of our model, which performs better than the state-of-the-art techniques for enhancing tumor and tumor core classes. In the case of segmenting the whole tumor, the proposed method achieves good results, comparable to state of art techniques. Overall, the proposed method seems to be a better technique to generate segmented images that are closest to the ground truth. Table 5: Comparison of proposed 3D Attention UNet model with state-of-the-art methods Reference /Study Image Dimension Dataset Dice Score (DSC ) Tumor Core Whole Tumor Enhancing Tumor (TC ) (WT ) (ET ) W. Wang et al. (Wang et al., 2021) 128 x 128 x 128 BraTS 2020 0.8173 0.9009 0.7873 O¨ .C¸ i¸cek et al. (Cicek et al., 2016) 128 x 128 x 128 BraTS 2020 0.7906 0.8411 0.6876 J. Colman et al. (Colman et al., 2021) 2DSlices with 240 x 240 BraTS 2020 0.7983 0.8673 0.7514 H. Messaoudi et al. (Messaoudi et al., 2021) 192 x 160 x 108 BraTS 2020 0.7520 0.8068 0.6959 L. M. Ballestar et al. (Ballestar and Vilaplana, 2020) 64 x 64 x 64 BraTS 2020 0.7526 0.8463 0.6215 M. Ghaffari et al. (Ghaffari et al., 2021) 128 x 128 x 128 BraTS 2020 0.82 0.90 0.78 F. Wang et al. (Wang et al., 2020) 128 x 128 x 128 BraTS 2019 0.798 0.852 0.778 J. Zhang et al. (Zhang et al., 2020) 128 x 128 x 128 BraTS 2019 0.777 0.870 0.709 A. Myronenko (Myronenko, 2018) 160 x 192 x 128 BraTS 2018 0.8154 0.8839 0.7664 Yang et al. (Yang and Yang, 2018) 96 x 96 x 96 BraTS 2018 0.789 0.869 0.722 Raza et al. (Raza et al., 2023) 128 x 128 x 128 BraTS 2020 0.8357 0.8660 0.8004 Proposed model 120 x 120 x 96 BraTS 2020 0.866 0.889 0.828 Table 6 shows the training time, prediction time, and number of parame- ters. With 29.5 MB million parameters, our model takes about 11.8 hours to train and 52.65 seconds to predict. This demonstrates that our model is substantially lower in size and trains faster while maintaining the same per- formance. Due to their computationally efficient designs, new transformer- based architectures improved inference times relative to other architectures, however they still fall short of our suggested model in terms of performance. Table 6: Complexity of models used in comparison process. Methods Training Time (h ) Prediction Time (s ) Parameters Our method 11.8 hours 52.65 29.5 mln 3DU -Net 15.2 76.2 19mln 3DRes -Net 24.4 65.1 46.4mln Swin BTS 13.2 58.7 16.3mln VT U -Net 15.8 59.8 18.4mln 3DLI -Net ++ 25.6 64.5 51.6mln Attention 3DUI -Net 18.3 121.4 22.3mln 3DRes -U -Net 29.2 69.3 58.6mln . 3DDeepLab V3 35.8 78.6 76.1mln 3DPSP -Net 36.5 96.9 72.4.mln 5. Conclusion By combining the advantages of ResNet and DenseNet, this paper pro- poses a deep 3D Attention UNet with dense encoder blocks and residual de- coder blocks to increase the performance of brain tumor segmentation from MRI images. The use of dense blocks has the advantage of reducing network parameters, deepening network layers, strengthening feature propagation, avoiding the vanishing-gradient problem, and enlarging receptive fields. In comparison to typical convolution blocks, residual blocks address the issue of network performance degradation as network depth increases. We also in- troduced 3D attention units to produce 3D spatial and channel attention by leveraging 3D inter-spatial and inter-channel feature correlations. The proposed method surpassed state-of-the-art models, yielding dice scores of 0.866, 0.889, and 0.828 for TC, WT, and ET, respectively, without the use of substantial post-processing. Furthermore, our model takes approximately 11.8 hours to train and 52.65 seconds to predict with only 29.5 MB million parameters. This shows that while maintaining the same performance, our model is significantly smaller and trains faster than other models. By using this approach in clinical oncology, radiologists and oncologists can better un- derstand the different tumor locations, sizes, and shapes, which will aid them with tumor diagnosis, treatment planning, and prognosis. More 3D-based ar- chitectures should be investigated in order to minimize computational costs while maximizing contextual information. Post-processing approaches can be utilized to improve the performance of the model, and this research can be applied to clinically challenging medical imaging problems and other seg- mentation applications. Declarations Acknowledgement We want to thank Northwestern Polytechinical University for helping us to conduct research in this area and providing resource. Competing interests: The authors declare no competing interests. References Al-qazzaz, S.A.L.I.R., 2020. Deep learning-based brain tumour image seg- mentation and its extension to stroke lesion segmentation . Alagarsamy, S., Zhang, Y.D., Govindaraj, V., Murugan, P.R., Sankaran, S., 2020. Smart identification of topographically variant anomalies in brain magnetic resonance imaging using a fish school based fuzzy clustering ap- proach. IEEE Transactions on Fuzzy Systems. Azhari, E.E.M., Hatta, M.M., Htike, Z.Z., Win, S.L., 2014. Tumor detection in medical imaging: a survey. Int. J. Adv. Inf. Technol. 4, 21–30. Baid, U., Talbar, S., Rane, S., Gupta, S., Thakur, M., Moiyadi, A.and Sable, N.A.M., Mahajan, A., 2020. A novel approach for fully automatic intra- tumor segmentation with 3d u-net architecture for gliomas. Front. Comput. Neurosci. 14. Ballestar, L.M., Vilaplana, V., 2020. Brain tumor segmentation using 3d- cnns with uncertainty estimation. ArXiv, 1–11. Bauer, S., Nolte, L.P., Reyes, M., 2011. Fully automatic segmentation of brain tumor images using support vector machine classification in combi- nation with hierarchical conditional random field regularization. Medical image computing and computer-assisted intervention 6893, 354–361. Bauer, S., Wiest, R., Nolte, L.P., Reyes, M., 2013. A survey of mri-based medical image analysis for brain tumor studies. Phys. Med. Biol. 58, R97–R129. Breiman, L., 2001. Random forests. Machine Learning 45(1), 5–32. Caselles, V., Catt´e, F., Coll, T., Dibos, F., 1993. A geometric model for active contours in image processing. Numerische Mathematik 66(1), 1–31. Chen, W., Liu, B., Peng, S., Sun, J., Qiao, X., 2019. S3d-unet: Separable 3d u-net for brain tumor segmentation, in: In: Crimi, A., Bakas, S., Kuijf, H., Keyvan, F., Reyes, M., van Walsum, T. (eds) Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. BrainLes 2018. Lecture Notes in Computer Science(.), Springer, Cham. Cicek, O., Abdulkadir, A., Lienkamp, S., Brox, T., Ronneberger, O., 2016. 3d u-net: Learning dense volumetric segmentation from sparse annotation. Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics) 9901 LNCS, 424–432. Colman, J., Zhang, L., Duan, W., Ye, X., 2021. Dr-unet104 for multimodal mri brain tumor segmentation. Lect. Notes Comput. Sci. (including Sub- ser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics) 12659 LNCS, 410–419. Crimi, A., Bakas, S., Kuijf, H., Keyvan, F., Reyes, et al., M., 2019. Brainle- sion: Glioma, multiple sclerosis, stroke and traumatic brain injuries. 4th International Workshop, BrainLes 2018 11384. Deng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L., 2009. Imagenet: A large-scale hierarchical image database, in: In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p. 248–255. Ellison, J., 2020. Improving the generalizability of convolutional neural net- works for brain tumor segmentation in the post-treatment setting . Ghaffari, M., Sowmya, A., Oliver, R., 2020. Automated brain tumor segmen- tation using multimodal brain scans: A survey based on models submitted to the brats. 2012–2018 challenges. IEEE Reviews in Biomedical Engineer- ing 13, 156–168. Ghaffari, M., Sowmya, A., Oliver, R., 2021. Automated brain tumour seg- mentation using cascaded 3d densely-connected u-net. Lect. Notes Com- put. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinfor- matics) 12658 LNCS, 481–491. Gibbs, P., Buckley, D.L., Blackband, S.J., Horsman, A., 1996. Tumour volume determination from mr images by morphological segmentation. Physics in Medicine Biology 41(11), 2437–2446. He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition, in: In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p. 770–778. Huang, G., Liu, Z., Maaten, L.V.D., Weinberger, K.Q., 2017. Densely con- nected convolutional networks, in: Computer Vision and Pattern Recog- nition (CVPR), p. 2261–2269. Huimin, H., Lin, L., Tong, R., Hu, H., Zhang, Q., Iwamoto, Y., Han, X., Chen, Y.W., Wu, J., 2020. U-net 3+: A full-scale connected u-net for medical image segmentation, in: In Proceedings of the ICASSP 2020–2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona, Spain, 4–8 May, p. 1055–1059. Islam, M., Vibashan, V.S., Jose, V.J.M., Wijethilake, N., Utkarsh, U., Ren, H., 2020. Brain tumor segmentation and survival prediction using 3d at- tention unet. In Brainlesion: Glioma, Multiple Sclerosis, Stroke and Trau- matic Brain Injuries: 5th International Workshop, BrainLes 2019, Held in Conjunction with MICCAI 2019, Shenzhen, China, October 17, 2019, Revised Selected Papers, Part I 5, 262–272. Krizhevsky, A., Sutskever, I., Hinton, G., 2012. Imagenet classification with deep convolutional neural networks, in: In: Proceedings of the 25th International Conference on Neural Information Processing Systems, p. 1097–1105. Lin, Y.C., Tsai, Y.P., Hung, Y.P., Shih, Z.C., 2006. Comparison be- tween immersion-based and toboggan- based watershed image segmenta- tion. IEEE Transactions on Image Processing 15(3), 632–640. Liu, Z., Tong, L., Chen, L., Jiang, Z., Zhou, F., Zhang, Q., Zhang, X., Jin, Y., Zhou, H., 2022. Deep learning based brain tumor segmentation: A survey. Complex Intelligent Systems 9, 1001–1026. Long, J., Shelhamer, E., Darrell, T., 2015. Fully convolutional networks for semantic segmentation, in: In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p. 3431–3440. Maiti, I., Chakraborty, M., 2013. A new method for brain tumor segmen- tation based on watershed and edge detection algorithms in hsv colour model, in: 2012 National Conference on Computing and Communication Systems, p. 1–5. Meier, R., Bauer, S., Slotboom, J., Wiest, R., Reyes, M., 2014. Appearance- and context-sensitive features for brain tumor segmentation, in: Proceed- ings of MICCAI BraTS Challenge 2014, p. 20–26. Meier, R., Karamitsou, V., Habegger, S., Wiest, R., Reyes, M., 2015. Param- eter learning for crf-based tissue segmentation of brain tumors. MICCAI Brainlesion Workshop 9556, 156–167. Menze, B.H., Leemput, K.V., Lashkari, D., Weber, M.A., Ayache, N.e.a., 2010. A generative model for brain tumor segmentation in multimodal im- ages. Medical image computing and computer-assisted intervention 6362, 151–159. Menze, B.H.e.a., 2015. The multimodal brain tumor image segmentation benchmark (brats). IEEE Transactions on Medical Imaging 34(10), 1993–2024. Messaoudi, H., Belaid, A., Allaoui, M.L., Zetout, A., Allili, M.S., Tliba, S., Conze, P.H., 2021. Efficient embedding network for 3d brain tumor segmentation. Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics) 12658 LNCS, 252–262. Muthukrishnan, R., Radha, M., 2012. Edge detection techniques for image segmentation. International Journal of Computer Science Information Technology 3(6), 250–254. Myronenko, A., 2018. 3d mri brain tumor segmentation using autoencoder regularization, in: In: International MICCAI Brainlesion Workshop, p. 311–320. Myronenko, A., 2019. 3d mri brain tumor segmentation using autoencoder regularization. Inter- national MICCAI Brainlesion Workshop, Cham, Switzerland, Springer 11384, 311–320. Narayanan, A., Rajasekaran, M.P., Zhang, Y.D., Govindara, V., Thiyagara- jan, A., 2019. Multi-channeled mr brain image segmentation: A novel double optimization approach combined with clustering technique for tu- mor identification and tissue segmentation. Biocybernetics and Biomedical Engineering 39(2), 350–381. Oktay, O., Schlemper, J., Folgoc, L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N., Kainz, B.e.a., 2018. Attention u-net: Learning where to look for the pancreas. ArXiv abs/1804.03999. Olaf, R., Fischer, P., Brox, T., 2015. U-net: Convolutional networks for biomedical image segmentation, in: In International Conference on Med- ical Image Computing and Computer Assisted Intervention; Springer: Cham, Switzerland, p. 234–241. Prajit, R., Parmar, N., Vaswani, A., Bello, I., Levskaya, A., Shlens, J., 2019. Stand-alone self-attention in vision models, in: In Proceedings of the Advances in Neural Information Processing Systems 32, Vancouver, BC, Canada, p. 8–14. Prastawa, M., Bullitt, E., Ho, S., Gerig, G., 2004. A brain tumor segmenta- tion frame work based on outlier detection. Medical Image Analysis 8(3), 275–283. Prastawa, M., Bullitt, E., Moon, N., van Leemput, K., Gerig, G., 2003. Automatic brain tumor segmentation by subject specific modification of atlas priors 1. Academic Radiology 10(12), 1341–1348. Raza, R., Ijaz Bajwa, U., Mehmood, Y., Waqas Anwar, M., Hassan Jamal, M., 2023. dresu-net: 3d deep residual u-net based brain tumor segmen- tation from multimodal mri. Biomedical Signal Processing and Control 79(P1), 103861. Ronneberger, O., Fischer, P., Brox, T., 2015. Convolutional networks for biomedical image segmentation, in: In: International Conference on Med- ical image computing and computer assisted intervention, p. 234–241. Ruan, S., Lebonvallet, S., Merabet, A., Constans, J., 2007. Tumor segmen- tation from a multispectral mri images by using support vector machine classification, in: IEEE International Symposium on Biomedical Imaging: From Nano to Macro, p. 1236–1239. Simon, J., Drozdzal, M., Vazquez, D., Romero, A., Bengio, Y., 2017. The one hundred layers tiramisu: Fully convolutional densenets for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, Honolulu, HI, USA 21–26, 11–19. Stadlbauer, A., Moser, E., Gruber, S., Buslei, R., Nimsky, C.e.a., 2004. Improved delineation of brain tumors: An automated method for seg- mentation based on pathologic changes of 1h-mrsi metabolites in gliomas. Neuroimage 23(2), 454–461. Stupp, R., Taillibert, S., Kanner, A., Read, W., Ram, Z., 2017. Effect of tumor-treating fields plus main- tenance temozolomide vs maintenance temozolomide alone on survival in patients with glioblastoma: A random- ized clinical trial. JAMA 318(23), 2306–2316. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A., 2017. Inception-v4, inception- resnet and the impact of residual connections on learning, in: In: Thirty-First AAAI Conference on Artificial Intelligence, p. 4278–4284. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A., 2015. Going deeper with convolu- tions, in: In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p. 1–9. Tanneedi, R.V., Pedapati, P., Johansson, S., 2017. Brain tumour detection using hog by svm. Online. Tie, J., Peng, H., Zhou, J., 2021. Mri brain tumor segmentation using 3d u-net with dense encoder blocks and residual decoder blocks. CMES - Computer Modeling in Engineering and Sciences 128(2), 427–445. Tustison, N., Shrinidhi, K.L., Wintermark, M., Durst, C.R., 2015. Opti- mal symmetric multimodal templates and concatenated random forests for supervised brain tumor segmentation (simplified) with antsr. Neuroin- formatics 13(2), 209–225. Wang, C., Zhao, Z., Ren, Q., Xu, Y., Yu, Y., 2019. Dense u-net based on patch-based learning for retinal vessel segmentation. Entropy 21(2), 168. Wang, F., Jiang, R., Zheng, L., Meng, C., Biswal, B., 2020. 3d u-net based brain tumor segmentation and survival days prediction. 11992, 131–141. Wang, W., Chen, C., Ding, M., Yu, H., Zha, S., Li, J., 2021. Transbts: multi- modal brain tumor segmentation using transformer. Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformat- ics) 12901 LNCS, 109–119. Weglinski, T., Fabijanska, A., 2011. Brain tumor segmentation from mri data sets using region growing approach, in: 2011 Proceedings of 7th In- ternational Conference on Perspective Technologies and Methods in MEMS Design, MEMSTECH2011, p. 185–188. Xiang, Y., Wang, S.H., Zhang, Y.D., 2021. Cgnet: A graph-knowledge em- bedded convolutional neural network for detection of pneumonia. Infor- mation Processing Management 58(1), 1–25. Yang, H., Yang, J., 2018. Automatic brain tumor segmentation with con- tour aware residual network and adversarial training, in: In: International MICCAI Brainlesion Workshop, Springer, Cham, p. 267–278. Zhang, Y.D., Dong, Z., Wang, S.H., Yu, X., Gorriz, J.M., 2020. Advances in multimodal data fusion in neuroimaging: Overview, challenges, and novel orientation. Information Fusion 64, 149–187. Zhang, Y.D., Dong, Z.C., Wu, L., Wang, S.H., 2011. A hybrid method for mri brain image classification. Expert Systems with Applications 38, 10049–10053. Zhang, Y.D., Govindaraj, V., Tang, C.S., Zhu, W.G., Sun, J.D., 2019. High performance multiple sclerosis classification by data augmentation and alexnet transfer learning model. Medical Imaging and Health Informat- ics 9(9), 2012–2021. Zhang, Y.D., Satapathy, S.C., Guttery, D.S., Gorriz, J., Wang, S.H., 2021. Improved breast cancer classification through combining graph convolu- tional network and convolutional neural network. Information Processing and Management 58(2), 1–25. Ziang, Z., Wu, C., Coleman, S., Kerr, D., 2020. Dense-inception u-net for medical image segmentation. Computer Methods andPrograms in Biomedicine 192, 1–15. Zongwei, Z., Siddiquee, M., Tajbakhsh, N., Liang, J., 2018. U-net++: A nested u-net architecture for medical image segmentation, in: In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support; Springer: Cham, Switzerland, p. 3–11. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-2717573","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Method Article","associatedPublications":[],"authors":[{"id":185855767,"identity":"11ec824c-e285-4a02-b379-c3c592dbd044","order_by":0,"name":"Tewodros Megabiaw Tassew","email":"data:image/png;base64,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","orcid":"","institution":"Northwestern Polytechnical University","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Tewodros","middleName":"Megabiaw","lastName":"Tassew","suffix":""},{"id":185855943,"identity":"3a369596-1c26-4007-895c-1952338a2e78","order_by":1,"name":"Betelihem Asfaw Ashamo","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCElEQVRIiWNgGAWjYFACxgcMjA1Amp2xDUjagEQaD+DXwmwA0iLBwAzWkgbS0kCsFgY2IO8wWAyvFv7+w4yfeXfY1PEzM7c9+Jhz3m5t+2GgLTU20bi0SNxIZpbmPZMmIdnM2G44c9vt5G1nEoFajqXlNuDSc4P/gHRu22EJg8OMbdK8QC1mB4BaGBsO49Qif/4w8+/ctv8S9iAtf7edSzY7/xC/FoMDyWxAWw5IGABDTJpx2wE7sxsEbDG8kcxm/fdMsuQMoC2SvduSE8xuAG1JwOMXOaDDbs7cYcfP397+TOLnNjt7s/PpDx98qLHB7X10kAhWmUCschCwJ0XxKBgFo2AUjAwAADoHYyrrKu6QAAAAAElFTkSuQmCC","orcid":"","institution":"Northwestern Polytechnical University","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Betelihem","middleName":"Asfaw","lastName":"Ashamo","suffix":""},{"id":185855944,"identity":"d781435b-340f-40f9-a8cd-4ea632fb8450","order_by":2,"name":"Xuan Nie","email":"","orcid":"","institution":"Northwestern Polytechnical University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Xuan","middleName":"","lastName":"Nie","suffix":""}],"badges":[],"createdAt":"2023-03-21 08:55:03","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":true,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false,"coiExplicitlySet":false},"doi":"10.21203/rs.3.rs-2717573/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-2717573/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":34882851,"identity":"aa3fb3f9-30b7-44ce-8357-9fc14dbe870e","added_by":"auto","created_at":"2023-03-27 20:55:38","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":195247,"visible":true,"origin":"","legend":"\u003cp\u003eFour MRI modalities axial slices of a patient data from the BraTS2020 training dataset. From (a) to (d) are T1, T1Gd, T2 and FLAIR. (e) The ground truth segmenta- tion, that purple indicates peritumoral edema, yellow indicates non enhancing tumor core, blue indicates GD-enhancing tumor\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-2717573/v1/66bed9cfbc8027eead08d3ab.png"},{"id":34882834,"identity":"0fbacabe-25ed-48ba-8732-fcf2aa12fe6a","added_by":"auto","created_at":"2023-03-27 20:55:37","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":166215,"visible":true,"origin":"","legend":"\u003cp\u003ePreparing the data and pre-processing workflow\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-2717573/v1/925007721df7f0ea001e8dea.png"},{"id":34882852,"identity":"53a29b6a-4d4a-4820-b3f8-201d974ec11d","added_by":"auto","created_at":"2023-03-27 20:55:38","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":269311,"visible":true,"origin":"","legend":"\u003cp\u003eNetwork Architecture\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-2717573/v1/bf3d193d40ad9adc9b082a3e.png"},{"id":34882855,"identity":"88b79bf7-5482-4da5-8b6a-0f00d3e17e48","added_by":"auto","created_at":"2023-03-27 20:55:38","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":476875,"visible":true,"origin":"","legend":"\u003cp\u003eBuilding Blocks Explanation\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-2717573/v1/eac8b329a334b6bbd75d5cd3.png"},{"id":34882847,"identity":"d36d179f-6d31-4574-9f87-4dd3f01cb9db","added_by":"auto","created_at":"2023-03-27 20:55:38","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":123503,"visible":true,"origin":"","legend":"\u003cp\u003eVisual representation of the 3D spatial and channel attention with skip connection.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-2717573/v1/145d427c6fa54528198ddd29.png"},{"id":34882849,"identity":"21636fe3-eef4-4d8e-adb2-0fda32b258ff","added_by":"auto","created_at":"2023-03-27 20:55:38","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":23687,"visible":true,"origin":"","legend":"\u003cp\u003eDice and Jaccard Coefficients Mean from Validation of our model.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-2717573/v1/6d7bbde93c0bdb4f638c3a47.png"},{"id":34882838,"identity":"acd3dba5-e27e-48fd-a075-ff9d4a8ac391","added_by":"auto","created_at":"2023-03-27 20:55:38","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":22918,"visible":true,"origin":"","legend":"\u003cp\u003eDice and Jaccard Coefficients Mean from Validation of 3D U-Net.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-2717573/v1/4305f3923541abf12c47004e.png"},{"id":34883398,"identity":"1c300e56-e6ec-41be-8001-184f43824f3f","added_by":"auto","created_at":"2023-03-27 21:03:38","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":234642,"visible":true,"origin":"","legend":"\u003cp\u003eQualitative segmentation results of our model on the BRATS-2020 dataset.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-2717573/v1/92c4665ecf8b2c557308d063.png"},{"id":34883399,"identity":"e4626334-14a6-427a-a02c-02dadf14932a","added_by":"auto","created_at":"2023-03-27 21:03:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1711498,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-2717573/v1/18f3b568-3d82-47d2-9d22-c994e8cfdaa2.pdf"}],"financialInterests":"","formattedTitle":"\u003cp\u003eMutltimodal MRI Brain Tumor Segmentation using 3D Attention UNet with Dense Encoder Blocks and Residual Decoder Blocks\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eBrain tumors are a serious medical condition that can have devastating effects on the nervous system and surrounding healthy brain tissue. They are divided into two main categories: primary and metastatic. Primary brain tumors originate from the cells of the brain or its immediate vicinity, while metastatic ones form elsewhere in the body then spread to the brain through circulation. Primary tumors can be classified further as glial, composed of glial cells, or non-glial, formed on or in other structures such as nerves, blood vessels, and glands; these may be benign (non-cancerous)or malignant (can- cerous). Treatment for both types varies depending on how advanced they are but typically involves surgery followed by radiation therapy or chemother- apy depending upon individual circumstances. In the United States alone, 700,000 people had primary brain tumors, with over 88,000 diagnosed by 2022. According to the World Health Organization, cancer causes 9 million deaths globally each year, accounting for 22 percent of all chronic diseases and ranking second only to cardiovascular disease, which causes 17.9 million deaths each year. Primary malignant brain tumors, for example, account for 1.4 percent of all new cancer diagnoses in the United States each year. Primary malignant brain tumors have a global mortality rate of 2.0/100,000 in women and 2.8/100,000 in men (Ellison, 2020).\u003c/p\u003e\n\u003cp\u003eMedical imaging is an essential tool in the diagnosis and treatment of tumors. It allows for accurate tumor segmentation, assessment of tumor size and shape, and evaluation of response to treatments such as radiation or chemotherapy. The most commonly used medical imaging techniques are computed tomography (CT) scans, ultrasound examinations, X-ray exami- nations, and magnetic resonance imaging (MRI). These technologies allow healthcare professionals to observe the structure of tissues within a patient\u0026rsquo;s\u003c/p\u003e\n\u003cp\u003ebody in order to accurately diagnose tumors or assess how well a treatment is working. Medical imaging plays an important role in providing early de- tection of cancerous lesions which can lead to better prognosis outcomes for patients with tumors (Raza et al., 2023). In clinical medicine, MRI has be- come the most widely used imaging examination for brain tumor diagnosis, because of its sharpness and high tissue resolution, ability to produce images of the brain\u0026rsquo;s tissues in clear contrast, and versatility in allowing different parameters to be set in order to obtain specific anatomical information (Raza et al., 2023; Alagarsamy et al., 2020). It is a safe and advanced imaging pro- cedure that makes use of nuclear magnetic resonance to produce a reliable depiction of internal body structure and tissues without causing harm to the organ being imaged or causing high levels of ionization or radiation effect. It creates an accurate portrayal of internal body structure and tissues by combining strong magnetic waves ranging from 1.5 T to 3 T with radiofre- quency waves (Azhari et al., 2014; Tanneedi et al., 2017). The use of MRI for glioma diagnoses allows physicians access to detailed information about the cells within a patient\u0026rsquo;s brain by offering up sub-regions such as peritumoral edema, necrotic core, enhancing tumor core, and non-enhancing tumor core depending on factors such as degree of invasion or prognosis. This technology also helps healthcare professionals determine if any additional treatments are necessary after the initial diagnosis has been made since it can provide fur- ther insight into any changes that may have occurred over time regarding the tumor\u0026rsquo;s size or shape prior to beginning treatment (Alagarsamy et al., 2020; Crimi et al., 2019; Ghaffari et al., 2020).\u003c/p\u003e\n\u003cp\u003eMultimodal MRI scans are a powerful tool for diagnosing and monitoring diseases of the brain. By combining four different modalities, T1 weighted (T1), T1 weighted with post-contrast (T1Gd), T2 weighted (T2), and Fluid Attenuated Inversion Recovery (FLAIR) scans, it is possible to gain an ac- curate representation of the various histological sub-regions within the brain (Ghaffari et al., 2020; Zhang et al., 2011). The purpose of each modality can be broadly understood based on their distinct intensity distributions (Zhang et al., 2020). For example, T1w is mainly used to measure healthy tissues. While T2w has a bright tumor region visible on scan results, the FLAIR scan helps distinguish edema from cerebrospinal fluid (CSF) (Bauer et al., 2013). Meanwhile, both the whole tumor area as well as its core can be seen clearly when viewing axial slices through both a FLAIR and T2 scan, respectively. Additionally, a post-contrast enhanced view via a T1Gd allows for the vi-\u003c/p\u003e\n\u003cp\u003esualization of enhancing tumors. This combination provides clinicians with valuable insight into disease states that would otherwise remain undetected or overlooked.\u003c/p\u003e\n\u003cp\u003eAn example demonstrating these four MRI modalities is shown in Figure 1, where ground truth segmentation was applied to patient data from the BraTS2020 training dataset. This demonstrates how multimodal MRIs pro- vide invaluable information regarding tissue structure, which helps build an understanding of normal physiology versus pathological conditions. In con- clusion, multimodal MRIs are essential tools used by healthcare professionals across many disciplines due to their ability accurately reflect differences be- tween healthy tissues and diseased states without resorting to invasive pro- cedures such as biopsy or surgery.\u003c/p\u003e\n\u003cp\u003eAutomation plays an important role when it comes to accurately iden- tifying areas affected by cancerous cells without damaging healthy parts of the body during surgery. Manual segmentation of brain tumors from MRI images is a time-consuming process that can be easily influenced by the subjective judgement of experts. Therefore, automated and semi-automated approaches have been developed over the last two decades to improve accu- racy and reduce the time needed for this task. The challenge with automated segmentation lies in its complexity as tumorous cells may appear anywhere inside the brain tissues with varying size, shape or appearance which makes it difficult to distinguish them from normal tissue boundaries due to poor contrast (Al-qazzaz, 2020; Baid et al., 2020). Moreover, different imaging modalities such as CT scans or Magnetic Resonance Imaging (MRI) pro- vide different types of information about a tumor\u0026rsquo;s location within the brain which requires specialized algorithms tailored specifically for each type. This has led researchers towards developing more sophisticated techniques such as deep learning models that incorporate prior knowledge into their predictions, allowing radiologists to speed up their workflow while achieving consistent results across multiple cases (Liu et al., 2022).\u003c/p\u003e\n"},{"header":"2. Related Works","content":"\u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe early detection and treatment of a brain tumor are essential because it is a potentially fatal condition. In the past, traditional machine learning methods that relied on hand-crafted feature engineering were used to acquire representations from brain images. However, these methods are not sufficient for accurate brain tumor segmentation (BTS) due to their limited ability in capturing local information of the tumor regions (Raza et al., \u003cspan class=\"CitationRef\"\u003e2023\u003c/span\u003e). Recent studies show that there are currently three different types of brain tumor segmentation algorithms: classical approaches, machine learning approaches (excluding deep learning approaches), and deep learning approaches (Tie et al., \u003cspan class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eThe most popular traditional approaches are the edge detection method (Caselles et al., \u003cspan class=\"CitationRef\"\u003e1993\u003c/span\u003e; Muthukrishnan and Radha, \u003cspan class=\"CitationRef\"\u003e2012\u003c/span\u003e), threshold method (Prastawa et al., \u003cspan class=\"CitationRef\"\u003e2004\u003c/span\u003e; Stadlbauer et al., \u003cspan class=\"CitationRef\"\u003e2004\u003c/span\u003e), region growth method (Gibbs et al., \u003cspan class=\"CitationRef\"\u003e1996\u003c/span\u003e; Weglinski and Fabijanska, \u003cspan class=\"CitationRef\"\u003e2011\u003c/span\u003e), watershed approach (Lin et al., \u003cspan class=\"CitationRef\"\u003e2006\u003c/span\u003e; Maiti and Chakraborty, \u003cspan class=\"CitationRef\"\u003e2013\u003c/span\u003e), and level set approach (Menze et al., 2010; Prastawa et al., \u003cspan class=\"CitationRef\"\u003e2003\u003c/span\u003e). These approaches are rapid and inexpensive to deploy, but they have limits due to their inability to generate good results in difficult cases, such as large or irregularly shaped tumors. Edge detection works by extracting boundary information from images while threshold in- volves setting a certain intensity value that divides objects from background pixels; both have limited accuracy when dealing with noisy or blurry images. Region growing begins at a point within the object and expands outward until it reaches the boundaries defined by user input parameters; however, this technique does not always guarantee accurate results because it may miss small regions within an image or include parts outside its borders due to noise levels in the image data. Watersheds use gradient magnitude maps in combination with markers manually placed at the center of each object. Level sets employ partial differential equations that evolve contours over time according to how they interact with edges detected using gradient oper- ators. Both require more computational power than simpler techniques like edge/thresholding yet still lack robustness when faced against challenging scenarios where multiple objects exist close together sharing similar intensi- ties across their boundaries making them difficult distinguishable from each other.\u003c/p\u003e\n \u003cp\u003eMachine learning methods are increasingly being utilized for the purpose of object classification and detection. These machine learning algorithms use training data to autonomously learn the features of a given dataset, allowing them to accurately classify objects within it. This has been especially useful in medical imaging applications such as brain tumor segmentation, where re- searchers have used machine learning methods like support vector machines (SVM) (Bauer et al., \u003cspan class=\"CitationRef\"\u003e2011\u003c/span\u003e; Ruan et al., \u003cspan class=\"CitationRef\"\u003e2007\u003c/span\u003e), clustering-based approaches (Stupp et al., \u003cspan class=\"CitationRef\"\u003e2017\u003c/span\u003e; Narayanan et al., \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e), conditional random fields (CRF) (Meier et al., \u003cspan class=\"CitationRef\"\u003e2014\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e2015\u003c/span\u003e) and random forest techniques (RF) (Tustison et al., \u003cspan class=\"CitationRef\"\u003e2015\u003c/span\u003e; Breiman, \u003cspan class=\"CitationRef\"\u003e2001\u003c/span\u003e). Bauer et al (Bauer et al., \u003cspan class=\"CitationRef\"\u003e2011\u003c/span\u003e), for example, devel- oped a classification method based on hierarchical SVM in conjunction with a CRF-based technique to segment a brain tumor from 3D MRI images. And Ruan et al. (Ruan et al., \u003cspan class=\"CitationRef\"\u003e2007\u003c/span\u003e) employed support vector machines (SVMs) to detect the abnormal region and then used multi-scales to extract the brain tumor field. Spatially Constrained Fish School Optimization (SCFSO) and Interval type-II Fuzzy Logic System (IT2FLS) algorithms were proposed by Stupp et al. (Stupp et al., \u003cspan class=\"CitationRef\"\u003e2017\u003c/span\u003e) for segmenting tumor and non-tumor re- gions in MR brain images. The segmentation results obtained by applying\u003c/p\u003e\n \u003cp\u003eIT2FLS to the subsequent clustering assignment following the use of SCFOS were impressive, taking less processing time while delivering good results.\u003c/p\u003e\n \u003cp\u003eFor segmenting brain tumors in multi-modal MR images, Narayanan et al. (Narayanan et al., \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e) combined particle swarm optimization (PSO) and bacterial foraging optimization (BFO) with an upgraded clustering method based on fuzzy c means. Whereas Tustison et al. (Tustison et al., \u003cspan class=\"CitationRef\"\u003e2015\u003c/span\u003e) suggested a supervised brain tumor segmentation approach based on ran- dom forest (RF) and Markov random field that produced good performance when applied to multimodal MRI scans, Meier et al. (Meier et al., \u003cspan class=\"CitationRef\"\u003e2015\u003c/span\u003e) em- ployed decision forest for classification and conditional random field (CRF) for spatial regularization in brain tumor segmentation. The main benefit of using these types of machine learning algorithms is that they require mini- mal human intervention; once trained on a given dataset they can be applied quickly and efficiently without requiring additional input from operators or clinicians. Furthermore, their accuracy tends to improve over time as more training data becomes an important factor when dealing with rare diseases or conditions which may not have sufficient datasets readily available for anal- ysis at first instance.\u003c/p\u003e\n \u003cp\u003eIn recent years, deep learning which is a particular kind of machine learn- ing has made remarkable achievements in many fields, such as pattern recog- nition, image segmentation, image detection and image classification (Zhang et al., \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e2021\u003c/span\u003e; Xiang et al., \u003cspan class=\"CitationRef\"\u003e2021\u003c/span\u003e). It has reentered the spotlight ever since AlexNet won the ImageNet competition in 2012 (Krizhevsky et al., \u003cspan class=\"CitationRef\"\u003e2012\u003c/span\u003e; Deng et al., \u003cspan class=\"CitationRef\"\u003e2009\u003c/span\u003e). Many researchers have applied deep learning meth- ods for medical imaging applications such as the automatic segmentation of brain tumors from MRI scans. Deep Learning models have been shown to outperform traditional machine learning algorithms when it comes to ac- curacy rate for automated BTS segmentation tasks with minimal human intervention required during the training phase itself.\u003c/p\u003e\n \u003cp\u003eThe early application of deep learning in brain tumor segmentation is mainly to modify the convolution network. It is able to extract features from volumetric scans using convolutional neural networks (CNNs). CNNs can capture both global and local information about the scanned image, allow- ing them to accurately detect subtle changes in tissue structure. Convolu- tional neural networks (CNNs) have revolutionized the field of deep learning and made tremendous progress in areas such as computer vision, natural language processing, and robotics. CNNs are a type of artificial neural net- work for automatically learning high-dimensional hierarchical features that use convolution operations to extract features from an input image or video. This has enabled researchers to create several specialized deep learning mod- els such as ResNet (He et al., \u003cspan class=\"CitationRef\"\u003e2016\u003c/span\u003e), VGG-Net (Szegedy et al., \u003cspan class=\"CitationRef\"\u003e2015\u003c/span\u003e), and Inception networks (Szegedy et al., \u003cspan class=\"CitationRef\"\u003e2017\u003c/span\u003e) for various tasks. These architec- tures can be used together with open-source frameworks like TensorFlow and PyTorch which makes development easier in academic fields like computer vision where semantic segmentation is used to classify objects within im- ages or videos accurately. Fully convolutional networks (FCN) enable CNNs to label each pixel using a straightforward upsampling technique which has proven invaluable when it comes to recognizing patterns at multiple resolu- tions without needing additional training data for different scales or sizes of inputs (Long et al., \u003cspan class=\"CitationRef\"\u003e2015\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eBecause of the rapid progress of machine and deep learning techniques, we are now able to generate 3D volumetric data from 2D brain imaging slices in a timely, accurate, and cost-effective manner. This enables the adoption of more complicated models like U-Net, VNet, and other 3D models that can extract a greater number of features than typical 2D models. U-Net is one of the most widely used deep learning architectures for biomedical image seg- mentation. It was first introduced by Ranneberger et al. (Ronneberger et al., \u003cspan class=\"CitationRef\"\u003e2015\u003c/span\u003e) and has since gained popularity because of its ability to deliver good segmentation results despite limited training data. The U-Net architecture is a powerful tool for segmentation tasks, particularly in the medical field. Its left contracting (encoder) part is responsible for feature extraction while its right expanding (decoder) part produces an output and maps it back into raw image pixels. Skip connections are essential to this model, as they allow the extracted features from the encoder block to be passed on to the decoder block so that tumors can be detected accurately. In recent years, different studies have used combinations of skip connections with models such as U- Net++ (Zongwei et al., \u003cspan class=\"CitationRef\"\u003e2018\u003c/span\u003e)d Net3+ (Huimin et al., \u003cspan class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eFurthermore, attention models (Prajit et al., \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e) have also been ac- tively used in vision tasks due to their potential accuracy increase when compared with traditional methods. Attention models filter out irrelevant features from input images which save computational power and improve performance; thus making them a popular choice among researchers using U-Nets in medical applications. Several studies have been undertaken in the medical industry, particularly with U-Net, to boost the accuracy of the model with attention models filtering away unimportant elements from the input images to conserve computational power and accelerate the model. For instance, the authors of (Oktay et al., \u003cspan class=\"CitationRef\"\u003e2018\u003c/span\u003e) developed a unique attention gating approach based on the U-Net (Olaf et al., \u003cspan class=\"CitationRef\"\u003e2015\u003c/span\u003e) architecture. In their research effort, the experts found that the attention module suppresses the unnecessary areas of the input image, while accentuating the salient features. They added that an attention module-enhanced model can operate with lit- tle additional computational burden while improving the model\u0026rsquo;s sensitivity and predictive power.\u003c/p\u003e\n \u003cp\u003eThe use of 2D or 3D convolutions during the training of a deep CNN model is essential for accurate and effective segmentation. U-Net, Res-Net, and DenseNets (Simon et al., \u003cspan class=\"CitationRef\"\u003e2017\u003c/span\u003e) were previously proposed, but they only utilized two-dimensional input images thus missing out on additional con- textual information between slices. 2D convolutional networks are limited in their ability to capture spatial information from medical images, while 3D convolutions require more memory and computational resources. To address this issue, W. Chen et al (Chen et al., \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e) proposed a separable 3D U-Net model that uses separate 2D and 1D layers to learn both spatial features (2D) as well as temporal features (1D). This approach allows for full utiliza- tion of all three dimensions within the brain volume without overburdening memory requirements or taking up too much computing power. In order to train their proposed approach for each orthogonal view (axial, sagittal, and coronal) individually, the authors added separable temporal convolution to the residual inception model. For improved performance, they also used a multi-view fusion approach to merge all of the convolutional outputs. In terms of efficiency, the suggested model performed well on the BraTS 2018 test dataset. For patients with brain cancer or other illnesses that generate tumor growths inside the skull cavity, accurate brain image segmentation models are crucial for early tumor detection and enhancing survival rates. Local and global characteristics are vital for making judgments regarding tu- mor identification; however, gradients of low-level features approach zero as we move deeper into this process, making it impossible to identify borders or edges effectively without utilizing more sophisticated methods such as deep learning techniques like those utilized by these researchers in their study.\u003c/p\u003e\n \u003cp\u003eThe residual ResNet with a residual block was proposed by He et al. (He et al., \u003cspan class=\"CitationRef\"\u003e2016\u003c/span\u003e) to simplify neural network optimization and address the degra- dation problem in deep learning architectures. This type of network has been widely applied for image classification, recognition and segmentation tasks. Myronenko et al. (Myronenko, \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e) further improved this architec- ture by adding a variational auto-encoder (VAE) branch based on 3D U-Net while replacing the original U-Net convolutional blocks with residual blocks. Huang et al. (Huang et al., \u003cspan class=\"CitationRef\"\u003e2017\u003c/span\u003e) proposed Dense Convolutional Networks (DenseNets), which are designed such that all layers are directly connected while ensuring maximum information transmission between them; thereby allowing more effective feature propagation, alleviating vanishing gradients issues, and reducing parameter numbers at the same time. Subsequently, many researchers have combined DenseNets with U-nets for medical image segmentation tasks as well due to their effectiveness in these areas of appli- cation.\u003c/p\u003e\n \u003cp\u003eMDU-Net (Zhang et al., \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e) is an improved version of U-net that uses three different multi-scale dense connections in the encoder, decoder and con- nection between them. This type of network was first proposed by Wang et al. (Wang et al., \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e) to segment retinal vessels and achieved good results. Building on this success, Ziang et al.(Ziang et al., \u003cspan class=\"CitationRef\"\u003e2020\u003c/span\u003e) combined Inception Residuals with dense connections to further improve the performance of U- net for medical image segmentation tasks such as tumor detection or brain tissue classification.\u003c/p\u003e\n \u003cp\u003eBy combining the advantages of DenseNet and ResNet, Tie et al. (Tie et al., \u003cspan class=\"CitationRef\"\u003e2021\u003c/span\u003e) proposed a new 3D U-Net with dense encoder blocks and resid- ual decoder blocks for segmenting the brain tumor multi-class 3D MR im- ages and achieved better results. Raza et al. (Raza et al., \u003cspan class=\"CitationRef\"\u003e2023\u003c/span\u003e) proposed a combined model of deep residual network and U-Net model (dResU-Net) to make predictions using both low-level and high-level information at the same time. Moreover, shortcut connections between residual networks are used to maintain low-level characteristics at each level and to enhance the training in the brain tumor segmentation process. According to results that were obtained by comparing the works, dResU-Net has greatly enhanced the segmentation performance of brain tumor sub-regions.\u003c/p\u003e\n \u003cp\u003eDue to the enormous variability of MRI brain tumor images and the small proportion of brain tumor, segmentation is an extremely tough task. Although deep learning has produced significant improvement and impres- sive results in segmentation with top results of dice scores 0.8660, 0.8357, and 0.8004 for whole tumor, tumor core, and enhancing tumor core respectively on the BraTS2020 validation data set, there is no doubt that this segmen- tation accuracy could not be used for clinical diagnosis. MRI brain tumor segmentation still has to improve a lot until completely addressed. This pa- per proposes a new network architecture based on 3D Attention U-Net to solve the accuracy problem of MRI brain tumor segmentation.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"3. Methodology","content":"\u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThis section discusses the dataset\u0026rsquo;s details, preprocessing procedures, and the proposed method\u0026rsquo;s implementation in detail. The proposed 3D Attention U-Net architecture is also covered in this part, along with detail on the loss functions that were used during the training stage.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv class=\"Section2\" id=\"Sec4\"\u003e\n \u003ch2\u003e3.1. Dataset\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe dataset that we used to train our model is taken from \u0026ldquo;Medical Im- age Computing and Computer-Assisted Intervention (MICCAI) Multimodal Brain Tumor Segmentation Challenge (BraTS) 2020\u0026rdquo; which is publicly avail- able in BraTS 2020 challenge dataset. It was gathered by medical experts from the \u0026rdquo;University of Pennsylvania and UPenn\u0026rsquo;s Center for Biomedical Image Computing and Analysis (CBICA)\u0026rdquo; (Menze, \u003cspan class=\"CitationRef\"\u003e2015\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eIt is divided into training and validation datasets which are manually annotated by neuroradiologists. It includes 3D MRI brain scans from 369 patients with gliomas, 76 of them had LGG, while the rest images are from patients with HGG. Each 3D scan consists of 155 slices, with each image measuring 240 x 240 dimensions. There are four main types of labels used to describe the different areas of a brain tumor including Label 0, Label 1, Label 2, and Label 4. Label 0 is reserved for non-tumor areas such as background or normal tissue area while Labels 1 and 4 refer to neurotic (NCR) or non- enhancing core (NET), edema (ED), respectively as well as enhancing tumor cores (ET). The regions labeled greater than 0 are referred to collectively as \u0026ldquo;whole tumor\u0026rdquo; regions with those labeled specifically 1 and 4 is called \u0026ldquo;tumor core\u0026rdquo; regions containing the ET region.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv class=\"Section2\" id=\"Sec5\"\u003e\n \u003ch2\u003e3.2. Pre-Processing\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe BraTS 2020 dataset was preprocessed by the competition\u0026rsquo;s organiz- ers before it was made publicly available. The images were all co-registered to the same anatomical template, interpolated to the same resolution (1 mm\u003csup\u003e3\u003c/sup\u003e), and have isotropic resolution. Despite the fact that all MRI images have been pre-processed, the brightness of the images vary substantially. As a result, we normalized each 3D MR modality image using a Min-Max normalizer to re- move the impact of image intensity change on prediction accuracy. By doing so, all features will be converted into a range [0,1], with 0 and 1 serving as the minimum and maximum values for each feature/variable. It is calculated by the following formula:\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv class=\"Equation\" id=\"Equ1\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e$${x}_{\\text{scaled }}=\\frac{x-\\text{m}\\text{i}\\text{n}\\left(x\\right)}{\\text{m}\\text{a}\\text{x}\\left(x\\right)-\\text{m}\\text{i}\\text{n}\\left(x\\right)}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003ewhere x\u003csub\u003escaled\u003c/sub\u003e is the new value of each entry in the data and x is the old value of each entry in data. max(x) and min(x) are the maximum and mini- mum values of the features respectively. Due to memory constraints, all MR images and segmentation masks were reduced to 120 x 120 x 96 dimensions from 240 x 240 x 155.\u003c/p\u003e\n \u003cp\u003eNearest-neighbor interpolation was used to resize both the images and the masks. When resizing segmentation masks, it is necessary to apply \u0026rsquo;nearest neighbor\u0026rsquo; interpolation. This is because other techniques, such as Bilinear interpolation, consider the closest 2x2 neighborhood of known pixel values surrounding a pixel. It then takes a weighted average of those pixels to get the final interpolated value, which can be a float value. This creates a smoother image than nearest-neighbor interpolation, but it results in full black images for some slices having segmentation masks.\u003c/p\u003e\n \u003cp\u003eNearest-neighbor does not really interpolate the pixels. Instead, it looks for the closest one in the source image that matches the location of the target image. It\u0026rsquo;s the fastest interpolation method, and requires the least amount of processing time, and allows for sharper detail. The biggest drawback, how- ever, is that the resulting image may contain jagged edges. But it is suitable for this task as it wont affect the segmentation masks. All four modalities were stacked together to take advantage of the entire information given in four sequences. At the training time, training examples with dimensions of 120x120x96x4 (where 4 denotes the four modalities, i.e., T1, T1ce, T2, FLAIR) were provided to the model as input.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv class=\"Section2\" id=\"Sec6\"\u003e\n \u003ch2\u003e3.3. Network Architecture\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eDue to its local and global feature extraction technique, U-Net is a promi- nent convolutional neural network used for biomedical semantic image seg- mentation. It is made up of two paths: an encoder path that extracts features from the input image and a decoder path that combines these extracted fea- tures to form the output mask. It employs skip connections between each level of encoders and their associated decoders, allowing the classifier to con- sider both low-level and high-level features when determining the segmen- tation mask. However, various constraints impede U-Net training, such as a lack of data augmentation techniques or a restricted number of available layers. It is also challenging for real-time application scenarios where speed is crucial because it needs a substantial amount of computer resources for training. The gradient vanishing problem is a limitation of a U-Net network that can lead to inefficient learning at early layers. Additionally, a large num- ber of parameters present in the model can lead to overfitting if not trained properly. Low-level features are essential for segmentation masks, as they contain information about boundaries, edges, and location of tumorous re- gions. As the network goes deeper during down-sampling operations, richer higher-level details become available than lower-level ones, but local detail and location data are lost due to convolutional and nonlinearity operations.\u003c/p\u003e\n \u003cp\u003eThis study introduced a dense-encoder and residual-decoder based atten- tion 3D U-Net model for the task of BTS to address the problem of vanishing gradient in the encoder section of the U-Net model during deep network train- ing. We used 3D U-Net as a baseline in our proposed architecture to develop a novel attention 3D U-Net with encoder-decoder blocks to segment brain tumors from MRI images. Our network architecture is depicted in Figure\u003c/p\u003e\n \u003cp\u003e3. The encoder section gathers context information from the input image, while the decoder section performs accurate segmentation. In Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e, the network encoder is presented on the left, and the network decoder is shown on the right. We fed four modalities 3D MR images (T1,T1Gd ,T2 and FLAIR) with sizes of 120 \u0026times; 120 \u0026times; 96. Thus, the input data is a 4D patch with a size of 120 \u0026times; 120 \u0026times; 96 \u0026times; 4. The first convolutional block consists of a 7x7 3D convolution followed by a group norm and ReLu; this produces an output feature map with size 120\u0026times;120\u0026times;96\u0026times;24. The feature map is the fed to a max pool layer followed by a double convolution layer with a kernel size of 3, which produces an output with a size of 60 \u0026times; 60 \u0026times; 48 \u0026times; 24. Finally, the feature map goes through four dense and transition layers.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eBecause each dense layer in a dense block has access to the feature maps of all dense layers preceding it, it has global information about the entire block, which is useful for accurate segmentation. The contribution of the output feature maps of each dense layer to the global information is con- trolled by the growth rate of the dense blocks. A dense layer is made up of one 1 \u0026times; 1 \u0026times; 1 convolution layer and one 3 \u0026times; 3 \u0026times; 3 convolution layer. Following each convolution layer, group normalization and the ReLU activation function are utilized. We first perform a 1x1 convolution with 128 filters to reduce the size of the feature maps, and then perform a more expensive 3x3 convolution with a growth rate of 32 feature maps which uses padding to ensure the dimensions remain constant. Concatenation is then used to com- bine the output from each layer in order to preserve the information from earlier layers.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe size of the feature vectors is reduced by using a transition block. As opposed to a pooling layer, Transition Block performs a 1x1 convolution with 32 filters, followed by another 1x1 convolution with a stride of 2, which reduces the size of the volume and the number of feature maps by half. In order to reduce the information loss associated with pooling, convolution should be performed with a stride of 2. Each convolution is then followed by a group norm and ReLU layers.\u003c/p\u003e\n \u003cp\u003eA trilinear interpolation layer is used for upsampling, which is subse- quently concatenated with low-level features from the corresponding dense block. The output is then passed through a spatial and channel attention block first. Our proposed attention module combines channel, spatial, and skip connection in parallel. However, combining excited features at the same time runs the danger of making feature learning inconsistent. As demon- strated in Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e, we integrate skip connections to reduce network redun- dancy and sparsity.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eSimilar to (Islam et al., \u003cspan class=\"CitationRef\"\u003e2020\u003c/span\u003e) by utilizing 3D inter-spatial and inter- channel feature relationships, we introduce 3D attention units to generate 3D spatial and channel attention. We initially carry out a 1x1xC convo- lution to group all spatial feature correlations into the HxWx1 dimension, after which we carry out average pooling and feed it to the neural network to produce the 3D attention map. We combine skip-connection to reduce the singularity and sparsity brought on by these parallel excitations, improving segmentation prediction and allowing for more generalized learning.\u003c/p\u003e\n \u003cp\u003eThe structure of the residual block is shown in Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e, which is com- prised of two 3 \u0026times; 3 \u0026times; 3 convolution layers with stride of 1. Group normaliza- tion and ReLU activation functions are used after to each convolution layer.\u003c/p\u003e\n \u003cp\u003eThe output of the residual block is the addition of the output of the last convolution layer and the input of the residual block. The shortcut connec- tion is first passed through an additional 1 \u0026times; 1 \u0026times; 1 convolution layer with stride of 1. From the top to the bottom of the encoder, the number of feature\u003c/p\u003e\n \u003cp\u003emap increases and the size of feature map reduces progressively. Specifically, Group Norm layer (GN) and activation layer were applied after each convo- lution layer. The Group Norm aids in creating deeper and wider networks. It first groups the channels and normalize inside each group. It is a trade-off between LayerNorm and InstanceNorm. When batch sizes are small, group normalization outperforms batch normalization. The batch size in our work is one. As a result, we applied group normalization across the entire net- work. However, when the batch size is significantly large, GN does not scale as well as BN and might not be able to match the performance of BN. The activation layer implements a rectified linear unit (ReLU) layer. The ReLU is employed to optimize the proposed architecture and consequently enhance the performance.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv class=\"Section2\" id=\"Sec7\"\u003e\n \u003ch2\u003e3.4. Losses and Evaluation metrics\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe loss and evaluation metrics utilized to evaluate the performance of the proposed model are explored in detail in this section. By minimizing the loss function, the proposed model learns parameters/features from MR images. In the BraTS datasets, about 98 percent of total voxels are non- tumor (Label 0, commonly known as background), with only a few voxels belonging to tumor (Labels 2, 3, 4). We used a BCE-Dice loss function-based region to overcome the class imbalance problem. This loss combines the Dice loss with the standard binary cross-entropy (BCE) loss, which is commonly used in segmentation models. Mixing the two approaches allows for some loss variability while benefiting from BCE\u0026rsquo;s stability. The equation for multi-class BCE by itself will be familiar to anyone who has studied logistic regression:\u003c/p\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eAs evaluation metrics, we used the IoU metric, also known as the Jaccard Index, and the Dice coefficient, also known as the Dice-S\u0026oslash;rensen coefficient. The Dice score, which calculates the discrepancies between the segmentation result and the ground truth, is a widely used metric for pixel segmentation. The dice score can be calculated as follows for a given predicted and ground truth segmentation X and Y, where each tumor voxel is labeled as 1, and non-tumor voxels are labeled as 0.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv class=\"Equation\" id=\"Equ2\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e$$DSC(X,Y)=\\frac{2|X\\cap Y|}{\\left|X\\right|+\\left|Y\\right|}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eThe Jaccard metric is similar to the Dice metric in that it calculates the ratio between the overlap of positive instances between two sets by evaluating the performance of pixel segmentation models and their mutual combined values:\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv class=\"Equation\" id=\"Equ3\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e$$J(A,B)=\\frac{|A\\cap B|}{|A\\cup B|}=\\frac{|A\\cap B|}{\\left|A\\right|+\\left|B\\right|-|A\\cap B|}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"4. Experimental Results And Analysis","content":"\u003cp\u003e\u003cem\u003e4.1 Implementation details\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe Python programming language and PyTorch as its backend were used to implement the suggested model. Due to constrained memory and compu- tational resources, the model was trained for 100 epochs with a batch size of\u003c/p\u003e\n\u003cp\u003e1. Lion (EvoLved Sign Momentum) optimizer with a learning rate of 0.0001 and weight decay of 0.01 was used to replace Adam. It is a new optimizer discovered by Google Brain that is purportedly to be better than Adam(w), in Pytorch. Each layer of the network was normalized and made more sta- ble by using the activation function ReLU with group normalization. The studies were carried out using the BraTS 2020 benchmark dataset, with 201 images used for training, 34 images for validation, and 134 images for testing purposes. The detailed information about the hyperparameter values that were provided during the training of the model is given in Table 1.\u003c/p\u003e\n\u003cp\u003eTable 1: List of hyperparameters used for 3D Attention UNet\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cp\u003e3D Attention U-Net clearly outperforms other approaches when it comes to model complexity. It took around 11.8 hours to train the proposed model through 100 epochs, with an execution time of 7.12 minutes per epoch. It can be deployed in a real-world clinical setting because our suggested model\u0026rsquo;s online test time for one person was less than one second. By leveraging more effective computational resources, the model\u0026rsquo;s efficiency may be improved even more. Detailed information about the hardware requirement and com- putational efficiency of the proposed model is given in Table 2.\u003c/p\u003e\n\u003cp\u003eTable 2: Hardware requirement and computational efficiency of the proposed model\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003e4.2 Results and discussion\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eFigure 6 shows the bar plot of dice and jaccard score obtained by our network on BraTS2020 validation dataset. Table 3 lists the value of dice and jaccard of ET, WT, and TC in the three regions of our segmentation results. The average value, standard deviation, median value, 25 percent quantile value, and 75 percent quantile value are given for each indicator value. It can be seen from Table 3 that the dice mean of WT, TC, and ET are 0.889, 0.866, and 0.828, respectively, and the Jaccard mean of ET, WT, and TC are 0.718, 0.809, and 0.779, respectively. The WT region has the highest prediction accuracy in our method, whereas the ET region has the lowest. This is due to the fact that the WT region is the largest and the ET region is the smallest. The greater the region, the easier it is to predict. And the smaller the region, the more difficult it is to predict.\u003c/p\u003e\n\u003cp\u003eTable 3: The performance of our proposed method on BraTS 2020 validation dataset.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cp\u003eThe performance of our method is compared with the original 3D U-Net method. Figure 7 and Table 4 list the dice and jaccard score obtained by 3D U-Net on the BraTS2020 validation dataset for each class. The proposed technique achieved a higher dice score for each sub-region of the brain tumor than the baseline work because of the use of the dense block with a tran- sition at the decoder part and residual convolutional blocks at the encoder part. The proposed Attention 3D U-Net model significantly improves the segmentation accuracy of the ET class which was the most difficult class to segment.\u003c/p\u003e\n\u003cp\u003eTable 4: The performance of original 3D U-Net on BraTS 2020 validation dataset.\u003c/p\u003e\n\u003ctable border=\"0\" cellpadding=\"0\" cellspacing=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"21.717171717171716%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e\u003cstrong\u003eDice\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.161616161616163%\"\u003e\n \u003cp\u003e\u003cstrong\u003eJaccard\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"21.717171717171716%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003eWT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003eTC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003eET\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003eWT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.161616161616163%\"\u003e\n \u003cp\u003eTC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003eET\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"21.717171717171716%\"\u003e\n \u003cp\u003eMean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.885\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.793\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.761\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.161616161616163%\"\u003e\n \u003cp\u003e0.682\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.646\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"21.717171717171716%\"\u003e\n \u003cp\u003eStdDev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.169\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.206\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.107\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.161616161616163%\"\u003e\n \u003cp\u003e0.189\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.202\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"21.717171717171716%\"\u003e\n \u003cp\u003eMedian\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.895\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.849\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.815\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.161616161616163%\"\u003e\n \u003cp\u003e0.738\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.687\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"21.717171717171716%\"\u003e\n \u003cp\u003e25quantile\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.847\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.762\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.763\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.735\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.161616161616163%\"\u003e\n \u003cp\u003e0.616\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.616\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"21.717171717171716%\"\u003e\n \u003cp\u003e75quantile\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.896\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"13.131313131313131%\"\u003e\n \u003cp\u003e0.872\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.887\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.161616161616163%\"\u003e\n \u003cp\u003e0.811\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.772\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eOur method was qualitatively compared with original 3D U-Net. We randomly selected a patient and used both our method and original 3D U- Net to segment the brain tumor. Figure 8 shows slices of the four modalities of ground truth of the patient and the slices of the predicted segmentation image using our method and the ground truth.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e4.2 Comparison with state-of-the-art studies\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eIn comparison to state of art models for brain tumor semantic segmenta- tion, the proposed method is able to segment these individual tumor regions (WT, TC, and ET) close to the ground truth, as shown in the findings. The most challenging areas to segment were the enhancing tumor and its dis- persion with necrosis. When it comes to segmenting augmenting enhancing tumor and tumor core, many existing segmentation models performed badly. The proposed Attention 3D U-Net model, on the other hand, successfully seg- mented these regions. Table 5 also illustrates the quantitative performance of our model, which performs better than the state-of-the-art techniques for enhancing tumor and tumor core classes. In the case of segmenting the whole tumor, the proposed method achieves good results, comparable to state of art techniques. Overall, the proposed method seems to be a better technique to generate segmented images that are closest to the ground truth.\u003c/p\u003e\n\u003cp\u003eTable 5: Comparison of proposed 3D Attention UNet model with state-of-the-art methods\u003c/p\u003e\n\u003ctable border=\"0\" cellpadding=\"0\" cellspacing=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003e\u003cstrong\u003eReference\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e/Study\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e\u003cstrong\u003eImage\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eDimension\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003e\u003cstrong\u003eDataset\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e\u003cstrong\u003eDice Score (DSC )\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"top\" width=\"25.606060606060606%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e\u003cstrong\u003eTumor\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eCore\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e\u003cstrong\u003eWhole\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eTumor\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e\u003cstrong\u003eEnhancing\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eTumor\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e\u003cstrong\u003e(TC )\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e\u003cstrong\u003e(WT )\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e\u003cstrong\u003e(ET )\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eW. Wang et al. (Wang et al., 2021)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e128 x 128 x 128\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2020\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.8173\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9009\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.7873\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eO\u0026uml; .C\u0026cedil; i\u0026cedil;cek et al. (Cicek et al., 2016)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e128 x 128 x 128\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2020\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.7906\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.8411\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.6876\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eJ. Colman et al. (Colman et al., 2021)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e2DSlices with 240 x 240\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2020\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.7983\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.8673\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.7514\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eH. Messaoudi et al. (Messaoudi et al., 2021)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e192 x 160 x 108\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2020\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.7520\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.8068\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.6959\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eL. M. Ballestar et al. (Ballestar and Vilaplana, 2020)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e64 x 64 x 64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2020\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.7526\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.8463\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.6215\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eM. Ghaffari et al. (Ghaffari et al., 2021)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e128 x 128 x 128\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2020\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eF. Wang et al. (Wang et al., 2020)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e128 x 128 x 128\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.798\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.852\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.778\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eJ. Zhang et al. (Zhang et al., 2020)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e128 x 128 x 128\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.777\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.870\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.709\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eA. Myronenko (Myronenko, 2018)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e160 x 192 x 128\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2018\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.8154\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.8839\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.7664\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eYang et al. (Yang and Yang, 2018)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e96 x 96 x 96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2018\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.789\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.869\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.722\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eRaza et al. (Raza et al., 2023)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e128 x 128 x 128\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2020\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e0.8357\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.8660\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e0.8004\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"34.54545454545455%\"\u003e\n \u003cp\u003eProposed model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"16.363636363636363%\"\u003e\n \u003cp\u003e120 x 120 x 96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"8.787878787878787%\"\u003e\n \u003cp\u003eBraTS 2020\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.696969696969697%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.866\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"11.363636363636363%\"\u003e\n \u003cp\u003e0.889\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"14.242424242424242%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.828\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 6 shows the training time, prediction time, and number of parame- ters. With 29.5 MB million parameters, our model takes about 11.8 hours to train and 52.65 seconds to predict. This demonstrates that our model is substantially lower in size and trains faster while maintaining the same per- formance. Due to their computationally efficient designs, new transformer- based architectures improved inference times relative to other architectures, however they still fall short of our suggested model in terms of performance.\u003c/p\u003e\n\u003cp\u003eTable 6: Complexity of models used in comparison process.\u003c/p\u003e\n\u003ctable border=\"0\" cellpadding=\"0\" cellspacing=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e\u003cstrong\u003eTraining Time (h )\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrediction Time (s )\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e\u003cstrong\u003eParameters\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003eOur method\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e\u003cstrong\u003e11.8 hours\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e\u003cstrong\u003e52.65\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e29.5 mln\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003e3DU -Net\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e15.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e76.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e19mln\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003e3DRes -Net\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e24.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e65.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e46.4mln\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003eSwin BTS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e13.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e58.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e\u003cstrong\u003e16.3mln\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003eVT U -Net\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e15.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e59.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e18.4mln\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003e3DLI -Net ++\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e25.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e64.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e51.6mln\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003eAttention 3DUI -Net\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e18.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e121.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e22.3mln\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003e3DRes -U -Net\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e29.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e69.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e58.6mln .\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003e3DDeepLab V3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e35.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e78.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e76.1mln\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" width=\"26.255707762557076%\"\u003e\n \u003cp\u003e3DPSP -Net\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"27.168949771689498%\"\u003e\n \u003cp\u003e36.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"29.22374429223744%\"\u003e\n \u003cp\u003e96.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" width=\"17.35159817351598%\"\u003e\n \u003cp\u003e72.4.mln\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eBy combining the advantages of ResNet and DenseNet, this paper pro- poses a deep 3D Attention UNet with dense encoder blocks and residual de- coder blocks to increase the performance of brain tumor segmentation from MRI images. The use of dense blocks has the advantage of reducing network parameters, deepening network layers, strengthening feature propagation, avoiding the vanishing-gradient problem, and enlarging receptive fields. In comparison to typical convolution blocks, residual blocks address the issue of network performance degradation as network depth increases. We also in- troduced 3D attention units to produce 3D spatial and channel attention by leveraging 3D inter-spatial and inter-channel feature correlations. The proposed method surpassed state-of-the-art models, yielding dice scores of 0.866, 0.889, and 0.828 for TC, WT, and ET, respectively, without the use of substantial post-processing. Furthermore, our model takes approximately\u003c/p\u003e \u003cp\u003e11.8 hours to train and 52.65 seconds to predict with only 29.5 MB million parameters. This shows that while maintaining the same performance, our model is significantly smaller and trains faster than other models. By using this approach in clinical oncology, radiologists and oncologists can better un- derstand the different tumor locations, sizes, and shapes, which will aid them with tumor diagnosis, treatment planning, and prognosis. More 3D-based ar- chitectures should be investigated in order to minimize computational costs while maximizing contextual information. Post-processing approaches can be utilized to improve the performance of the model, and this research can be applied to clinically challenging medical imaging problems and other seg- mentation applications.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAcknowledgement\u003c/p\u003e\n\u003cp\u003eWe want to thank Northwestern Polytechinical University for helping us to conduct research in this area and providing resource.\u003c/p\u003e\n\u003cp\u003eCompeting interests: The authors declare no competing interests.\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003e\u003cspan\u003eAl-qazzaz, S.A.L.I.R., 2020. Deep learning-based brain tumour image seg- mentation and its extension to stroke lesion segmentation .\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eAlagarsamy, S., Zhang, Y.D., Govindaraj, V., Murugan, P.R., Sankaran, S., 2020. Smart identification of topographically variant anomalies in brain\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003emagnetic resonance imaging using a fish school based fuzzy clustering ap- proach. IEEE Transactions on Fuzzy Systems.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eAzhari, E.E.M., Hatta, M.M., Htike, Z.Z., Win, S.L., 2014. Tumor detection in medical imaging: a survey. Int. J. Adv. Inf. Technol. 4, 21\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eBaid, U., Talbar, S., Rane, S., Gupta, S., Thakur, M., Moiyadi, A.and Sable, N.A.M., Mahajan, A., 2020. A novel approach for fully automatic intra- tumor segmentation with 3d u-net architecture for gliomas. Front. Comput. Neurosci. 14.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eBallestar, L.M., Vilaplana, V., 2020. Brain tumor segmentation using 3d- cnns with uncertainty estimation. ArXiv, 1\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eBauer, S., Nolte, L.P., Reyes, M., 2011. Fully automatic segmentation of brain tumor images using support vector machine classification in combi- nation with hierarchical conditional random field regularization. Medical image computing and computer-assisted intervention 6893, 354\u0026ndash;361.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eBauer, S., Wiest, R., Nolte, L.P., Reyes, M., 2013. A survey of mri-based medical image analysis for brain tumor studies. Phys. Med. Biol. 58, R97\u0026ndash;R129.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eBreiman, L., 2001. Random forests. Machine Learning 45(1), 5\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eCaselles, V., Catt\u0026acute;e, F., Coll, T., Dibos, F., 1993. A geometric model for active contours in image processing. Numerische Mathematik 66(1), 1\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eChen, W., Liu, B., Peng, S., Sun, J., Qiao, X., 2019. S3d-unet: Separable 3d u-net for brain tumor segmentation, in: In: Crimi, A., Bakas, S., Kuijf, H., Keyvan, F., Reyes, M., van Walsum, T. (eds) Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. BrainLes 2018. Lecture Notes in Computer Science(.), Springer, Cham.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eCicek, O., Abdulkadir, A., Lienkamp, S., Brox, T., Ronneberger, O., 2016. 3d u-net: Learning dense volumetric segmentation from sparse annotation. Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics) 9901 LNCS, 424\u0026ndash;432.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eColman, J., Zhang, L., Duan, W., Ye, X., 2021. Dr-unet104 for multimodal mri brain tumor segmentation. Lect. Notes Comput. Sci. (including Sub- ser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics) 12659 LNCS, 410\u0026ndash;419.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eCrimi, A., Bakas, S., Kuijf, H., Keyvan, F., Reyes, et al., M., 2019. Brainle- sion: Glioma, multiple sclerosis, stroke and traumatic brain injuries. 4th International Workshop, BrainLes 2018 11384.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eDeng, J., Dong, W., Socher, R., Li, L., Li, K., Fei-Fei, L., 2009. Imagenet: A large-scale hierarchical image database, in: In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p.\u0026nbsp;248\u0026ndash;255.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eEllison, J., 2020. Improving the generalizability of convolutional neural net- works for brain tumor segmentation in the post-treatment setting .\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eGhaffari, M., Sowmya, A., Oliver, R., 2020. Automated brain tumor segmen- tation using multimodal brain scans: A survey based on models submitted to the brats. 2012\u0026ndash;2018 challenges. IEEE Reviews in Biomedical Engineer- ing 13, 156\u0026ndash;168.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eGhaffari, M., Sowmya, A., Oliver, R., 2021. Automated brain tumour seg- mentation using cascaded 3d densely-connected u-net. Lect. Notes Com- put. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinfor- matics) 12658 LNCS, 481\u0026ndash;491.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eGibbs, P., Buckley, D.L., Blackband, S.J., Horsman, A., 1996. Tumour volume determination from mr images by morphological segmentation. Physics in Medicine Biology 41(11), 2437\u0026ndash;2446.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eHe, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition, in: In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p.\u0026nbsp;770\u0026ndash;778.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eHuang, G., Liu, Z., Maaten, L.V.D., Weinberger, K.Q., 2017. Densely con- nected convolutional networks, in: Computer Vision and Pattern Recog- nition (CVPR), p.\u0026nbsp;2261\u0026ndash;2269.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eHuimin, H., Lin, L., Tong, R., Hu, H., Zhang, Q., Iwamoto, Y., Han, X., Chen, Y.W., Wu, J., 2020. U-net 3+: A full-scale connected u-net for medical image segmentation, in: In Proceedings of the ICASSP 2020\u0026ndash;2020\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona, Spain, 4\u0026ndash;8 May, p.\u0026nbsp;1055\u0026ndash;1059.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eIslam, M., Vibashan, V.S., Jose, V.J.M., Wijethilake, N., Utkarsh, U., Ren, H., 2020. Brain tumor segmentation and survival prediction using 3d at- tention unet. In Brainlesion: Glioma, Multiple Sclerosis, Stroke and Trau- matic Brain Injuries: 5th International Workshop, BrainLes 2019, Held in Conjunction with MICCAI 2019, Shenzhen, China, October 17, 2019, Revised Selected Papers, Part I 5, 262\u0026ndash;272.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eKrizhevsky, A., Sutskever, I., Hinton, G., 2012. Imagenet classification with deep convolutional neural networks, in: In: Proceedings of the 25th International Conference on Neural Information Processing Systems, p.\u0026nbsp;1097\u0026ndash;1105.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eLin, Y.C., Tsai, Y.P., Hung, Y.P., Shih, Z.C., 2006. Comparison be- tween immersion-based and toboggan- based watershed image segmenta- tion. IEEE Transactions on Image Processing 15(3), 632\u0026ndash;640.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eLiu, Z., Tong, L., Chen, L., Jiang, Z., Zhou, F., Zhang, Q., Zhang, X., Jin, Y., Zhou, H., 2022. Deep learning based brain tumor segmentation: A survey. Complex Intelligent Systems 9, 1001\u0026ndash;1026.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eLong, J., Shelhamer, E., Darrell, T., 2015. Fully convolutional networks for semantic segmentation, in: In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p.\u0026nbsp;3431\u0026ndash;3440.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMaiti, I., Chakraborty, M., 2013. A new method for brain tumor segmen- tation based on watershed and edge detection algorithms in hsv colour model, in: 2012 National Conference on Computing and Communication Systems, p.\u0026nbsp;1\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMeier, R., Bauer, S., Slotboom, J., Wiest, R., Reyes, M., 2014. Appearance- and context-sensitive features for brain tumor segmentation, in: Proceed- ings of MICCAI BraTS Challenge 2014, p.\u0026nbsp;20\u0026ndash;26.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMeier, R., Karamitsou, V., Habegger, S., Wiest, R., Reyes, M., 2015. Param- eter learning for crf-based tissue segmentation of brain tumors. MICCAI Brainlesion Workshop 9556, 156\u0026ndash;167.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMenze, B.H., Leemput, K.V., Lashkari, D., Weber, M.A., Ayache, N.e.a., 2010. A generative model for brain tumor segmentation in multimodal im- ages. Medical image computing and computer-assisted intervention 6362, 151\u0026ndash;159.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMenze, B.H.e.a., 2015. The multimodal brain tumor image segmentation benchmark (brats). IEEE Transactions on Medical Imaging 34(10), 1993\u0026ndash;2024.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMessaoudi, H., Belaid, A., Allaoui, M.L., Zetout, A., Allili, M.S., Tliba, S., Conze, P.H., 2021. Efficient embedding network for 3d brain tumor segmentation. Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics) 12658 LNCS, 252\u0026ndash;262.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMuthukrishnan, R., Radha, M., 2012. Edge detection techniques for image segmentation. International Journal of Computer Science Information Technology 3(6), 250\u0026ndash;254.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMyronenko, A., 2018. 3d mri brain tumor segmentation using autoencoder regularization, in: In: International MICCAI Brainlesion Workshop, p.\u0026nbsp;311\u0026ndash;320.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMyronenko, A., 2019. 3d mri brain tumor segmentation using autoencoder regularization. Inter- national MICCAI Brainlesion Workshop, Cham, Switzerland, Springer 11384, 311\u0026ndash;320.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eNarayanan, A., Rajasekaran, M.P., Zhang, Y.D., Govindara, V., Thiyagara- jan, A., 2019. Multi-channeled mr brain image segmentation: A novel double optimization approach combined with clustering technique for tu- mor identification and tissue segmentation. Biocybernetics and Biomedical Engineering 39(2), 350\u0026ndash;381.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eOktay, O., Schlemper, J., Folgoc, L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N., Kainz, B.e.a., 2018. Attention u-net: Learning where to look for the pancreas. ArXiv abs/1804.03999.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eOlaf, R., Fischer, P., Brox, T., 2015. U-net: Convolutional networks for biomedical image segmentation, in: In International Conference on Med- ical Image Computing and Computer Assisted Intervention; Springer: Cham, Switzerland, p.\u0026nbsp;234\u0026ndash;241.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003ePrajit, R., Parmar, N., Vaswani, A., Bello, I., Levskaya, A., Shlens, J., 2019. Stand-alone self-attention in vision models, in: In Proceedings of the Advances in Neural Information Processing Systems 32, Vancouver, BC, Canada, p.\u0026nbsp;8\u0026ndash;14.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003ePrastawa, M., Bullitt, E., Ho, S., Gerig, G., 2004. A brain tumor segmenta- tion frame work based on outlier detection. Medical Image Analysis 8(3), 275\u0026ndash;283.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003ePrastawa, M., Bullitt, E., Moon, N., van Leemput, K., Gerig, G., 2003. Automatic brain tumor segmentation by subject specific modification of atlas priors 1. Academic Radiology 10(12), 1341\u0026ndash;1348.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eRaza, R., Ijaz Bajwa, U., Mehmood, Y., Waqas Anwar, M., Hassan Jamal, M., 2023. dresu-net: 3d deep residual u-net based brain tumor segmen- tation from multimodal mri. Biomedical Signal Processing and Control 79(P1), 103861.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eRonneberger, O., Fischer, P., Brox, T., 2015. Convolutional networks for biomedical image segmentation, in: In: International Conference on Med- ical image computing and computer assisted intervention, p.\u0026nbsp;234\u0026ndash;241.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eRuan, S., Lebonvallet, S., Merabet, A., Constans, J., 2007. Tumor segmen- tation from a multispectral mri images by using support vector machine classification, in: IEEE International Symposium on Biomedical Imaging: From Nano to Macro, p.\u0026nbsp;1236\u0026ndash;1239.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eSimon, J., Drozdzal, M., Vazquez, D., Romero, A., Bengio, Y., 2017. The one hundred layers tiramisu: Fully convolutional densenets for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, Honolulu, HI, USA 21\u0026ndash;26, 11\u0026ndash;19.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eStadlbauer, A., Moser, E., Gruber, S., Buslei, R., Nimsky, C.e.a., 2004. Improved delineation of brain tumors: An automated method for seg- mentation based on pathologic changes of 1h-mrsi metabolites in gliomas. Neuroimage 23(2), 454\u0026ndash;461.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eStupp, R., Taillibert, S., Kanner, A., Read, W., Ram, Z., 2017. Effect of tumor-treating fields plus main- tenance temozolomide vs maintenance temozolomide alone on survival in patients with glioblastoma: A random- ized clinical trial. JAMA 318(23), 2306\u0026ndash;2316.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eSzegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A., 2017. Inception-v4, inception- resnet and the impact of residual connections on learning, in: In: Thirty-First AAAI Conference on Artificial Intelligence, p.\u0026nbsp;4278\u0026ndash;4284.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eSzegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A., 2015. Going deeper with convolu- tions, in: In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, p.\u0026nbsp;1\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eTanneedi, R.V., Pedapati, P., Johansson, S., 2017. Brain tumour detection using hog by svm. Online.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eTie, J., Peng, H., Zhou, J., 2021. Mri brain tumor segmentation using 3d u-net with dense encoder blocks and residual decoder blocks. CMES - Computer Modeling in Engineering and Sciences 128(2), 427\u0026ndash;445.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eTustison, N., Shrinidhi, K.L., Wintermark, M., Durst, C.R., 2015. Opti- mal symmetric multimodal templates and concatenated random forests for supervised brain tumor segmentation (simplified) with antsr. Neuroin- formatics 13(2), 209\u0026ndash;225.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eWang, C., Zhao, Z., Ren, Q., Xu, Y., Yu, Y., 2019. Dense u-net based on patch-based learning for retinal vessel segmentation. Entropy 21(2), 168.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eWang, F., Jiang, R., Zheng, L., Meng, C., Biswal, B., 2020. 3d u-net based brain tumor segmentation and survival days prediction. 11992, 131\u0026ndash;141.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eWang, W., Chen, C., Ding, M., Yu, H., Zha, S., Li, J., 2021. Transbts: multi- modal brain tumor segmentation using transformer. Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformat- ics) 12901 LNCS, 109\u0026ndash;119.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eWeglinski, T., Fabijanska, A., 2011. Brain tumor segmentation from mri data sets using region growing approach, in: 2011 Proceedings of 7th In- ternational Conference on Perspective Technologies and Methods in MEMS Design, MEMSTECH2011, p.\u0026nbsp;185\u0026ndash;188.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eXiang, Y., Wang, S.H., Zhang, Y.D., 2021. Cgnet: A graph-knowledge em- bedded convolutional neural network for detection of pneumonia. Infor- mation Processing Management 58(1), 1\u0026ndash;25.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eYang, H., Yang, J., 2018. Automatic brain tumor segmentation with con- tour aware residual network and adversarial training, in: In: International MICCAI Brainlesion Workshop, Springer, Cham, p.\u0026nbsp;267\u0026ndash;278.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eZhang, Y.D., Dong, Z., Wang, S.H., Yu, X., Gorriz, J.M., 2020. Advances in multimodal data fusion in neuroimaging: Overview, challenges, and novel orientation. Information Fusion 64, 149\u0026ndash;187.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eZhang, Y.D., Dong, Z.C., Wu, L., Wang, S.H., 2011. A hybrid method for mri brain image classification. Expert Systems with Applications 38, 10049\u0026ndash;10053.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eZhang, Y.D., Govindaraj, V., Tang, C.S., Zhu, W.G., Sun, J.D., 2019. High performance multiple sclerosis classification by data augmentation and alexnet transfer learning model. Medical Imaging and Health Informat- ics 9(9), 2012\u0026ndash;2021.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eZhang, Y.D., Satapathy, S.C., Guttery, D.S., Gorriz, J., Wang, S.H., 2021. Improved breast cancer classification through combining graph convolu- tional network and convolutional neural network. Information Processing and Management 58(2), 1\u0026ndash;25.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eZiang, Z., Wu, C., Coleman, S., Kerr, D., 2020. Dense-inception u-net for medical image segmentation. Computer Methods andPrograms in Biomedicine 192, 1\u0026ndash;15.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eZongwei, Z., Siddiquee, M., Tajbakhsh, N., Liang, J., 2018. U-net++: A nested u-net architecture for medical image segmentation, in: In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support; Springer: Cham, Switzerland, p. 3\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Northwestern Polytechnical University","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Brain tumor segmentation, MRI images, Attention U-Net, Dense block, Residual block","lastPublishedDoi":"10.21203/rs.3.rs-2717573/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-2717573/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eMedical image segmentation is essential for disease diagnosis and for support- ing medical decision systems. Automatic segmentation of brain tumors from Magnetic Resonance Imaging (MRI) is crucial for treatment planning and timely diagnosis. Due to the enormous amount of data that MRI provides as well as the variability in the location and size of the tumor, automatic seg- mentation is a difficult process. Consequently, a current outstanding problem in the field of deep learning-based medical image analysis is the development of an accurate and trustworthy way to separate the tumorous region from healthy tissues. In this paper, we propose a novel 3D Attention U-Net with dense encoder blocks and residual decoder blocks, which combines the bene- fits of both DenseNet and ResNet. Dense blocks with transition layers help to strengthen feature propagation, reduce vanishing gradient, and increase the receptive field. Because each layer receives feature maps from all previous layers, the network can be made thinner and more compact. To make predic- tions, it considers both low-level and high-level features at the same time. In addition, shortcut connections between the residual network are used to pre- serve low-level features at each level. As part of the proposed architecture, skip connections between dense and residual blocks are utilized along with an attention layer to speed up the training process. The proposed architecture was trained and validated using BraTS 2020 dataset, it showed promising results with dice scores of 0.866, 0.889, and 0.828 for the tumor core (TC), whole tumor (WT), and enhancing tumor (ET), respectively. In compar- ison to the original 3D U-Net, our approach performs better. According to the findings of our experiment, our approach is a competitive automatic brain tumor segmentation method when compared to some state-of-the-art techniques.\u003c/p\u003e","manuscriptTitle":"Mutltimodal MRI Brain Tumor Segmentation using 3D Attention UNet with Dense Encoder Blocks and Residual Decoder Blocks","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2023-03-27 20:55:32","doi":"10.21203/rs.3.rs-2717573/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e727cd8e-da31-4c85-874d-a663b4f25f3d","owner":[],"postedDate":"March 27th, 2023","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":20142763,"name":"Artificial Intelligence and Machine Learning"},{"id":20142764,"name":"Nuclear Medicine \u0026 Medical Imaging"}],"tags":[],"updatedAt":"2023-03-27T20:55:33+00:00","versionOfRecord":[],"versionCreatedAt":"2023-03-27 20:55:32","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-2717573","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-2717573","identity":"rs-2717573","version":["v1"]},"buildId":"cBFmMYwuxLRRLfASyISRj","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}