Explainable AI in brain tumor diagnosis: A critical review of ML and DL techniques

Explainable AI in brain tumor diagnosis: A critical review of ML and DL techniques | 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 Systematic Review Explainable AI in brain tumor diagnosis: A critical review of ML and DL techniques Krishan Kumar, Dr.Kiran Jyoti, Krishan Kumar This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5580195/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 Brain tumors, caused by abnormal tissue growth within the brain, can severely disrupt brain functions and pose significant health risks. As the tumor progresses to higher stages, the patient's prognosis and survival decrease, resulting in a high mortality rate. With the advancements in medical imaging, especially the use of MRI, AI approaches have emerged as strong tools for detecting, segmenting, and classifying brain cancers. CNN and hybrid models, such as Vision Transformers (ViTs), have produced promising findings in this area. Although AI models exhibit high accuracy, they suffer from a lack of transparency and interpretability, paving the way for the development of eXplainable AI (XAI) methods in brain disease diagnosis. This paper investigates the utilization of machine learning, deep learning, and explainable AI (XAI) in brain tumor detection, segmentation, and classification. In this study, we have utilized the Preferred Reporting Items for Systematic Reviews and Meta-Analyses checklist and diagram. Peer-reviewed articles from PubMed, IEEE Explore, ScienceDirect, Google Scholar, Springer, and Wilay online libraries were searched, and only those papers were selected that were published in Scopus, SCIE, and ESCI-indexed journals. We have identified the 20 research papers published between 2020 and 2024 that used machine learning, deep learning and explainable AI to detect, segment, and classify the brain tumor. This review provides a comprehensive survey the of explainable artificial intelligence (XAI) in biomedical imaging, focusing on its role in the detection, segmentation and classification of brain tumors. It examines various machine learning, deep learning and XAI techniques, addresses current challenges, and suggests future directions. The objective is to provide clinicians, regulators and AI developers with valuable insights to improve the transparency and reliability of these methods in medical diagnostics. Artificial Intelligence and Machine Learning Machine Learning Deep Learning Explainable AI Brain tumor Detection Segmentation Classification Figures Figure 1 Figure 2 Figure 3 1. Introduction A brain tumor is an abnormal tissue mass that develops within brain cells and disrupts normal brain functions. It is classified as benign or malignant. Benign tumors are usually classified as grade 1 or 2, while malignant tumors are more serious and fall into grades 3 and 4. Malignant tumors are also categorized by their level of aggressiveness, with some being less aggressive and others being very aggressive. Important factors that help determine the grade of a tumor are its vascularity, invasiveness and growth rate. As a tumor progresses to a higher stage, the patient's prognosis and chances of survival worsen significantly. Therefore, diagnosis of brain tumor and early treatment will certainly improve the patient's chances of survival [ 1 – 3 ]. In 2023, an estimated 24,810 adults (14,280 men and 10,530 women) in the United States were diagnosed with primary brain or spinal cord cancer. These brain tumors account for approximately 85–90% of central nervous system (CNS) tumors. It is estimated that 5,230 children under 20 years of age will be diagnosed with CNS cancer in the United States by 2023 [ 4 ]. MRI is widely used to diagnose primary tumors because it provides highly detailed images that are essential for the clinical detection and diagnosis of brain tumors [ 5 ]. This imaging technique is very accurate and provides clear images, which is why it is an important tool in clinical practice for diagnosing brain diseases [6]. Properly diagnosing diseases requires a lot of medical training and can take a long time. Many areas in the medical field suffer from a shortage of medical professionals, which leads to delays in life-saving measures [7] [8]. The field of artificial intelligence (AI) has made significant progress in recent years, particularly in medical imaging and diagnostics. AI models based on deep learning have shown remarkable performance in detecting, segmenting and classifying brain tumors from medical images such as MRI and CT scans [ 1 ]. Convolutional Neural Networks (CNN) are state-of-the-art models for image segmentation and classification [9], although encoder-decoder-transformer architectures have also been proposed [10–12]. These AI-driven tools promise to expand the capabilities of physicians by providing rapid, accurate, and reproducible diagnoses that are important for early intervention and treatment planning in brain tumor cases [13]. In addition, AI algorithms such as machine learning and deep learning have many disadvantages. They are also black boxes for predictive interpretation, making decisions that are not easily interpreted by human experts. This lack of transparency represents a major challenge in healthcare, where understanding the reasons behind a diagnosis is critical to gaining the trust of doctors and patients alike. To solve this problem, the science of explainable AI (XAI) has grown exponentially with its successful application in brain tumor diagnosis [14]. However, a thorough review of existing studies is still needed to help researchers and practitioners gain insights and understanding of the field. Therefore, this systematic review is conducted. The main contributions of this review are : 1. Investigating the best current ML and DL techniques used in the detection, segmentation, and classification of brain tumors using MRI. 2. Investigating the XAI methods used in detection, segmentation, and classification of brain tumors using MRI. 3. Studying the impact of a code of paper is publicly accessible. 4. Investing the most widely used datasets in brain tumor detection, segmentation, and classification. 5. Summarizing the common limitations of existing studies. This article presents a comprehensive review of the latest brain tumor detection, segmentation, classification, and explainable artificial intelligence (XAI) methods to assist healthcare professionals. The rest of the paper is organized as follows: Section 2 presents the methodology, Section 3 presents the literature review, Section 4 presents results and discussion, and Section 5 presents conclusions of the work. 2. Methodology This section presents the paper's methodology for conducting the review study on brain tumor detection, segmentation, classification and explainable AI (XAI), which helps make AI more transparent in medical diagnoses. This study follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) [15] guidelines for conducting systematic reviews. PRISMA 2020 was adopted because of the precise parameters it provides for conducting rigorous systematic reviews. This review paper therefore follows the recommendations of the guidelines. 2.1 Papers selection We carefully selected the articles for review by conducting keyword searches and closely analyzing the titles, abstracts, and conclusions. We prioritized recent studies that offered something new or significantly different, all based on specific inclusion and exclusion criteria. Only the articles that met these criteria are discussed in this article. A detailed breakdown of these criteria can be found in Table II. Peer-reviewed manuscripts and conference proceedings from the published online library databases PubMed, IEEE Explore, ScienceDirect, Google Scholar, Springer, and Wilay were searched. 2.2 Search strategy Six databases (PubMed, IEEE Explore, ScienceDirect, Google Scholar, Springer, and Wilay online library) were searched systemically to filter papers. 2.3 Inclusion and exclusion criteria After applying the search equation, the criteria for inclusion and exclusion are mentioned in Table I. Table I Inclusions and Exclusions criteria Criteria No. Description Inclusions IC1 Studies related to brain tumor detection, segmentation, classification employing ML, DL and XAI techniques. IC2 Published within 2020 to 2024. IC3 Published in English language. IC4 Paper published in SCI/ ESCI and Scopus indexed journals were selected. Exclusions EC1 Studies that not conducted based of Brain MRI. EC2 Review studies. EC3 Data from books, presentations, technical documentations, and website 2.4 Study Selection The papers were selected using the inclusion and exclusion criteria defined in Table I. A bibliography with 343 papers was compiled from databases. We excluded duplicate studies, systematic reviews, scoping reviews, brain tumor diagnosis without explainable AI, and explainable AI without brain tumor diagnosis. Finally, 20 studies that used XAI for brain tumor detection, segmentation, and classification were included and included in the systematic review. Figure 1 PRISMA flow demonstrating the procedure for choosing the most suitable 20 papers 3. Results 3.1 Brain tumor Detection Tumor detection is the process of recognizing abnormal growths, or tumors, in the brain with the aid of MRI. The soft tissues of the brain can be seen in extremely detailed images by MRI, which makes it simpler to identify any issues. Scanning : High-resolution images of the brain from various angles are taken by MRI machine. Analysis : The images are analyzed for any abnormalities by a radiologist, or AI tools. Identification : The goal is to identify areas that differ from normal brain tissue and could indicate a tumor. Brain tumor detection has made dramatic progress through the application of deep learning techniques, particularly convolutional neural networks (CNNs), which eliminate the need for manual feature extraction and improve diagnostic accuracy. Recent studies have demonstrated the effectiveness of various CNN architectures such as DarkNet53, EfficientNet-B0 and DenseNet201, achieving accuracies up to 99.6% [14–19]. The major challenges in brain tumor detection include early detection of small tumors, ensuring accuracy to avoid false positives and negatives, and the limited availability of diverse, high-quality medical imaging datasets. Precise tumor segmentation is crucial for accurate diagnosis and treatment planning. Additionally, the black box nature of deep learning models can lead to trust issues among physicians due to a lack of transparency. To overcome these issues, recent studies on brain tumor detection using machine learning and deep learning have shown promising results, especially given the increasing emphasis on explainable AI (XAI). For example, Windisch et al. [20] combined ResNet50 and Bayesian neural networks with MRI data from the TCGA-GBM dataset. Using Grad-CAM, they achieved an incredible 93% accuracy and interpretation. Although the accuracy was excellent, the study also showed that the model sometimes focused on irrelevant areas, raising concerns about its generalizability and reliability across different data types. Similarly, Esmaeili et al. [21] tested DenseNet-121 on MRI data and achieved a classification accuracy of 92.1%. Using Grad-CAM helped them recognize the model's decisions and showed clear visualizations of tumor regions. However, the model struggled when images contained multiple objects, and the small sample of patients limited its broader applicability. In 2023, Ullah et al. [22] created the DeepEBTDNet model, which achieved an astounding accuracy of 98.96% for brain tumor identification. This model was extremely effective at detecting tumors; However, validation for other diseases or image types was lacking. Eder et al. [ 23 ] used CNNs with SHAP and achieved 94% accuracy on BraTS 2020 dataset. The SHAP visualizations facilitated understanding of the model's predictions and increased clinical relevance. However, the study acknowledged that more data sets would be needed to ensure the generalizability of the model. In 2024, several studies including those by Brima, Appiah, Tehsin and Pasvantis examined various CNN architectures such as VGG16, ResNet50, MobileNetV2 and InceptionV3. Brima et al. [ 24 ] used DenseNet121 and saliency maps to achieve 98.10% accuracy on brain MRI images. However, generalizability remained an issue due to a single dataset. Appiah et al. [ 25 ] used the transfer learning models MobileNetV2, Inception-v3, ResNet101 and VGG-19 on the Brats 2018 dataset and MobileNetV2 achieved 99.82% accuracy using SHAP. Despite these results, the study was limited to binary classifications and required extensive data preparation. Tehsin et al. [ 26 ] developed the DaSAM model, which achieved 99% accuracy in brain tumor detection, but had problems with data imbalance and a lack of accurate tumor segmentation. Finally, Pasvantis et al. [ 27 ] found that ResNet50V2 improved tumor segmentation, although concerns remain regarding edge detection and mask reliability. Table II summarizes the selected studies and highlights the techniques for brain tumor detection using ML, DL, and XAI. 3.2 Brain tumor segmentation Brain tumor segmentation is important for brain tumor diagnosis, treatment planning, and monitoring because it distinguishes between healthy and malignant tissue in MRI or CT scans. Manual segmentation by radiologists can be time-consuming and subjective, but AI-based algorithms, particularly those based on machine learning and deep learning, have significantly increased accuracy and speed. Recent research has shown that models such as U-Net, ResNet and DenseNet can achieve greater than 98% accuracy. In addition to speed, explainability has also become a priority, with tools like Grad-CAM and SHAP providing visual explanations of how AI models make decisions. These explainability tools help physicians understand and trust AI results, increasing their likelihood of putting these technologies into practice. Finally, the combination of high-performance models and transparent decision-making paves the way for broader use of AI in medical imaging [ 28 – 29 ]. Table III summarizes the selected studies and highlights the brain tumor segmentation techniques using ML, DL and XAI techniques. In recent years, many advances have been made in brain tumor segmentation and classification by using deep learning approaches combined with explainable AI methods to increase physician confidence. Zeineldin et al. [ 28 ] used a variety of state-of-the-art XAI approaches, including Grad-CAM and SmoothGrad, coupled with ResNet-50 to evaluate brain gliomas and segment tumors. Their model exhibits a remarkable classification accuracy of 98.62%, with Dice scores of 84.10 (enhanced tumor), 87.33 (tumor core), and 92.00 (whole tumor). However, their research found that deep learning models in particular continue to lack interpretability with multimodal MRI data and highlighted that most XAI approaches are still designed for 2D MRI slices rather than entire 3D volumes. Yang et al. [ 29 ] proposed an ODE neural model with good segmentation accuracy (more than 99%) across all criteria. They emphasized the importance of specific MRI modalities – T1-Ce for tumor enhancement and T2 and FLAIR for complete tumor segmentation. However, their results highlighted the need for additional data set validation and the high computational cost of the model. Table II: Brain Tumor detection techniques. Sr. No Study Imaging Modality ML/DL/TF Dataset XAI Results Limitations 1 Windisch et al.,[ 20] MRI T1-weighted, T2-weighted Resnet50, and Bayesian neural network TCGA-GBM GradCAM The model achieved a categorical accuracy of 93%. 1.The model showed a tendency to focus on irrelevant areas. 2.The MRI data used for training came from different views and acquisition sequences. 3.Limited Generalizability. 2 Esmaeili et al., [ 21] MRI T1-weighted, T2-weighted, and FLAIR DenseNet-121, GoogLeNet, and MobileNet, TCGA CIA Grad-CAM DenseNet-121 achieved the highest classification (92.1%) and localization accuracy (81.1% hits and 79.1% IoU. 1.A significant number of tumor brains were incorrectly classified. 2.Grad-CAM struggled with complex images containing multiple objects. 3.The study used limited patient data. 3 Ullah et al.,[22] MRI DeepEBTDNet BTDMRI database and the Brain MRI Images for BTD dataset. LIME The model achieved a validation accuracy of 98.96% and a testing accuracy of 94.0% 1.The model is specifically designed for brain tumor detection and has not been validated on other types of medical images or conditions. 2.The study does not explicitly differentiate between benign and malignant tumors. 4 Eder et al.,[ 23 ] MRI T1-weighted, T2-weighted, FLAIR, and T1ce CNN BraTS 2020 SHAP The model achieved up to 94% accuracy. 1.The lack of additional datasets, particularly for network validation. 2.optimization of network structures was not explored in-depth. 3.Limited generalizability. 5 Brima et al., [24 ] MRI T1-weighted VGG16, VGG19, ResNet50, ResNet50V2, InceptionV3, Xception, EfficientNetB0, DenseNet121 Figshare and the COVID-19 X-ray Saliency maps With the brain MRI dataset, the DenseNet121 model achieved 98.10% accuracy 1.Limited generalizability. 2.The Vanilla Gradient method and SmoothGrad produced noisy and less reliable saliency maps compared to XRAI. 3.No Quantitative Metrics for Explainability 6 Appiah et al.,[ 25] MRI MobileNetV2, Inception-v3, ResNet101, VGG-19 Brats 2018 SHAP 1.CNN predicted tumors with 99.21% accuracy. 2.MobileNetV2 delivered the highest accuracy of 99.82%, outperforming other models. 1.The study focuses on binary classification only. 2.Limited generalizability. 3.The extensive pre-processing required. 7 Tehsin et al., [ 26 ] MRI DaSAM Based on CNN Kaggle: Brain Tumor Classification and Figshare brain tumor dataset. 2D Grad Cam The method achieved precision values of 99% and 96% on the Figshare and Kaggle datasets. 1.Limited generalizability. 2.The model focuses on classification rather than precise tumor segmentation, 3.The authors mention an imbalance in the data distribution. 8 Pasvantis et al., [ 27 ] MRI InceptionV3, ResNet50V2,NasNetLarge Kaggle:Brain Tumor Dataset LIME ResNet50V2 outperformed both InceptionV3 and NasNetLarge in terms of F1-scores, as confirmed by statistical tests (p = 0.02 for InceptionV3 and p = 0.008 for NasNetLarge). 1.Edge detection algorithms sometimes produce unreliable masks, affecting the refinement quality. 2.Limited generalizability. Similarly, Yan et al. [ 30 ] presented a 3D nnU-Net model that was superior in segmentation, with high cube similarity coefficients and a classification accuracy of 95.46%. They used Grad-CAM + + to create visual heatmaps that helped interpret the regions of interest in MRI data. Despite the impressive results, their model suffered from generalizability across datasets and required significant processing resources. Zeineldin et al. [ 31 ] extended their work by developing the TransXAI model, which used a hybrid architecture that included CNNs and transformers. Table III Summary of Segmentation techniques. Sr. No Study Imaging Modality ML/DL/TF Dataset XAI Results limitations 1 Zeineldin et al.,[28 ] MRI native T1W, Gadolinium T1Gd, T2W, and FLAIR ResNet-50,3D DeepSeg,Vanilla gradient, guided backpropagation, BraTS 2019 and 2021 Integrated gradients, Guided integrated gradients, SmoothGrad, Grad-CAM, and guided Grad-CAM. 1.ResNet 50 achieved a classification accuracy of 98.62% for brain glioma grading. 2.For BT segmentation, the model achieved Dice scores of 84.10 (enhancing tumor), 87.33 (tumor core), and 92.00 (whole tumor). 1.Lack of Interpretability in Deep Learning Models. 2.Current explainability methods struggle to effectively interpret complex multi-modal MRI data. 3.Many existing XAI methods are focused on 2D MRI slices. 2 Yang et al.,[ 29] MRI T1-weighted, T2-weighted, FLAIR, and T1ce Neural ODE, U-Net and Cascaded U-Net BraTS 2020 Neural Ordinary Differential Equation (ODE) model The proposed model achieved accuracy across all segmentation tasks: ET (99.649%), TC (99.425%), and WT (99.383%). 1.Limited generalizability. 2.Model requires significant computational resources. 3.The model is based on 2D MRI slices rather than full 3D volumes. 3 Yan et al., [ 30 ] MRI T1-weighted, T2-weighted, FLAIR, and T1ce 3D nnU-Net model BraTS 2018 Grad Cam 1.The nnU-Net model achieved high Dice Similarity Coefficients (ET: 0.79, WT: 0.91, TC: 0.85), demonstrating strong segmentation performance. 1.Required significant computational resources 2.Framework heavily rely on the availability of large datasets. 3.Limited generalizability. 4 Zeineldin et al.,[ 31 ] MRI T1-weighted, T2-weighted, FLAIR, and T1ce CNN and Grad Cam BraTS 2019 and FeTS2022 datasets. Grad-Cam The TransXAI model achieved a Dice Similarity Coefficient: Enhancing Tumor (ET): 0.745 Tumor Core (TC): 0.782 Whole Tumor (WT): 0.882 1.The current implementation focuses on 2D axial MRI slices. 2.The hybrid architecture may require significant computational resources. 5 Hassan et al.,[ 32 ] MRI T1-weighted, T2-weighted, FLAIR, and T1ce - Dataset source is not mentioned. NSL It introduces the concept of neuro-symbolic learning as a solution to enhance the interpretability and explainability of AI in clinical practices, particularly in brain tumor segmentation . 1.Paper proposed a strong case for NSL, it remains theoretical. 2.Dataset source is not mentioned. This model increased explainability even further by visualizing its decisions using Grad-CAM. However, as with previous studies, the implementation focused on two-dimensional axial MRI slices and the computational requirements remained significant. Finally, Hassan et al. [ 32 ] proposed a more theoretical technique called Neuro-Symbolic Learning (NSL), which aims to improve interpretability by merging neural networks and symbolic AI. Although this methodology has the potential to improve the explainability of AI models in clinical practice, the work is still theoretical and the dataset used is not disclosed. 3.3 Brain Tumor Classification Brain tumor classification is a significant issue in medical imaging, particularly using magnetic resonance imaging, which is preferred due to its non-invasive nature and enhanced soft tissue visualization. Recent research has used various machine learning approaches, including convolutional neural networks (CNNs) and capsule networks (CapsNets), to improve classification accuracy. Table IV summarizes the selected works focusing on the classification of brain tumors using ML, DL and XAI approaches. Table IV Summary of Brain tumor classification techniques. Sr. No Study Imaging Modality ML/DL/TF Dataset XAI Results Limitations 1 Jin et al.,[33 ] MRI T1-weighted, T2-weighted, FLAIR, and T1ce CNN, BraTS 2020 Grad Cam and SHAP 1.The 16 examined heatmap algorithms failed to meet clinical requirements for indicating the AI model’s decision process or quality. 1.Current XAI Algorithms Are Not Clinically Designed. 2.Most XAI algorithms struggle to highlight the importance of different imaging modalities, such as T1, T2, or FLAIR. 2 Dasanayaka et al.,[ 34 ] MRI and WSI DenseNet, ResNet MRI sequences and WSI Grad-Cam The proposed method achieved an accuracy of 86% during classification. 1.Source of dataset is not mentioned. 2.Limited generalizability. 3.The tool currently supports a limited number of imaging modalities. 3 Marmolejo et al.,[ 35 ] MRI T1-weighted, T2-weighted, FLAIR, and T1ce CNN BRATS 2017 . numGrad Cam The proposed model achieved an accuracy of 97.11%, sensitivity of 95.58%, and specificity of 96.81%. 1.Limited generalizability. 2.There is a need for further validation. 3.The model requires significant computational resources. 4 Mandloi et al., [ 36 ] MRI Conditional GAN, MobileNet, Inception ResNet, Efficient Net, and VGGNet, MobileNet Kaggle: Brain MRI Images for B T Detection Kaggle: B T Classification LRP 1.InceptionResNetV2 achieved training, validation, and testing accuracies of 99.6%, 99.2%, and 99.0%, respectively, for tumor detection. 1.InceptionResNetV2 is computationally expensive. 2.Validation thorugh additional dataset is required. 2.Limited generalizability. 5 Šefčík et al., [ 37] T1-weighted and T2-weighted, and FLAIR DNN BraTS 2020 LRP, Grad Cam The proposed method increased classification accuracy from 72–79%. 1.Limited generalizability. 2.Model neglect other potentially relevant features in the MRI images that could contribute to classification. 6 Bhuvaneswari et al.,[ 38 ] MRI InceptionV3, ResNet-50, VGG16, DenseNet TCGA LGG Dataset LRP ResNet-50 achieved the highest accuracy (95.5%) with superior efficiency, followed by DenseNet (94.65%) and VGG16 (91.04%) 1.Limited generalizability. 2.The model's performance heavily depends on the quality and diversity of the training dataset. 3.The model lacks comprehensive clinical validation. 7 Nag et al.,[ 39 ] MRI ResNet50,Generative Adversarial Networks Kaggle: Brain Tumor MRI dataset LIME The proposed model, TumorGANet, achieved exceptional results with an accuracy of 99.53%, precision and recall rates of 100%, F1-scores of 99%, and a Hamming loss of 0.2%. 1.Limited generalizability. 2.There is a risk that the GAN-generated data could introduce artifacts or features that do not accurately represent real tumors. Jin et al. [ 33 ] investigated the limitations of current explainable AI (XAI) approaches in brain tumor classification. Despite testing numerous heatmap algorithms, they discovered that they failed to fulfill clinical criteria for decision transparency, leaving clinicians struggling to interpret the data. The study underscores the need for clinically driven advancements in XAI. Dasanayaka et al. [ 34 ] created a web program that uses MRI and whole slide imaging (WSI) to classify brain cancers with 86% accuracy. Though it closely resembles genuine diagnostic workflows, the tool's limited generalizability and imaging modality support emphasize the difficulties of implementing AI in real-world settings. Marmolejo et al. [ 35 ] developed numGrad-CAM to improve explainability in brain tumor classification models and achieved 97.11% accuracy. Despite receiving encouraging feedback from clinicians, the model confronts limitations in generalizability and computing demands, necessitating additional validation. Mandloi et al. [ 36 ] employed InceptionResNetV2 and EfficientNet to classify brain tumors with above 99% accuracy. However, the high computational cost and the necessity for further validation limit its practical adoption. Šefčík et al. [ 37 ] suggested an approach that improved classification accuracy from 72–79% using explanation-guided training. However, the model may omit relevant MRI features. Generalizability remains an issue. Bhuvaneswari et al. [ 38 ] demonstrated the effectiveness of a hybrid CNN model, with ResNet-50 leading in accuracy (95.5%). The addition of LRP increased transparency, but the model's reliance on dataset quality and lack of clinical validation limit its general applicability. Nag et al. [ 39 ] created TumorGANet, a GAN-based model capable of generating synthetic MRI data with excellent accuracy (99.53%). While highly effective, worries regarding potential artifacts in synthetic data indicate that more refinement is required for real-world application. 3.4 Explainable AI techniques used in Brain tumor detection, segmentation and Classification Explainable AI (XAI) methods play a critical role in improving the interpretability of models used for brain tumor detection, segmentation, and classification. Methods such as Grad-CAM and layer-wise relevance propagation (LRP) have been used to provide insights into deep learning models' decision-making processes, thereby solving the black box issue common in AI systems. For example, the Disease and Spatial Attention Model (DaSAM) uses attention mechanisms to highlight important aspects in MRI images, allowing for a better comprehension of tumor characteristics [ 26 ]. Furthermore, saliency maps have been used to detect important areas in brain MRIs, increasing the dependability and accountability of AI systems in clinical settings [ 40 ]. These methods not only improve model transparency but also build confidence among medical practitioners, resulting in more quick and accurate diagnoses [ 36 ]. Table VI depicts the spectrum of explainability strategies employed in brain tumor detection, segmentation, and classification studies: Table V: Summery of XAI techniques used. XAI Method GradCam/Varient LRP LIME SHAP Others Count 11 3 3 3 2 Grad-CAM/Variants The most commonly utilized approach, as seen in 11 investigations. Grad-CAM and its variants are well-known for their ability to show the sections of the input image that are most important to the model's prediction, making them valuable for understanding model decisions. Grad-CAM and its variants are popular explainability techniques in deep learning, particularly for interpreting CNNs in visual tasks such as medical image analysis. Grad Cam has several variations, including SmoothGrad, Grad-CAM, guided Grad-CAM, numGrad Cam, and 2D Grad Cam. SHAP (Shapley Additive Explanations ): Used in three investigations. SHAP uses a unifying strategy to explain the output of machine learning models by assigning relevance ratings to each feature, which aids in evaluating the contributions of various input sources. LIME (Local Interpretable Model-Agnostic Explanations) is also used in three research studies. LIME is an explanation technique that localizes a complex model with an interpretable one, providing insights into specific predictions. LRP (Layer-wise Relevance Propagation) Used in two studies. LRP explains model decisions by propagating prediction scores backward and highlighting key elements. Others Saliency maps and the Neural Ordinary Differential Equation (ODE) model, which are less often utilized explanatory approaches, were used in two investigations. 4. Discussion In this section, we provide a detailed discussion of the key findings related to machine learning (ML), deep learning (DL), and explainable AI techniques for brain tumor detection, segmentation, and classification using MRI data. We also examine the role of Explainable AI (XAI) techniques, dataset availability, and common limitations in the current literature. 4.1 Investigating the best current ML and DL techniques used in detection, segmentation, and classification of brain tumors using MRI. Convolutional neural networks (CNNs) have become a foundation in brain tumor detection and classification using MRI. Widely used models like ResNet, DenseNet, VGG, and U-Net have consistently shown strong performance. For example, ResNet-50 achieved a high classification accuracy of 95.5% in a study by Ullah et al. [22], with DenseNet and VGG16 following closely behind. Hybrid models, which combine CNNs with more advanced methods, such as generative adversarial networks (GANs) and vision transformers (ViTs), are gaining recognition from the scientific community due to their potential to improve detection and segmentation. A significant example is the TransXAI model, which combines CNNs and ViTs and has produced excellent results, particularly for tumor core and entire tumor segmentation. Bayesian Neural Networks (BNNs) are also becoming increasingly widely employed, particularly for uncertainty estimates, which are important in medical decision-making. Windisch et al. [20] used a BNN with ResNet50 to achieve a categorical accuracy of 93%, providing insight into the model's prediction confidence. Finally, neural ODEs (ordinary differential equations) are a recent approach to tumor segmentation. They have had good results, with accuracy rates surpassing 99% for a variety of tumor locations, including the enhancing tumor, tumor core, and entire tumor. 4.2 Investigating the XAI methods used in detection, segmentation, and classification of brain tumors using MRI. Explainable AI (XAI) techniques are essential for making deep learning models more transparent, especially in critical areas like brain tumor detection and classification. Some of the most commonly used XAI methods include: GradCAM and its variants Grad-CAM (Gradient-weighted Class Activation Mapping) and its upgraded versions, such as Guided Grad-CAM and Grad-CAM++, are widely used for determining which regions of an MRI scan are most relevant for a model's decision. These algorithms create heatmaps that emphasize tumor locations, allowing doctors to better understand how the model operates. Windisch et al. [20] and Esmaeili et al. [21] used Grad-CAM to explain CNN judgments, resulting in good classification accuracy. SHAP (Shapley Additive Explanations) SHAP assigns relevance scores to individual features, describing how each feature contributes to the model's prediction. For example, Appiah et al. [ 25 ] employed SHAP and MobileNetV2 to classify tumors, attaining great accuracy (99.82%) while keeping transparency in decision-making. LIME (Local Interpretable Model-Agnostic Explanations) LIME generates localized explanations for each situation, assisting in explaining specific predictions. Ullah et al. [22] used it to improve the interpretability of a variety of deep learning models, guaranteeing that the model's behavior is intelligible while maintaining accuracy. Layer-wise Relevance Propagation (LRP) LRP reveals pixel-level details on how neural networks generate judgments. Brima et al. [ 24 ] used LRP to determine which areas of an MRI image were most useful for tumor detection and classification. 4.3 Studying the impact of a code of paper that is publicly accessible. The availability of code is critical to reproducibility and future study. According to our evaluation, the majority of the publications did not openly publish their code. As shown in Table VII, only six of the 20 reviewed publications specifically acknowledged their code and were made publicly available. The lack of open-source implementations hinders other researchers' capacity to reproduce or expand on the findings of these investigations, which is a typical difficulty in medical AI research. However, some recent studies have begun to share their code, indicating a favorable trend toward open science. Table VI depicting the code availability count in research studies. Code Availability Yes No Count 6 14 4.4 Investing the most widely used datasets in brain tumor detection, segmentation, and classification. Several well-established datasets have been consistently used for brain tumor analysis in the literature. Based on the provided data, the most widely used datasets in brain tumor detection, segmentation, and classification are: BraTS (2017, 2018, 2019, 2020, 2021) This dataset appears frequently and is the most popular for brain tumor segmentation tasks. TCGA (GBM, LGG, CIA) The Cancer Genome Atlas datasets are widely used for segmentation and classification applications. Kaggle Brain Tumor Datasets Several studies use Kaggle datasets, including "Brain MRI Images for Brain Tumor Detection" and "Brain Tumor Classification (MRI)." Figshare Brain tumor detection dataset is used in study. BTDMRI Another dataset used for MRI tumor identification. These datasets, particularly BraTS and Kaggle, are fundamental to brain tumor analysis research. 4.5 Summarizing the common limitations of existing studies. Despite the progress in deep learning for brain tumor detection, segmentation, and classification, several challenges remain: Generalization Issues Many models, such as BraTS, perform well on some datasets but struggle on others, indicating a lack of generalizability, which is critical for real-world clinical usage. Limited Interpretability Although techniques such as Grad-CAM and LRP provide some interpretability, the explanations might be ambiguous and may not always match clinical standards. Heatmaps frequently do not correspond to what medical practitioners expect. High computational demands Advanced models, particularly those based on hybrid methods such as GANs or Transformers, necessitate a significant amount of processing resources, making them challenging to apply in real-time or in resource-constrained environments, such as smaller medical institutions. Limited Code Availability Many researchers do not make their code available to the public, making it difficult to duplicate and validate results. The absence of standardization throughout research impacts comparisons of different studies. Data Imbalance Many brain tumor databases contain class imbalances, with fewer examples of unusual tumor forms. This makes it more difficult for models to perform well across all classes, sometimes resulting in lower outcomes for underrepresented tumor classifications. 5. Conclusions This study focuses on advances in applying machine learning (ML) and deep learning (DL) to detect, segment, and classify brain tumors using MRI. CNN-based models like ResNet, DenseNet, and U-Net have proven to be extremely effective, while newer hybrid models that combine CNNs with GANs and Vision Transformers improve performance even further. Explainable AI (XAI) approaches such as Grad-CAM, SHAP, and LIME have increased model interpretability by visualizing decision-making processes. However, these explanations remain insufficiently thorough to meet clinical needs, restricting their applicability in real-world healthcare settings. Despite advancements, some difficulties persist. A major difficulty is that many models perform well on specialized datasets, such as BraTS, but struggle to generalize to other ones. The complexity of some models also makes them too resource-intensive for real-time use in many hospitals. Furthermore, only six of the 20 research studies made their code publicly available, making it difficult to replicate results and progress the area. BraTS, TCGA, and Kaggle are the most extensively utilized datasets in this field, and while they provide rich data, they also suffer from data imbalance, particularly for rarer tumor types. In conclusion, while deep learning models show great potential for brain tumor analysis, they still confront challenges in interpretability, generalizability, and practical application. Addressing these issues and making models more accessible will be critical in bringing these technologies closer to routine clinical use. References Al-Galal, S.A.Y.; Alshaikhli, I.F.T.; Abdulrazzaq, M.M. MRI brain tumor medical images analysis using deep learning techniques:A systematic review. Health Technol. 2021, 11, 267–282. Louis, D.N.; Perry, A.; Wesseling, P.; Brat, D.J.; Cree, I.A.; Figarella-Branger, D.; Hawkins, C.; Ng, H.K.; Pfister, S.M.;Reifenberger, G. The 2021 WHO classification of tumors of the central nervous system: A summary. Neuro-Oncology 2021, 23,1231–1251. Nodirov, J.; Abdusalomov, A.B.; Whangbo, T.K. Attention 3D U-Net with Multiple Skip Connections for Segmentation of Brain Tumor Images. Sensors 2022, 22, 6501. https://www.cancer.net/cancer-types/brain-tumor/statistics C. Wang et al., “Phenotypic and genetic associations of quantitative magnetic susceptibility in U.K. biobank brain imaging,” Nature Neurosci., vol. 562, pp. 1–14, May 2022. Z. Liu et al., “Deep learning based brain tumor segmentation: A survey,” Complex Intell. Syst., vol. 9, no. 1, pp. 1001–1026, 2020. M. A. Ottom, H. A. Rahman, and I. D. Dinov, “ZNet: Deep learning approach for 2D MRI brain tumor segmentation,” IEEE J. Transl. Eng. Health Med., vol. 10, May 2022, Art. no. 1800508. H. S. Abdulbaqi, K. N. Mutter, M. Z. M. Jafri, and Z. A. Al-Khafaji, “Estimation of brain tumour volume using expanded computed tomography scan images,” in Proc. IEEE 23rd Iranian Conf. Biomed. Eng. 1st Int. Iranian Conf. Biomed. Eng., 2016, pp. 117–121. A. Barredo Arrieta, N. Díaz-Rodríguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garcia, S. Gil-Lopez, D. Molina, R. Benjamins, R. Chatila, F. Herrera, Explainable Artificial Intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI, Inf. Fusion 58 (2020) 82–115. M. Ahamed, A. Imran, Joint learning with local and global consistency for improved medical image segmentation, in: Annual Conference on Medical Image Understanding and Analysis, 2022. A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2010. R. Strudel, R. Garcia, I. Laptev, C. Schmid, Segmenter: transformer for semantic segmentation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021 H, A., MO, D., & B, S. (2023). Current challenges of the state-of-the-art of AI techniques for diagnosing brain tumor. Material Science & Engineering International Journal, 7(4), 196-208. https://doi.org/10.15406/mseij.2023.07.00224 Ali, S., Abuhmed, T., El-Sappagh, S.,et al. (2023). Explainable artificial intelligence (XAI): What we know and what is left to attain trustworthy artificial intelligence. Information Fusion, 99, 101805. https://doi.org/10.1016/j.inffus.2023.101805 Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021;372:n71. Gillies, R. J., Kinahan, P. E., & Hricak, H. (2016). Radiomics: Images are more than pictures, they are data. Radiology, 278(2), 563-577. Zhou, S. K., Greenspan, H., & Shen, D. (2017). Deep learning for medical image analysis. Academic Press. Zhou, M., Scott, J., Chaudhury, B., Hall, L., Goldgof, D., Yeom, K., Iv, M., & Erickson, B. (2018). Radiomics in brain tumor: image assessment, quantitative feature descriptors, and machine-learning approaches. AJNR American Journal of Neuroradiology, 39(2), 208-216. Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A., Ciompi, F., Ghafoorian, M., & van der Laak, J. A. (2017). A survey on deep learning in medical image analysis. Medical Image Analysis, 42, 60-88. Cheng, J., Huang, W., Cao, S., Yang, R., Yang, W., Yun, Z., Wang, Z., & Feng, Q. (2016). Enhanced performance of brain tumor classification via tumor region augmentation and partition. PLOS ONE, 10(10), e0140381. Windisch, P., Weber, P., Fürweger, C., Ehret, F., Kufeld, M., Zwahlen, D., & Muacevic, A. (2020). Implementation of model explainability for a basic brain tumor detection using convolutional neural networks on MRI slices. Neuroradiology, 62(11), 1515-1518. https://doi.org/10.1007/s00234-020-02465-1 Esmaeili, M., Vettukattil, R., Banitalebi, H., Krogh, N. R., & Geitung, J. T. (2021). Explainable artificial intelligence for human-machine interaction in brain tumor localization. Journal of Personalized Medicine, 11(11), 1213. https://doi.org/10.3390/jpm11111213 Ullah, N., Hassan, M., Khan, J. A., Anwar, M. S., & Aurangzeb, K. (2023). Enhancing explainability in brain tumor detection: A novel DeepEBTDNet model with LIME on MRI images. International Journal of Imaging Systems and Technology, 34(1). https://doi.org/10.1002/ima.23012 Eder, M., Moser, E., Holzinger, A., Jean-Quartier, C., & Jeanquartier, F. (2022). Interpretable machine learning with brain image and survival data. BioMedInformatics, 2(3), 492-510. https://doi.org/10.3390/biomedinformatics2030031 Brima, Y., & Atemkeng, M. (2023). Visual interpretable and explainable deep learning models for brain tumor MRI and COVID-19 chest X-ray images. https://doi.org/10.21203/rs.3.rs-3241888/v1 Appiah, R., Pulletikurthi, V., Esquivel-Puentes, H. A., Cabrera, C., Hasan, N. I., Dharmarathne, S., Gomez, L. J., & Castillo, L. (2024). Brain tumor detection using proper orthogonal decomposition integrated with deep learning networks. Computer Methods and Programs in Biomedicine, 250, 108167. https://doi.org/10.1016/j.cmpb.2024.108167 Tehsin, S., Nasir, I. M., Damaševičius, R., & Maskeliūnas, R. (2024). DaSAM: Disease and spatial attention module-based explainable model for brain tumor detection. Big Data and Cognitive Computing, 8(9), 97. https://doi.org/10.3390/bdcc8090097 Pasvantis, K., & Protopapadakis, E. (2024). Enhancing deep learning model Explainability in brain tumor datasets using post-heuristic approaches. Journal of Imaging, 10(9), 232. https://doi.org/10.3390/jimaging10090232 Zeineldin, R. A., Karar, M. E., Elshaer, Z., Coburger, ·., Wirtz, C. R., Burgert, O., & Mathis-Ullrich, F. (2022). Explainability of deep neural networks for MRI analysis of brain tumors. International Journal of Computer Assisted Radiology and Surgery, 17(9), 1673-1683. https://doi.org/10.1007/s11548-022-02619-x Yang, Z., Hu, Z., Ji, H., Lafata, K., Vaios, E., Floyd, S., Yin, F., & Wang, C. (2023). A neural ordinary differential equation model for visualizing deep neural network behaviors in multi‐parametric MRI‐based glioma segmentation. Medical Physics, 50(8), 4825-4838. https://doi.org/10.1002/mp.16286 Yan, F.; Chen, Y.; Xia, Y.; Wang, Z.; Xiao, R. An Explainable Brain Tumor Detection Framework for MRI Analysis. Appl. Sci. 2023, 13, 3438. https://doi.org/10.3390/app13063438 Zeineldin, R. A., Karar, M. E., Elshaer, Z., Coburger, J., Wirtz, C. R., Burgert, O., & Mathis-Ullrich, F. (2024). Explainable hybrid vision transformers and convolutional network for multimodal glioma segmentation in brain MRI. Scientific Reports, 14(1). https://doi.org/10.1038/s41598-024-54186-7 Hassan, M., Fateh, A. A., Lin, J., Zhuang, Y., Lin, G., Xiong, H., You, Z., Qin, P., & Zeng, H. (2024). Unfolding explainable AI for brain tumor segmentation. Neurocomputing, 599, 128058. https://doi.org/10.1016/j.neucom.2024.128058 Jin, W., Li, X., & Hamarneh, G. (2022). Evaluating explainable AI on a multi-modal medical imaging task: Can existing algorithms fulfill clinical requirements? Proceedings of the AAAI Conference on Artificial Intelligence, 36(11), 11945-11953. https://doi.org/10.1609/aaai.v36i11.21452 Dasanayaka, S., Shantha, V., Silva, S., Meedeniya, D., & Ambegoda, T. (2022). Interpretable machine learning for brain tumour analysis using MRI and whole slide images. Software Impacts, 13, 100340. https://doi.org/10.1016/j.simpa.2022.100340 Marmolejo-Saucedo, J. A., & Kose, U. (2022). Numerical grad-cam based explainable Convolutional neural network for brain tumor diagnosis. Mobile Networks and Applications. https://doi.org/10.1007/s11036-022-02021-6 Mandloi, S., Zuber, M., & Gupta, R. K. (2023). An explainable brain tumor detection and classification model using deep learning and layer-wise relevance propagation. Multimedia Tools and Applications, 83(11), 33753-33783. https://doi.org/10.1007/s11042-023-16708-9 Šefčík, F., & Benesova, W. (2023). Improving a neural network model by explanation-guided training for glioma classification based on MRI data. International Journal of Information Technology, 15(5), 2593-2601. https://doi.org/10.1007/s41870-023-01289-5 Bhuvaneswari Ramakrishnan, A., Sridevi, M., Vasudevan, S. K., Manikandan, R., & Gandomi, A. H. (2024). Optimizing brain tumor classification with hybrid CNN architecture: Balancing accuracy and efficiency through oneAPI optimization. Informatics in Medicine Unlocked, 44, 101436. https://doi.org/10.1016/j.imu.2023.101436 Nag, A., Mondal, H., Mehedi Hassan, M., Al-Shehari, T., Kadrie, M., Al-Razgan, M., Alfakih, T., Biswas, S., & Kumar Bairagi, A. (2024). TumorGANet: A transfer learning and generative adversarial network- Based data augmentation model for brain tumor classification. IEEE Access, 12, 103060-103081. https://doi.org/10.1109/access.2024.3429633 Keles, A., Akcay, O., Kul, H., & Bendechache, M. (2023, August 1). Saliency Maps as an Explainable AI Method in Medical Imaging: A Case Study on Brain Tumor Classification. Irish Mahine Vision and Image Processing Conference 2023 (IMVIP2023), University of Galway, Ireland. https://doi.org/10.5281/zenodo.8199333 Additional Declarations The authors declare no competing interests. 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-5580195","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":386206463,"identity":"900de046-ec52-4313-8507-8ac27ed0bdf1","order_by":0,"name":"Krishan Kumar","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0002-1087-3724","institution":"Guru Nanak Dev Engineering College, Ludhiana","correspondingAuthor":true,"prefix":"","firstName":"Krishan","middleName":"","lastName":"Kumar","suffix":""},{"id":386206964,"identity":"90debba4-6165-43ee-b968-ce4468350bf0","order_by":1,"name":"Dr.Kiran Jyoti","email":"","orcid":"","institution":"Guru Nanak Dev Engineering College, Ludhiana","correspondingAuthor":false,"prefix":"Dr.","firstName":"Kiran","middleName":"","lastName":"Jyoti","suffix":""},{"id":386206967,"identity":"39c5ad4c-3333-4b73-852f-e1020ca49e91","order_by":2,"name":"Krishan Kumar","email":"","orcid":"","institution":"Hindu College, Amritsar","correspondingAuthor":false,"prefix":"","firstName":"Krishan","middleName":"","lastName":"Kumar","suffix":""}],"badges":[],"createdAt":"2024-12-04 13:37:12","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-5580195/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5580195/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":71148346,"identity":"65f16f4e-8180-4b57-9122-e6737b765059","added_by":"auto","created_at":"2024-12-11 14:32:55","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":62745,"visible":true,"origin":"","legend":"\u003cp\u003ePRISMA flow demonstrating the procedure for choosing the most suitable 20 papers\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5580195/v1/157b96801cc0824d2b91adea.png"},{"id":71147534,"identity":"5def97e1-61a2-4beb-b9bd-db6a877be96c","added_by":"auto","created_at":"2024-12-11 14:24:54","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":9326,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBrain tumor detection process\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5580195/v1/b2a1dff4c69b56232edf348c.png"},{"id":71147537,"identity":"9e70eca2-0873-400f-bdf4-efc042392ace","added_by":"auto","created_at":"2024-12-11 14:24:55","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":12571,"visible":true,"origin":"","legend":"\u003cp\u003eShowing distribution of XAI techniques used across the various studies.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5580195/v1/2a8ea594a8a75c53437951c0.png"},{"id":71151653,"identity":"21f61a31-3354-41a0-86ad-d14bc4b8b05d","added_by":"auto","created_at":"2024-12-11 14:56:55","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1073611,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5580195/v1/f69efab3-c4d6-4413-9952-472d010eebdd.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eExplainable AI in brain tumor diagnosis: A critical review of ML and DL techniques\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eA brain tumor is an abnormal tissue mass that develops within brain cells and disrupts normal brain functions. It is classified as benign or malignant. Benign tumors are usually classified as grade 1 or 2, while malignant tumors are more serious and fall into grades 3 and 4. Malignant tumors are also categorized by their level of aggressiveness, with some being less aggressive and others being very aggressive. Important factors that help determine the grade of a tumor are its vascularity, invasiveness and growth rate. As a tumor progresses to a higher stage, the patient's prognosis and chances of survival worsen significantly. Therefore, diagnosis of brain tumor and early treatment will certainly improve the patient's chances of survival [\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. In 2023, an estimated 24,810 adults (14,280 men and 10,530 women) in the United States were diagnosed with primary brain or spinal cord cancer. These brain tumors account for approximately 85\u0026ndash;90% of central nervous system (CNS) tumors. It is estimated that 5,230 children under 20 years of age will be diagnosed with CNS cancer in the United States by 2023 [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMRI is widely used to diagnose primary tumors because it provides highly detailed images that are essential for the clinical detection and diagnosis of brain tumors [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. This imaging technique is very accurate and provides clear images, which is why it is an important tool in clinical practice for diagnosing brain diseases [6]. Properly diagnosing diseases requires a lot of medical training and can take a long time. Many areas in the medical field suffer from a shortage of medical professionals, which leads to delays in life-saving measures [7] [8].\u003c/p\u003e \u003cp\u003eThe field of artificial intelligence (AI) has made significant progress in recent years, particularly in medical imaging and diagnostics. AI models based on deep learning have shown remarkable performance in detecting, segmenting and classifying brain tumors from medical images such as MRI and CT scans [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Convolutional Neural Networks (CNN) are state-of-the-art models for image segmentation and classification [9], although encoder-decoder-transformer architectures have also been proposed [10\u0026ndash;12]. These AI-driven tools promise to expand the capabilities of physicians by providing rapid, accurate, and reproducible diagnoses that are important for early intervention and treatment planning in brain tumor cases [13]. In addition, AI algorithms such as machine learning and deep learning have many disadvantages. They are also black boxes for predictive interpretation, making decisions that are not easily interpreted by human experts. This lack of transparency represents a major challenge in healthcare, where understanding the reasons behind a diagnosis is critical to gaining the trust of doctors and patients alike. To solve this problem, the science of explainable AI (XAI) has grown exponentially with its successful application in brain tumor diagnosis [14]. However, a thorough review of existing studies is still needed to help researchers and practitioners gain insights and understanding of the field. Therefore, this systematic review is conducted.\u003c/p\u003e \u003cp\u003e \u003cb\u003eThe main contributions of this review are\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e1. Investigating the best current ML and DL techniques used in the detection, segmentation, and classification of brain tumors using MRI.\u003c/p\u003e\u003cp\u003e2. Investigating the XAI methods used in detection, segmentation, and classification of brain tumors using MRI.\u003c/p\u003e \u003cp\u003e3. Studying the impact of a code of paper is publicly accessible.\u003c/p\u003e\u003cp\u003e4. Investing the most widely used datasets in brain tumor detection, segmentation, and classification.\u003c/p\u003e \u003cp\u003e5. Summarizing the common limitations of existing studies.\u003c/p\u003e \u003cp\u003eThis article presents a comprehensive review of the latest brain tumor detection, segmentation, classification, and explainable artificial intelligence (XAI) methods to assist healthcare professionals. The rest of the paper is organized as follows: Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents the methodology, Section \u003cspan refid=\"Sec7\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents the literature review, Section \u003cspan refid=\"Sec12\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents results and discussion, and Section \u003cspan refid=\"Sec17\" class=\"InternalRef\"\u003e5\u003c/span\u003e presents conclusions of the work.\u003c/p\u003e"},{"header":"2. Methodology","content":"\u003cp\u003eThis section presents the paper's methodology for conducting the review study on brain tumor detection, segmentation, classification and explainable AI (XAI), which helps make AI more transparent in medical diagnoses. This study follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) [15] guidelines for conducting systematic reviews. PRISMA 2020 was adopted because of the precise parameters it provides for conducting rigorous systematic reviews. This review paper therefore follows the recommendations of the guidelines.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Papers selection\u003c/h2\u003e \u003cp\u003eWe carefully selected the articles for review by conducting keyword searches and closely analyzing the titles, abstracts, and conclusions. We prioritized recent studies that offered something new or significantly different, all based on specific inclusion and exclusion criteria. Only the articles that met these criteria are discussed in this article. A detailed breakdown of these criteria can be found in Table II. Peer-reviewed manuscripts and conference proceedings from the published online library databases PubMed, IEEE Explore, ScienceDirect, Google Scholar, Springer, and Wilay were searched.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Search strategy\u003c/h2\u003e \u003cp\u003eSix databases (PubMed, IEEE Explore, ScienceDirect, Google Scholar, Springer, and Wilay online library) were searched systemically to filter papers.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Inclusion and exclusion criteria\u003c/h2\u003e \u003cp\u003eAfter applying the search equation, the criteria for inclusion and exclusion are mentioned in Table I.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTable I Inclusions and Exclusions criteria\u003c/b\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCriteria\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNo.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eInclusions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIC1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStudies related to brain tumor detection, segmentation, classification employing ML, DL and XAI techniques.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIC2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePublished within 2020 to 2024.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIC3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePublished in English language.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIC4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePaper published in SCI/ ESCI and Scopus indexed journals were selected.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eExclusions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEC1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStudies that not conducted based of Brain MRI.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEC2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReview studies.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEC3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eData from books, presentations, technical documentations, and website\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Study Selection\u003c/h2\u003e \u003cp\u003eThe papers were selected using the inclusion and exclusion criteria defined in Table I. A bibliography with 343 papers was compiled from databases. We excluded duplicate studies, systematic reviews, scoping reviews, brain tumor diagnosis without explainable AI, and explainable AI without brain tumor diagnosis. Finally, 20 studies that used XAI for brain tumor detection, segmentation, and classification were included and included in the systematic review.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFigure\u0026nbsp;1 PRISMA flow demonstrating the procedure for choosing the most suitable 20 papers\u003c/b\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Brain tumor Detection\u003c/h2\u003e \u003cp\u003eTumor detection is the process of recognizing abnormal growths, or tumors, in the brain with the aid of MRI. The soft tissues of the brain can be seen in extremely detailed images by MRI, which makes it simpler to identify any issues.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cem\u003eScanning\u003c/em\u003e: High-resolution images of the brain from various angles are taken by MRI machine.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cem\u003eAnalysis\u003c/em\u003e: The images are analyzed for any abnormalities by a radiologist, or AI tools.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cem\u003eIdentification\u003c/em\u003e: The goal is to identify areas that differ from normal brain tissue and could indicate a tumor.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eBrain tumor detection has made dramatic progress through the application of deep learning techniques, particularly convolutional neural networks (CNNs), which eliminate the need for manual feature extraction and improve diagnostic accuracy. Recent studies have demonstrated the effectiveness of various CNN architectures such as DarkNet53, EfficientNet-B0 and DenseNet201, achieving accuracies up to 99.6% [14\u0026ndash;19]. The major challenges in brain tumor detection include early detection of small tumors, ensuring accuracy to avoid false positives and negatives, and the limited availability of diverse, high-quality medical imaging datasets. Precise tumor segmentation is crucial for accurate diagnosis and treatment planning. Additionally, the black box nature of deep learning models can lead to trust issues among physicians due to a lack of transparency.\u003c/p\u003e \u003cp\u003eTo overcome these issues, recent studies on brain tumor detection using machine learning and deep learning have shown promising results, especially given the increasing emphasis on explainable AI (XAI). For example, Windisch et al. [20] combined ResNet50 and Bayesian neural networks with MRI data from the TCGA-GBM dataset. Using Grad-CAM, they achieved an incredible 93% accuracy and interpretation. Although the accuracy was excellent, the study also showed that the model sometimes focused on irrelevant areas, raising concerns about its generalizability and reliability across different data types. Similarly, Esmaeili et al. [21] tested DenseNet-121 on MRI data and achieved a classification accuracy of 92.1%. Using Grad-CAM helped them recognize the model's decisions and showed clear visualizations of tumor regions. However, the model struggled when images contained multiple objects, and the small sample of patients limited its broader applicability. In 2023, Ullah et al. [22] created the DeepEBTDNet model, which achieved an astounding accuracy of 98.96% for brain tumor identification. This model was extremely effective at detecting tumors; However, validation for other diseases or image types was lacking. Eder et al. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e23\u003c/span\u003e] used CNNs with SHAP and achieved 94% accuracy on BraTS 2020 dataset. The SHAP visualizations facilitated understanding of the model's predictions and increased clinical relevance.\u003c/p\u003e \u003cp\u003eHowever, the study acknowledged that more data sets would be needed to ensure the generalizability of the model. In 2024, several studies including those by Brima, Appiah, Tehsin and Pasvantis examined various CNN architectures such as VGG16, ResNet50, MobileNetV2 and InceptionV3. Brima et al. [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e24\u003c/span\u003e] used DenseNet121 and saliency maps to achieve 98.10% accuracy on brain MRI images. However, generalizability remained an issue due to a single dataset. Appiah et al. [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e25\u003c/span\u003e] used the transfer learning models MobileNetV2, Inception-v3, ResNet101 and VGG-19 on the Brats 2018 dataset and MobileNetV2 achieved 99.82% accuracy using SHAP. Despite these results, the study was limited to binary classifications and required extensive data preparation. Tehsin et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e26\u003c/span\u003e] developed the DaSAM model, which achieved 99% accuracy in brain tumor detection, but had problems with data imbalance and a lack of accurate tumor segmentation. Finally, Pasvantis et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e27\u003c/span\u003e] found that ResNet50V2 improved tumor segmentation, although concerns remain regarding edge detection and mask reliability. Table II summarizes the selected studies and highlights the techniques for brain tumor detection using ML, DL, and XAI.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Brain tumor segmentation\u003c/h2\u003e \u003cp\u003eBrain tumor segmentation is important for brain tumor diagnosis, treatment planning, and monitoring because it distinguishes between healthy and malignant tissue in MRI or CT scans. Manual segmentation by radiologists can be time-consuming and subjective, but AI-based algorithms, particularly those based on machine learning and deep learning, have significantly increased accuracy and speed. Recent research has shown that models such as U-Net, ResNet and DenseNet can achieve greater than 98% accuracy. In addition to speed, explainability has also become a priority, with tools like Grad-CAM and SHAP providing visual explanations of how AI models make decisions. These explainability tools help physicians understand and trust AI results, increasing their likelihood of putting these technologies into practice. Finally, the combination of high-performance models and transparent decision-making paves the way for broader use of AI in medical imaging [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Table III summarizes the selected studies and highlights the brain tumor segmentation techniques using ML, DL and XAI techniques.\u003c/p\u003e \u003cp\u003eIn recent years, many advances have been made in brain tumor segmentation and classification by using deep learning approaches combined with explainable AI methods to increase physician confidence. Zeineldin et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e28\u003c/span\u003e] used a variety of state-of-the-art XAI approaches, including Grad-CAM and SmoothGrad, coupled with ResNet-50 to evaluate brain gliomas and segment tumors. Their model exhibits a remarkable classification accuracy of 98.62%, with Dice scores of 84.10 (enhanced tumor), 87.33 (tumor core), and 92.00 (whole tumor). However, their research found that deep learning models in particular continue to lack interpretability with multimodal MRI data and highlighted that most XAI approaches are still designed for 2D MRI slices rather than entire 3D volumes. Yang et al. [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e29\u003c/span\u003e] proposed an ODE neural model with good segmentation accuracy (more than 99%) across all criteria. They emphasized the importance of specific MRI modalities \u0026ndash; T1-Ce for tumor enhancement and T2 and FLAIR for complete tumor segmentation. However, their results highlighted the need for additional data set validation and the high computational cost of the model.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTable II: Brain Tumor detection techniques.\u003c/b\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabc\" border=\"1\"\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSr. No\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eImaging Modality\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eML/DL/TF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eXAI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eResults\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eLimitations\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWindisch et al.,[ 20]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI T1-weighted, T2-weighted\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eResnet50, and Bayesian neural network\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTCGA-GBM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eGradCAM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eThe model achieved a categorical accuracy of 93%.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.The model showed a tendency to focus on irrelevant areas.\u003c/p\u003e \u003cp\u003e2.The MRI data used for training came from different views and acquisition sequences.\u003c/p\u003e \u003cp\u003e3.Limited Generalizability.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEsmaeili et al., [ 21]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI T1-weighted, T2-weighted, and FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDenseNet-121, GoogLeNet, and MobileNet,\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTCGA CIA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eGrad-CAM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eDenseNet-121 achieved the highest classification (92.1%) and localization accuracy (81.1% hits and 79.1% IoU.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.A significant number of tumor brains were incorrectly classified.\u003c/p\u003e \u003cp\u003e2.Grad-CAM struggled with complex images containing multiple objects.\u003c/p\u003e \u003cp\u003e3.The study used limited patient data.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUllah et al.,[22]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDeepEBTDNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBTDMRI database and the Brain MRI Images for BTD dataset.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLIME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eThe model achieved a validation accuracy of 98.96% and a testing accuracy of 94.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.The model is specifically designed for brain tumor detection and has not been validated on other types of medical images or conditions.\u003c/p\u003e \u003cp\u003e2.The study does not explicitly differentiate between benign and malignant tumors.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEder et al.,[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI T1-weighted, T2-weighted, FLAIR, and T1ce\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBraTS 2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSHAP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eThe model achieved up to 94% accuracy.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.The lack of additional datasets, particularly for network validation.\u003c/p\u003e \u003cp\u003e2.optimization of network structures was not explored in-depth.\u003c/p\u003e \u003cp\u003e3.Limited generalizability.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBrima et al., [24 ]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI T1-weighted\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eVGG16, VGG19, ResNet50, ResNet50V2, InceptionV3, Xception, EfficientNetB0, DenseNet121\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eFigshare and the COVID-19 X-ray\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSaliency maps\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eWith the brain MRI dataset, the DenseNet121 model achieved 98.10% accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Limited generalizability.\u003c/p\u003e \u003cp\u003e2.The Vanilla Gradient method and SmoothGrad produced noisy and less reliable saliency maps compared to XRAI.\u003c/p\u003e \u003cp\u003e3.No Quantitative Metrics for Explainability\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAppiah et al.,[ 25]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMobileNetV2, Inception-v3, ResNet101, VGG-19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBrats 2018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSHAP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.CNN predicted tumors with 99.21% accuracy.\u003c/p\u003e \u003cp\u003e2.MobileNetV2 delivered the highest accuracy of 99.82%, outperforming other models.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.The study focuses on binary classification only.\u003c/p\u003e \u003cp\u003e2.Limited generalizability.\u003c/p\u003e \u003cp\u003e3.The extensive pre-processing required.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTehsin et al., [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDaSAM Based on CNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eKaggle: Brain Tumor Classification and Figshare brain tumor dataset.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2D Grad Cam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eThe method achieved precision values of 99% and 96% on the Figshare and Kaggle datasets.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Limited generalizability.\u003c/p\u003e \u003cp\u003e2.The model focuses on classification rather than precise tumor segmentation,\u003c/p\u003e \u003cp\u003e3.The authors mention an imbalance in the data distribution.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePasvantis et al., [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e27\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eInceptionV3, ResNet50V2,NasNetLarge\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eKaggle:Brain Tumor Dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLIME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eResNet50V2 outperformed both InceptionV3 and NasNetLarge in terms of F1-scores, as confirmed by statistical tests (p\u0026thinsp;=\u0026thinsp;0.02 for InceptionV3 and p\u0026thinsp;=\u0026thinsp;0.008 for NasNetLarge).\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Edge detection algorithms sometimes produce unreliable masks, affecting the refinement quality.\u003c/p\u003e \u003cp\u003e2.Limited generalizability.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eSimilarly, Yan et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e30\u003c/span\u003e] presented a 3D nnU-Net model that was superior in segmentation, with high cube similarity coefficients and a classification accuracy of 95.46%. They used Grad-CAM\u0026thinsp;+\u0026thinsp;+\u0026thinsp;to create visual heatmaps that helped interpret the regions of interest in MRI data. Despite the impressive results, their model suffered from generalizability across datasets and required significant processing resources. Zeineldin et al. [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e31\u003c/span\u003e] extended their work by developing the TransXAI model, which used a hybrid architecture that included CNNs and transformers.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTable III Summary of Segmentation techniques.\u003c/b\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabd\" border=\"1\"\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSr.\u003c/p\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eImaging Modality\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eML/DL/TF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eXAI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eResults\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003elimitations\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZeineldin et al.,[28 ]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI native T1W, Gadolinium T1Gd, T2W, and FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eResNet-50,3D DeepSeg,Vanilla gradient, guided backpropagation,\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBraTS 2019 and 2021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eIntegrated gradients, Guided integrated gradients, SmoothGrad, Grad-CAM, and guided Grad-CAM.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.ResNet 50 achieved a classification accuracy of 98.62% for brain glioma grading.\u003c/p\u003e \u003cp\u003e2.For BT segmentation, the model achieved Dice scores of 84.10 (enhancing tumor), 87.33 (tumor core), and 92.00 (whole tumor).\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Lack of Interpretability in Deep Learning Models.\u003c/p\u003e \u003cp\u003e2.Current explainability methods struggle to effectively interpret complex multi-modal MRI data.\u003c/p\u003e \u003cp\u003e3.Many existing XAI methods are focused on 2D MRI slices.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYang et al.,[ 29]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI T1-weighted, T2-weighted, FLAIR, and T1ce\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNeural ODE, U-Net and Cascaded U-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBraTS 2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNeural Ordinary Differential Equation (ODE) model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eThe proposed model achieved accuracy across all segmentation tasks: ET (99.649%), TC (99.425%), and WT (99.383%).\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Limited generalizability.\u003c/p\u003e \u003cp\u003e2.Model requires significant computational resources.\u003c/p\u003e \u003cp\u003e3.The model is based on 2D MRI slices rather than full 3D volumes.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYan et al., [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e30\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI T1-weighted, T2-weighted, FLAIR, and T1ce\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3D nnU-Net model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBraTS 2018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eGrad Cam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.The nnU-Net model achieved high Dice Similarity Coefficients (ET: 0.79, WT: 0.91, TC: 0.85), demonstrating strong segmentation performance.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Required significant computational resources\u003c/p\u003e \u003cp\u003e2.Framework heavily rely on the availability of large datasets.\u003c/p\u003e \u003cp\u003e3.Limited generalizability.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eZeineldin et al.,[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e31\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI T1-weighted, T2-weighted, FLAIR, and T1ce\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN and Grad Cam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBraTS 2019 and FeTS2022 datasets.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eGrad-Cam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eThe TransXAI model achieved a Dice Similarity Coefficient:\u003c/p\u003e \u003cp\u003eEnhancing Tumor (ET): 0.745\u003c/p\u003e \u003cp\u003eTumor Core (TC): 0.782\u003c/p\u003e \u003cp\u003eWhole Tumor (WT): 0.882\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.The current implementation focuses on 2D axial MRI slices.\u003c/p\u003e \u003cp\u003e2.The hybrid architecture may require significant computational resources.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHassan et al.,[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e32\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI T1-weighted, T2-weighted, FLAIR, and T1ce\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDataset source is not mentioned.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNSL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eIt introduces the concept of neuro-symbolic learning as a solution to enhance the interpretability and explainability of AI in clinical practices, particularly in brain tumor segmentation .\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Paper proposed a strong case for NSL, it remains theoretical.\u003c/p\u003e \u003cp\u003e2.Dataset source is not mentioned.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThis model increased explainability even further by visualizing its decisions using Grad-CAM. However, as with previous studies, the implementation focused on two-dimensional axial MRI slices and the computational requirements remained significant. Finally, Hassan et al. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e32\u003c/span\u003e] proposed a more theoretical technique called Neuro-Symbolic Learning (NSL), which aims to improve interpretability by merging neural networks and symbolic AI. Although this methodology has the potential to improve the explainability of AI models in clinical practice, the work is still theoretical and the dataset used is not disclosed.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Brain Tumor Classification\u003c/h2\u003e \u003cp\u003eBrain tumor classification is a significant issue in medical imaging, particularly using magnetic resonance imaging, which is preferred due to its non-invasive nature and enhanced soft tissue visualization. Recent research has used various machine learning approaches, including convolutional neural networks (CNNs) and capsule networks (CapsNets), to improve classification accuracy. Table IV summarizes the selected works focusing on the classification of brain tumors using ML, DL and XAI approaches.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTable IV Summary of Brain tumor classification techniques.\u003c/b\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabe\" border=\"1\"\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSr.\u003c/p\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eImaging Modality\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eML/DL/TF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eXAI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eResults\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eLimitations\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eJin et al.,[33 ]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI T1-weighted, T2-weighted, FLAIR, and T1ce\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN,\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBraTS 2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eGrad Cam and SHAP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.The 16 examined heatmap algorithms failed to meet clinical requirements for indicating the AI model\u0026rsquo;s decision process or quality.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Current XAI Algorithms Are Not Clinically Designed.\u003c/p\u003e \u003cp\u003e2.Most XAI algorithms struggle to highlight the importance of different imaging modalities, such as T1, T2, or FLAIR.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDasanayaka et al.,[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e34\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI and WSI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDenseNet, ResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMRI sequences and WSI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eGrad-Cam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eThe proposed method achieved an accuracy of 86% during classification.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Source of dataset is not mentioned.\u003c/p\u003e \u003cp\u003e2.Limited generalizability.\u003c/p\u003e \u003cp\u003e3.The tool currently supports a limited number of imaging modalities.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMarmolejo et al.,[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e35\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI T1-weighted, T2-weighted, FLAIR, and T1ce\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBRATS 2017 .\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003enumGrad Cam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eThe proposed model achieved an accuracy of 97.11%, sensitivity of 95.58%, and specificity of 96.81%.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Limited generalizability.\u003c/p\u003e \u003cp\u003e2.There is a need for further validation.\u003c/p\u003e \u003cp\u003e3.The model requires significant computational resources.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMandloi et al., [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e36\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eConditional GAN, MobileNet, Inception ResNet, Efficient Net, and VGGNet, MobileNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eKaggle: Brain MRI Images for B T Detection\u003c/p\u003e \u003cp\u003eKaggle: B T Classification\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLRP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.InceptionResNetV2 achieved training, validation, and testing accuracies of 99.6%, 99.2%, and 99.0%, respectively, for tumor detection. \u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.InceptionResNetV2 is computationally expensive.\u003c/p\u003e \u003cp\u003e2.Validation thorugh additional dataset is required.\u003c/p\u003e \u003cp\u003e2.Limited generalizability.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eŠefč\u0026iacute;k et al., [ 37]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eT1-weighted and T2-weighted, and FLAIR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBraTS 2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLRP, Grad Cam\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eThe proposed method increased classification accuracy from 72\u0026ndash;79%.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Limited generalizability.\u003c/p\u003e \u003cp\u003e2.Model neglect other potentially relevant features in the MRI images that could contribute to classification.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBhuvaneswari et al.,[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e38\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eInceptionV3, ResNet-50, VGG16, DenseNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTCGA LGG Dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLRP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eResNet-50 achieved the highest accuracy (95.5%) with superior efficiency, followed by DenseNet (94.65%) and VGG16 (91.04%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Limited generalizability.\u003c/p\u003e \u003cp\u003e2.The model's performance heavily depends on the quality and diversity of the training dataset.\u003c/p\u003e \u003cp\u003e3.The model lacks comprehensive clinical validation.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNag et al.,[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e39\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMRI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eResNet50,Generative Adversarial Networks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eKaggle: Brain Tumor MRI dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLIME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eThe proposed model, TumorGANet, achieved exceptional results with an accuracy of 99.53%, precision and recall rates of 100%, F1-scores of 99%, and a Hamming loss of 0.2%.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e1.Limited generalizability.\u003c/p\u003e \u003cp\u003e2.There is a risk that the GAN-generated data could introduce artifacts or features that do not accurately represent real tumors.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eJin et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e33\u003c/span\u003e] investigated the limitations of current explainable AI (XAI) approaches in brain tumor classification. Despite testing numerous heatmap algorithms, they discovered that they failed to fulfill clinical criteria for decision transparency, leaving clinicians struggling to interpret the data. The study underscores the need for clinically driven advancements in XAI. Dasanayaka et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e34\u003c/span\u003e] created a web program that uses MRI and whole slide imaging (WSI) to classify brain cancers with 86% accuracy. Though it closely resembles genuine diagnostic workflows, the tool's limited generalizability and imaging modality support emphasize the difficulties of implementing AI in real-world settings. Marmolejo et al. [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e35\u003c/span\u003e] developed numGrad-CAM to improve explainability in brain tumor classification models and achieved 97.11% accuracy. Despite receiving encouraging feedback from clinicians, the model confronts limitations in generalizability and computing demands, necessitating additional validation.\u003c/p\u003e \u003cp\u003eMandloi et al. [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e36\u003c/span\u003e] employed InceptionResNetV2 and EfficientNet to classify brain tumors with above 99% accuracy. However, the high computational cost and the necessity for further validation limit its practical adoption. Šefč\u0026iacute;k et al. [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e37\u003c/span\u003e] suggested an approach that improved classification accuracy from 72\u0026ndash;79% using explanation-guided training. However, the model may omit relevant MRI features. Generalizability remains an issue. Bhuvaneswari et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e38\u003c/span\u003e] demonstrated the effectiveness of a hybrid CNN model, with ResNet-50 leading in accuracy (95.5%). The addition of LRP increased transparency, but the model's reliance on dataset quality and lack of clinical validation limit its general applicability. Nag et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e39\u003c/span\u003e] created TumorGANet, a GAN-based model capable of generating synthetic MRI data with excellent accuracy (99.53%). While highly effective, worries regarding potential artifacts in synthetic data indicate that more refinement is required for real-world application.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Explainable AI techniques used in Brain tumor detection, segmentation and Classification\u003c/h2\u003e \u003cp\u003eExplainable AI (XAI) methods play a critical role in improving the interpretability of models used for brain tumor detection, segmentation, and classification. Methods such as Grad-CAM and layer-wise relevance propagation (LRP) have been used to provide insights into deep learning models' decision-making processes, thereby solving the black box issue common in AI systems. For example, the Disease and Spatial Attention Model (DaSAM) uses attention mechanisms to highlight important aspects in MRI images, allowing for a better comprehension of tumor characteristics [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Furthermore, saliency maps have been used to detect important areas in brain MRIs, increasing the dependability and accountability of AI systems in clinical settings [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. These methods not only improve model transparency but also build confidence among medical practitioners, resulting in more quick and accurate diagnoses [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Table VI depicts the spectrum of explainability strategies employed in brain tumor detection, segmentation, and classification studies:\u003c/p\u003e \u003cp\u003e \u003cb\u003eTable V: Summery of XAI techniques used.\u003c/b\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabf\" border=\"1\"\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eXAI Method\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGradCam/Varient\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLRP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLIME\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSHAP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eOthers\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCount\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eGrad-CAM/Variants\u003c/strong\u003e \u003cp\u003eThe most commonly utilized approach, as seen in 11 investigations. Grad-CAM and its variants are well-known for their ability to show the sections of the input image that are most important to the model's prediction, making them valuable for understanding model decisions. Grad-CAM and its variants are popular explainability techniques in deep learning, particularly for interpreting CNNs in visual tasks such as medical image analysis. Grad Cam has several variations, including SmoothGrad, Grad-CAM, guided Grad-CAM, numGrad Cam, and 2D Grad Cam.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eSHAP (Shapley Additive Explanations\u003c/b\u003e): Used in three investigations. SHAP uses a unifying strategy to explain the output of machine learning models by assigning relevance ratings to each feature, which aids in evaluating the contributions of various input sources.\u003c/p\u003e \u003cp\u003e \u003cb\u003eLIME (Local Interpretable Model-Agnostic Explanations)\u003c/b\u003e is also used in three research studies. LIME is an explanation technique that localizes a complex model with an interpretable one, providing insights into specific predictions.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eLRP (Layer-wise Relevance Propagation)\u003c/strong\u003e \u003cp\u003eUsed in two studies. LRP explains model decisions by propagating prediction scores backward and highlighting key elements.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eOthers\u003c/strong\u003e \u003cp\u003eSaliency maps and the Neural Ordinary Differential Equation (ODE) model, which are less often utilized explanatory approaches, were used in two investigations.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eIn this section, we provide a detailed discussion of the key findings related to machine learning (ML), deep learning (DL), and explainable AI techniques for brain tumor detection, segmentation, and classification using MRI data. We also examine the role of Explainable AI (XAI) techniques, dataset availability, and common limitations in the current literature.\u003c/p\u003e \u003cp\u003e \u003cb\u003e4.1 Investigating the best current ML and DL techniques used in detection, segmentation, and classification of brain tumors using MRI.\u003c/b\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eConvolutional neural networks (CNNs)\u003c/b\u003e have become a foundation in brain tumor detection and classification using MRI. Widely used models like ResNet, DenseNet, VGG, and U-Net have consistently shown strong performance. For example, ResNet-50 achieved a high classification accuracy of 95.5% in a study by Ullah et al. [22], with DenseNet and VGG16 following closely behind.\u003c/p\u003e \u003cp\u003e \u003cb\u003eHybrid models, which\u003c/b\u003e combine CNNs with more advanced methods, such as generative adversarial networks (GANs) and vision transformers (ViTs), are gaining recognition from the scientific community due to their potential to improve detection and segmentation. A significant example is the TransXAI model, which combines CNNs and ViTs and has produced excellent results, particularly for tumor core and entire tumor segmentation.\u003c/p\u003e \u003cp\u003e \u003cb\u003eBayesian Neural Networks (BNNs)\u003c/b\u003e are also becoming increasingly widely employed, particularly for uncertainty estimates, which are important in medical decision-making. Windisch et al. [20] used a BNN with ResNet50 to achieve a categorical accuracy of 93%, providing insight into the model's prediction confidence.\u003c/p\u003e \u003cp\u003eFinally, \u003cb\u003eneural ODEs (ordinary differential equations)\u003c/b\u003e are a recent approach to tumor segmentation. They have had good results, with accuracy rates surpassing 99% for a variety of tumor locations, including the enhancing tumor, tumor core, and entire tumor.\u003c/p\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Investigating the XAI methods used in detection, segmentation, and classification of brain tumors using MRI.\u003c/h2\u003e \u003cp\u003e \u003cb\u003eExplainable AI (XAI)\u003c/b\u003e techniques are essential for making deep learning models more transparent, especially in critical areas like brain tumor detection and classification. Some of the most commonly used XAI methods include:\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eGradCAM and its variants\u003c/strong\u003e \u003cp\u003eGrad-CAM (Gradient-weighted Class Activation Mapping) and its upgraded versions, such as Guided Grad-CAM and Grad-CAM++, are widely used for determining which regions of an MRI scan are most relevant for a model's decision. These algorithms create heatmaps that emphasize tumor locations, allowing doctors to better understand how the model operates. Windisch et al. [20] and Esmaeili et al. [21] used Grad-CAM to explain CNN judgments, resulting in good classification accuracy.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eSHAP (Shapley Additive Explanations)\u003c/strong\u003e \u003cp\u003eSHAP assigns relevance scores to individual features, describing how each feature contributes to the model's prediction. For example, Appiah et al. [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e25\u003c/span\u003e] employed SHAP and MobileNetV2 to classify tumors, attaining great accuracy (99.82%) while keeping transparency in decision-making.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eLIME (Local Interpretable Model-Agnostic Explanations)\u003c/strong\u003e \u003cp\u003eLIME generates localized explanations for each situation, assisting in explaining specific predictions. Ullah et al. [22] used it to improve the interpretability of a variety of deep learning models, guaranteeing that the model's behavior is intelligible while maintaining accuracy.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eLayer-wise Relevance Propagation (LRP)\u003c/strong\u003e \u003cp\u003eLRP reveals pixel-level details on how neural networks generate judgments. Brima et al. [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e24\u003c/span\u003e] used LRP to determine which areas of an MRI image were most useful for tumor detection and classification.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Studying the impact of a code of paper that is publicly accessible.\u003c/h2\u003e \u003cp\u003eThe availability of code is critical to reproducibility and future study. According to our evaluation, the majority of the publications did not openly publish their code. As shown in Table VII, only six of the 20 reviewed publications specifically acknowledged their code and were made publicly available. The lack of open-source implementations hinders other researchers' capacity to reproduce or expand on the findings of these investigations, which is a typical difficulty in medical AI research. However, some recent studies have begun to share their code, indicating a favorable trend toward open science.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTable VI\u003c/b\u003e depicting the code availability count in research studies.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabg\" border=\"1\"\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCode Availability\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCount\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Investing the most widely used datasets in brain tumor detection, segmentation, and classification.\u003c/h2\u003e \u003cp\u003eSeveral well-established datasets have been consistently used for brain tumor analysis in the literature. Based on the provided data, the most widely used datasets in brain tumor detection, segmentation, and classification are:\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eBraTS (2017, 2018, 2019, 2020, 2021)\u003c/strong\u003e \u003cp\u003eThis dataset appears frequently and is the most popular for brain tumor segmentation tasks.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eTCGA (GBM, LGG, CIA)\u003c/strong\u003e \u003cp\u003eThe Cancer Genome Atlas datasets are widely used for segmentation and classification applications.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eKaggle Brain Tumor Datasets\u003c/strong\u003e \u003cp\u003eSeveral studies use Kaggle datasets, including \"Brain MRI Images for Brain Tumor Detection\" and \"Brain Tumor Classification (MRI).\"\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eFigshare\u003c/strong\u003e \u003cp\u003eBrain tumor detection dataset is used in study.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eBTDMRI\u003c/strong\u003e \u003cp\u003eAnother dataset used for MRI tumor identification.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eThese datasets, particularly BraTS and Kaggle, are fundamental to brain tumor analysis research.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Summarizing the common limitations of existing studies.\u003c/h2\u003e \u003cp\u003eDespite the progress in deep learning for brain tumor detection, segmentation, and classification, several challenges remain:\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eGeneralization Issues\u003c/strong\u003e \u003cp\u003eMany models, such as BraTS, perform well on some datasets but struggle on others, indicating a lack of generalizability, which is critical for real-world clinical usage.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eLimited Interpretability\u003c/strong\u003e \u003cp\u003eAlthough techniques such as Grad-CAM and LRP provide some interpretability, the explanations might be ambiguous and may not always match clinical standards. Heatmaps frequently do not correspond to what medical practitioners expect.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eHigh computational demands\u003c/strong\u003e \u003cp\u003eAdvanced models, particularly those based on hybrid methods such as GANs or Transformers, necessitate a significant amount of processing resources, making them challenging to apply in real-time or in resource-constrained environments, such as smaller medical institutions.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eLimited Code Availability\u003c/strong\u003e \u003cp\u003eMany researchers do not make their code available to the public, making it difficult to duplicate and validate results. The absence of standardization throughout research impacts comparisons of different studies.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eData Imbalance\u003c/strong\u003e \u003cp\u003eMany brain tumor databases contain class imbalances, with fewer examples of unusual tumor forms. This makes it more difficult for models to perform well across all classes, sometimes resulting in lower outcomes for underrepresented tumor classifications.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003eThis study focuses on advances in applying machine learning (ML) and deep learning (DL) to detect, segment, and classify brain tumors using MRI. CNN-based models like ResNet, DenseNet, and U-Net have proven to be extremely effective, while newer hybrid models that combine CNNs with GANs and Vision Transformers improve performance even further. Explainable AI (XAI) approaches such as Grad-CAM, SHAP, and LIME have increased model interpretability by visualizing decision-making processes. However, these explanations remain insufficiently thorough to meet clinical needs, restricting their applicability in real-world healthcare settings. Despite advancements, some difficulties persist. A major difficulty is that many models perform well on specialized datasets, such as BraTS, but struggle to generalize to other ones. The complexity of some models also makes them too resource-intensive for real-time use in many hospitals. Furthermore, only six of the 20 research studies made their code publicly available, making it difficult to replicate results and progress the area. BraTS, TCGA, and Kaggle are the most extensively utilized datasets in this field, and while they provide rich data, they also suffer from data imbalance, particularly for rarer tumor types. In conclusion, while deep learning models show great potential for brain tumor analysis, they still confront challenges in interpretability, generalizability, and practical application. Addressing these issues and making models more accessible will be critical in bringing these technologies closer to routine clinical use.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAl-Galal, S.A.Y.; Alshaikhli, I.F.T.; Abdulrazzaq, M.M. MRI brain tumor medical images analysis using deep learning techniques:A systematic review. Health Technol. 2021, 11, 267\u0026ndash;282.\u003c/li\u003e\n\u003cli\u003eLouis, D.N.; Perry, A.; Wesseling, P.; Brat, D.J.; Cree, I.A.; Figarella-Branger, D.; Hawkins, C.; Ng, H.K.; Pfister, S.M.;Reifenberger, G. The 2021 WHO classification of tumors of the central nervous system: A summary. Neuro-Oncology 2021, 23,1231\u0026ndash;1251.\u003c/li\u003e\n\u003cli\u003eNodirov, J.; Abdusalomov, A.B.; Whangbo, T.K. Attention 3D U-Net with Multiple Skip Connections for Segmentation of Brain Tumor Images. Sensors 2022, 22, 6501.\u003c/li\u003e\n\u003cli\u003ehttps://www.cancer.net/cancer-types/brain-tumor/statistics\u003c/li\u003e\n\u003cli\u003eC. Wang et al., \u0026ldquo;Phenotypic and genetic associations of quantitative magnetic susceptibility in U.K. biobank brain imaging,\u0026rdquo; Nature Neurosci., vol. 562, pp. 1\u0026ndash;14, May 2022.\u003c/li\u003e\n\u003cli\u003eZ. Liu et al., \u0026ldquo;Deep learning based brain tumor segmentation: A survey,\u0026rdquo; Complex Intell. Syst., vol. 9, no. 1, pp. 1001\u0026ndash;1026, 2020.\u003c/li\u003e\n\u003cli\u003eM. A. Ottom, H. A. Rahman, and I. D. Dinov, \u0026ldquo;ZNet: Deep learning approach for 2D MRI brain tumor segmentation,\u0026rdquo; IEEE J. Transl. Eng. Health Med., vol. 10, May 2022, Art. no. 1800508.\u003c/li\u003e\n\u003cli\u003eH. S. Abdulbaqi, K. N. Mutter, M. Z. M. Jafri, and Z. A. Al-Khafaji, \u0026ldquo;Estimation of brain tumour volume using expanded computed tomography scan images,\u0026rdquo; in Proc. IEEE 23rd Iranian Conf. Biomed. Eng. 1st Int. Iranian Conf. Biomed. Eng., 2016, pp. 117\u0026ndash;121.\u003c/li\u003e\n\u003cli\u003eA. Barredo Arrieta, N. D\u0026iacute;az-Rodr\u0026iacute;guez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garcia, S. Gil-Lopez, D. Molina, R. Benjamins, R. Chatila, F. Herrera, Explainable Artificial Intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI, Inf. Fusion 58 (2020) 82\u0026ndash;115.\u003c/li\u003e\n\u003cli\u003eM. Ahamed, A. Imran, Joint learning with local and global consistency for improved medical image segmentation, in: Annual Conference on Medical Image Understanding and Analysis, 2022.\u003c/li\u003e\n\u003cli\u003eA. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2010. \u003c/li\u003e\n\u003cli\u003eR. Strudel, R. Garcia, I. Laptev, C. Schmid, Segmenter: transformer for semantic segmentation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021\u003c/li\u003e\n\u003cli\u003eH, A., MO, D., \u0026amp; B, S. (2023). Current challenges of the state-of-the-art of AI techniques for diagnosing brain tumor. Material Science \u0026amp; Engineering International Journal, 7(4), 196-208. https://doi.org/10.15406/mseij.2023.07.00224\u003c/li\u003e\n\u003cli\u003eAli, S., Abuhmed, T., El-Sappagh, S.,et al. (2023). Explainable artificial intelligence (XAI): What we know and what is left to attain trustworthy artificial intelligence. Information Fusion, 99, 101805. https://doi.org/10.1016/j.inffus.2023.101805\u003c/li\u003e\n\u003cli\u003ePage MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021;372:n71.\u003c/li\u003e\n\u003cli\u003eGillies, R. J., Kinahan, P. E., \u0026amp; Hricak, H. (2016). Radiomics: Images are more than pictures, they are data. Radiology, 278(2), 563-577.\u003c/li\u003e\n\u003cli\u003eZhou, S. K., Greenspan, H., \u0026amp; Shen, D. (2017). Deep learning for medical image analysis. Academic Press.\u003c/li\u003e\n\u003cli\u003eZhou, M., Scott, J., Chaudhury, B., Hall, L., Goldgof, D., Yeom, K., Iv, M., \u0026amp; Erickson, B. (2018). Radiomics in brain tumor: image assessment, quantitative feature descriptors, and machine-learning approaches. AJNR American Journal of Neuroradiology, 39(2), 208-216.\u003c/li\u003e\n\u003cli\u003eLitjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A., Ciompi, F., Ghafoorian, M., \u0026amp; van der Laak, J. A. (2017). A survey on deep learning in medical image analysis. Medical Image Analysis, 42, 60-88.\u003c/li\u003e\n\u003cli\u003eCheng, J., Huang, W., Cao, S., Yang, R., Yang, W., Yun, Z., Wang, Z., \u0026amp; Feng, Q. (2016). Enhanced performance of brain tumor classification via tumor region augmentation and partition. PLOS ONE, 10(10), e0140381.\u003c/li\u003e\n\u003cli\u003eWindisch, P., Weber, P., F\u0026uuml;rweger, C., Ehret, F., Kufeld, M., Zwahlen, D., \u0026amp; Muacevic, A. (2020). Implementation of model explainability for a basic brain tumor detection using convolutional neural networks on MRI slices. Neuroradiology, 62(11), 1515-1518. https://doi.org/10.1007/s00234-020-02465-1\u003c/li\u003e\n\u003cli\u003eEsmaeili, M., Vettukattil, R., Banitalebi, H., Krogh, N. R., \u0026amp; Geitung, J. T. (2021). Explainable artificial intelligence for human-machine interaction in brain tumor localization. Journal of Personalized Medicine, 11(11), 1213. https://doi.org/10.3390/jpm11111213\u003c/li\u003e\n\u003cli\u003eUllah, N., Hassan, M., Khan, J. A., Anwar, M. S., \u0026amp; Aurangzeb, K. (2023). Enhancing explainability in brain tumor detection: A novel DeepEBTDNet model with LIME on MRI images. International Journal of Imaging Systems and Technology, 34(1). https://doi.org/10.1002/ima.23012\u003c/li\u003e\n\u003cli\u003eEder, M., Moser, E., Holzinger, A., Jean-Quartier, C., \u0026amp; Jeanquartier, F. (2022). Interpretable machine learning with brain image and survival data. BioMedInformatics, 2(3), 492-510. https://doi.org/10.3390/biomedinformatics2030031\u003c/li\u003e\n\u003cli\u003eBrima, Y., \u0026amp; Atemkeng, M. (2023). Visual interpretable and explainable deep learning models for brain tumor MRI and COVID-19 chest X-ray images. https://doi.org/10.21203/rs.3.rs-3241888/v1\u003c/li\u003e\n\u003cli\u003eAppiah, R., Pulletikurthi, V., Esquivel-Puentes, H. A., Cabrera, C., Hasan, N. I., Dharmarathne, S., Gomez, L. J., \u0026amp; Castillo, L. (2024). Brain tumor detection using proper orthogonal decomposition integrated with deep learning networks. Computer Methods and Programs in Biomedicine, 250, 108167. https://doi.org/10.1016/j.cmpb.2024.108167\u003c/li\u003e\n\u003cli\u003eTehsin, S., Nasir, I. M., Dama\u0026scaron;evičius, R., \u0026amp; Maskeliūnas, R. (2024). DaSAM: Disease and spatial attention module-based explainable model for brain tumor detection. Big Data and Cognitive Computing, 8(9), 97. https://doi.org/10.3390/bdcc8090097\u003c/li\u003e\n\u003cli\u003ePasvantis, K., \u0026amp; Protopapadakis, E. (2024). Enhancing deep learning model Explainability in brain tumor datasets using post-heuristic approaches. Journal of Imaging, 10(9), 232. https://doi.org/10.3390/jimaging10090232\u003c/li\u003e\n\u003cli\u003eZeineldin, R. A., Karar, M. E., Elshaer, Z., Coburger, \u0026middot;., Wirtz, C. R., Burgert, O., \u0026amp; Mathis-Ullrich, F. (2022). Explainability of deep neural networks for MRI analysis of brain tumors. International Journal of Computer Assisted Radiology and Surgery, 17(9), 1673-1683. https://doi.org/10.1007/s11548-022-02619-x\u003c/li\u003e\n\u003cli\u003eYang, Z., Hu, Z., Ji, H., Lafata, K., Vaios, E., Floyd, S., Yin, F., \u0026amp; Wang, C. (2023). A neural ordinary differential equation model for visualizing deep neural network behaviors in multi‐parametric MRI‐based glioma segmentation. Medical Physics, 50(8), 4825-4838. https://doi.org/10.1002/mp.16286\u003c/li\u003e\n\u003cli\u003eYan, F.; Chen, Y.; Xia, Y.; Wang, Z.; Xiao, R. An Explainable Brain Tumor Detection Framework for MRI Analysis. Appl. Sci. 2023, 13, 3438. https://doi.org/10.3390/app13063438\u003c/li\u003e\n\u003cli\u003eZeineldin, R. A., Karar, M. E., Elshaer, Z., Coburger, J., Wirtz, C. R., Burgert, O., \u0026amp; Mathis-Ullrich, F. (2024). Explainable hybrid vision transformers and convolutional network for multimodal glioma segmentation in brain MRI. Scientific Reports, 14(1). https://doi.org/10.1038/s41598-024-54186-7\u003c/li\u003e\n\u003cli\u003eHassan, M., Fateh, A. A., Lin, J., Zhuang, Y., Lin, G., Xiong, H., You, Z., Qin, P., \u0026amp; Zeng, H. (2024). Unfolding explainable AI for brain tumor segmentation. Neurocomputing, 599, 128058. https://doi.org/10.1016/j.neucom.2024.128058\u003c/li\u003e\n\u003cli\u003eJin, W., Li, X., \u0026amp; Hamarneh, G. (2022). Evaluating explainable AI on a multi-modal medical imaging task: Can existing algorithms fulfill clinical requirements? Proceedings of the AAAI Conference on Artificial Intelligence, 36(11), 11945-11953. https://doi.org/10.1609/aaai.v36i11.21452\u003c/li\u003e\n\u003cli\u003eDasanayaka, S., Shantha, V., Silva, S., Meedeniya, D., \u0026amp; Ambegoda, T. (2022). Interpretable machine learning for brain tumour analysis using MRI and whole slide images. Software Impacts, 13, 100340. https://doi.org/10.1016/j.simpa.2022.100340\u003c/li\u003e\n\u003cli\u003eMarmolejo-Saucedo, J. A., \u0026amp; Kose, U. (2022). Numerical grad-cam based explainable Convolutional neural network for brain tumor diagnosis. Mobile Networks and Applications. https://doi.org/10.1007/s11036-022-02021-6\u003c/li\u003e\n\u003cli\u003eMandloi, S., Zuber, M., \u0026amp; Gupta, R. K. (2023). An explainable brain tumor detection and classification model using deep learning and layer-wise relevance propagation. Multimedia Tools and Applications, 83(11), 33753-33783. https://doi.org/10.1007/s11042-023-16708-9\u003c/li\u003e\n\u003cli\u003e\u0026Scaron;efč\u0026iacute;k, F., \u0026amp; Benesova, W. (2023). Improving a neural network model by explanation-guided training for glioma classification based on MRI data. International Journal of Information Technology, 15(5), 2593-2601. https://doi.org/10.1007/s41870-023-01289-5\u003c/li\u003e\n\u003cli\u003eBhuvaneswari Ramakrishnan, A., Sridevi, M., Vasudevan, S. K., Manikandan, R., \u0026amp; Gandomi, A. H. (2024). Optimizing brain tumor classification with hybrid CNN architecture: Balancing accuracy and efficiency through oneAPI optimization. Informatics in Medicine Unlocked, 44, 101436. https://doi.org/10.1016/j.imu.2023.101436\u003c/li\u003e\n\u003cli\u003eNag, A., Mondal, H., Mehedi Hassan, M., Al-Shehari, T., Kadrie, M., Al-Razgan, M., Alfakih, T., Biswas, S., \u0026amp; Kumar Bairagi, A. (2024). TumorGANet: A transfer learning and generative adversarial network- Based data augmentation model for brain tumor classification. IEEE Access, 12, 103060-103081. https://doi.org/10.1109/access.2024.3429633\u003c/li\u003e\n\u003cli\u003eKeles, A., Akcay, O., Kul, H., \u0026amp; Bendechache, M. (2023, August 1). Saliency Maps as an Explainable AI Method in Medical Imaging: A Case Study on Brain Tumor Classification. Irish Mahine Vision and Image Processing Conference 2023 (IMVIP2023), University of Galway, Ireland. https://doi.org/10.5281/zenodo.8199333\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Guru Nanak Dev Engineering College, Ludhiana","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":"Machine Learning, Deep Learning, Explainable AI, Brain tumor Detection, Segmentation, Classification","lastPublishedDoi":"10.21203/rs.3.rs-5580195/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5580195/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eBrain tumors, caused by abnormal tissue growth within the brain, can severely disrupt brain functions and pose significant health risks. As the tumor progresses to higher stages, the patient's prognosis and survival decrease, resulting in a high mortality rate. With the advancements in medical imaging, especially the use of MRI, AI approaches have emerged as strong tools for detecting, segmenting, and classifying brain cancers. CNN and hybrid models, such as Vision Transformers (ViTs), have produced promising findings in this area. Although AI models exhibit high accuracy, they suffer from a lack of transparency and interpretability, paving the way for the development of eXplainable AI (XAI) methods in brain disease diagnosis. This paper investigates the utilization of machine learning, deep learning, and explainable AI (XAI) in brain tumor detection, segmentation, and classification. In this study, we have utilized the Preferred Reporting Items for Systematic Reviews and Meta-Analyses checklist and diagram. Peer-reviewed articles from PubMed, IEEE Explore, ScienceDirect, Google Scholar, Springer, and Wilay online libraries were searched, and only those papers were selected that were published in Scopus, SCIE, and ESCI-indexed journals. We have identified the 20 research papers published between 2020 and 2024 that used machine learning, deep learning and explainable AI to detect, segment, and classify the brain tumor. This review provides a comprehensive survey the of explainable artificial intelligence (XAI) in biomedical imaging, focusing on its role in the detection, segmentation and classification of brain tumors. It examines various machine learning, deep learning and XAI techniques, addresses current challenges, and suggests future directions. The objective is to provide clinicians, regulators and AI developers with valuable insights to improve the transparency and reliability of these methods in medical diagnostics.\u003c/p\u003e","manuscriptTitle":"Explainable AI in brain tumor diagnosis: A critical review of ML and DL techniques","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-12-11 14:24:50","doi":"10.21203/rs.3.rs-5580195/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":"8f49dba5-4476-4788-83a4-7aa8bf66e0b6","owner":[],"postedDate":"December 11th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":41150297,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2024-12-11T14:24:50+00:00","versionOfRecord":[],"versionCreatedAt":"2024-12-11 14:24:50","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5580195","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5580195","identity":"rs-5580195","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}