Traumatic Meningeal Enhancement Detection by Deep Learning-based Biomedical Image Analysis and Handcrafted Features Extraction

Traumatic Meningeal Enhancement Detection by Deep Learning-based Biomedical Image Analysis and Handcrafted Features Extraction | 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 Short Report Traumatic Meningeal Enhancement Detection by Deep Learning-based Biomedical Image Analysis and Handcrafted Features Extraction Proloy Kanti Roy, Mohammad Sakib Uddin This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7383303/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 Traumatic Meningeal Enhancement (TME) is characterized by abnormal enhancement of the meninges following head trauma, detectable on MRI even without other cerebral abnormalities. Causes include falls, motor vehicle acci- dents, sports injuries, and assaults. This study investigates advanced imaging for TME detection using a customized 13-layer convolutional neural network (CNN) compared to transfer learning and handcrafted feature extraction methods. A data set of 7,800 MRI images was enhanced and classified into normal, early (pre), and acute (post) categories. Seventy percent of the data was used for training and 30% for testing. Four transfer learning models, VGG19, VGG16, InceptionV3, and MobileNet, achieved accuracies of 84%, 86%, 80%, and 89%, respectively; their ensemble achieved 88.83%. Handcrafted feature extraction using positional orientation (PO), histogram of oriented gradients (HOG), and mean pixel value (MPV) was tested with Support Vector Machine (SVM) and XGBoost, yield- ing lower accuracies (SVM: 65–81%; XGBoost: 57–72%). The customized CNN achieved the highest accuracy at 91%, outperforming all other models. These results demonstrate the potential of the proposed CNN for accurate and reliable identification of TME in medical imaging. TME CNN Pre-trained feature extraction Image Processing Transfer Learning Figures Figure 1 Figure 2 Figure 3 Figure 4 1 Introduction Detecting Traumatic Meningeal Enhancement (TME) remains a significant challenge in medical imaging. Current diagnosis depends heavily on subjective radiologist inter- pretation, which is time-consuming and prone to human error and inter-operator variability. TME often presents subtly, appearing as a biomarker on FLAIR MRI after post-contrast administration, even without other cerebral abnormalities. Prior studies [ 1 ] [ 2 ] have highlighted differences in diagnostic performance among MRI sequences. Convolutional Neural Networks (CNNs) are well-suited for complex brain imag- ing tasks, capable of automatically learning hierarchical patterns from raw MRI data, including FLAIR and contrast-enhanced images [ 3 ] [ 4 ]. CNN-based segmentation offers advantages over manual feature extraction and traditional machine learning by reduc- ing dependence on handcrafted features, adapting to diverse datasets, and enabling end-to-end learning. Depth in CNNs enhances abstract feature extraction but also increases risks of overfitting and computational cost. 1.1 Limitations in Current Research TME detection research faces several constraints: Scarcity of annotated datasets limits model robustness and generalizability. Complex lesion patterns—subtle enhancements and irregular shapes—complicate detection and segmentation. Interpretability of CNNs is low, which poses barriers to clinical adoption where transparency is vital. High computational demand for deep architectures makes the deployment challeng- ing. Inconsistent image quality across datasets reduces reliability. To address binary classification challenges, prior work has applied deep neural networks such as VGG16, showing improved accuracy but requiring substantial com- puting resources. This study builds on such approaches with a customized CNN tailored for TME detection. 1.2 Research Motivation Advances in AI have expanded opportunities for computer-based clinical decision support, enabling more precise diagnoses and personalized treatment planning. Our motivation is to develop an automated, accurate, and reliable method for detecting TME that supports clinicians by: Understanding and applying image processing principles. Using preprocessing techniques such as augmentation and reshaping to enhance datasets. Developing a CNN-based model to classify MRI images into normal, early (pre), and acute (post) TME categories. Leveraging deep learning to improve diagnostic performance. Assisting neurologists in faster, more precise detection through automated analysis. 1.3 Research Objective The primary aim is to implement a CNN-based detection and classification system for TME that integrates transfer learning with deep learning. Our proposed 13-layer CNN is designed to identify subtle features in MRI images and categorize them into normal, early (pre), and acute (post) stages. By automating detection, we aim to reduce subjectivity, increase diagnostic speed, and improve accuracy compared with manual interpretation. TME’s subtle and multifaceted presentation often delays intervention, potentially worsening patient outcomes. Manual assessment by radiologists is not only subjective but also limited by individual expertise and workload constraints. An automated, objective, and precise approach could transform the diagnostic process, improving both timeliness and reliability. Our system processes MRI scans through multiple stages: image acquisition, preprocessing, and classification using the proposed CNN model. This approach is expecvarious local hospitals that employ contrast-enhanced MRI to assess TME in patients with suspected traumatic brain injury (TBI). Ultimately, the system is intended to enhance clinical decision-making, reduce diagnostic variability, and improve outcomes for patients with traumatic brain injury. 2 Methodology This study applies deep learning and handcrafted feature extraction to detect trau- matic meningeal enhancement (TME) in biomedical images. Using a dataset from different local hospitals, which employed contrast-enhanced MRI to assess TME in suspected traumatic brain injury (TBI) patients, we developed a hybrid detection approach. Our method combines learned representations from deep learning with domain-specific features to improve accuracy and efficiency in TME detection. We implemented a 13-layer custom CNN and evaluated four pre-trained CNN models (InceptionV3, VGG16, VGG19, MobileNetV2) using transfer learning and ensemble techniques. Additionally, handcrafted features—Histogram of Oriented Gra- dients (HOG), Mean Pixel Value (MPV), and Positional Orientation (PO)—were extracted and used to train Support Vector Machine (SVM) and XGBoost models. Performance was assessed using accuracy, precision, recall, and F1-score. The cus- tomized CNN outperformed both pre-trained models and models using handcrafted features, underscoring the value of deep learning-based biomedical image analysis for TME detection. 2.1 Dataset and Pre-Processing The dataset, originally in DICOM format, obtained from publicly available sources [ 6 ], was converted to JPG using the Pydicom and Pillow libraries. A total of 7,800 images were prepared, with 6,000 used for training and 1,800 for testing. Pre-processing aimed to reduce inconsistencies and prepare the data for optimal model performance. Key steps included: Image Conversion: DICOM files were read, pixel arrays extracted, rescaled to 0–255 (uint8), and saved as JPG images. Resizing: All images were scaled to 128×128 pixels to match the input requirements of both the transfer learning models and the 13-layer CNN. Augmentation: To address the limited dataset size, various augmentation techniques were applied, including rescaling, shearing, zooming, shifting (width/height), and flipping (vertical/horizontal), thereby increasing training diversity and improving model generalization. These pre-processing steps ensured consistent image dimensions, expanded train- ing samples, and enhanced the dataset’s suitability for deep learning-based TME detection. 2.2 Architecture of Proposed Model We used the Keras neural network toolbox to build a sequential CNN model for the given system. The correctness of the model was examined. Thirteen layers make up the model. Additionally, there were three max-pooling layers, three 2D convolutional layers, and an equal number of batch normalization layers. Two layers of dense paint, then one layer of flattening and a dropout layer. Input Layer : This is the first convolutional layer, and it uses the Rectified Linear Unit (ReLU) activation function with 32 filters of size ( 3 , 3 ). The input shape is supplied as (img height, img width, 3), giving the input images’ height, width, and colour channels Pooling and Normalization Layers : The code repeats a similar pattern twice for additional convolutional blocks. The second convolutional block uses 32 filters, and the third one uses 64 filters. Each block is followed by max pooling and batch normalization. Flatten Layer : This layer flattens the output from the preceding convolutional blocks, transforming it into a one-dimensional vector. It provides the data for the completely interconnected layers. Fully Connected Layers : The fully linked layer features ReLU activation function and 256 neurons. With a rate of 0.5 for dropout, half of the neurons in this layer will be arbitrarily set to zero at each update during training. This aids in avoiding overfitting. Three neurons in the last layer, which represent the three classes in the classification job, are part of a densely linked output layer. The probability distributions over the classes are obtained using the softmax activation function. Table 1 13-Layer Custom CNN Model Layers Output Shape Parameter conv2d (Conv2D) (None, 126, 126, 32) 896 max pooling2d (MaxPooling2D) (None, 63, 63, 32) 0 conv2d 1 (Conv2D) (None, 61, 61, 32) 9248 max pooling2d 1 (MaxPooling2D) (None, 30, 30, 32) 0 conv2d 2 (Conv2D) (None, 28, 28, 64) 18496 max pooling2d 2 (MaxPooling2D) (None, 14, 14, 64) 0 flatten (Flatten) (None, 12544) 0 dense (Dense) (None, 256) 3211520 dropout (Dropout) (None, 256) 0 dense 1 (Dense) (None, 3) 771 2.3 Models Architecture We implemented a custom 13-layer CNN alongside transfer learning using VGG16, VGG19, InceptionV3, and MobileNetV2. VGG16 and VGG19 : Deep networks (16 and 19 layers) using 3×3 convolutions, max pooling, and ReLU activations, taking 128×128 RGB inputs. InceptionV3 : Optimized with smaller convolutions, auxiliary classifiers, and efficient grid-size reduction to lower parameter count and improve speed. MobileNetV2 : Lightweight architecture using depthwise separable convolutions for computational efficiency. An ensemble model averaged predictions from the four pre-trained networks to enhance accuracy. 2.4 Handcrafted Feature Models Histogram of Oriented Gradients (HOG), Mean Pixel Value (MPV), and Positional Orientation (PO) features were extracted and used to train: XGBoost : Gradient boosting trees with regularization and subsampling for robustness. SVM : Support Vector Classifier tuned for multi-class prediction. Model performance was evaluated using accuracy, precision, recall, and F1-score, with the custom CNN outperforming pre-trained models on handcrafted feature datasets. 3 Performance Study 3.1 Implementation Model training was regulated by parameters such as size, epochs, layers, groups, trans- former layers, filters, and matrix settings. After preprocessing, training began using convolutional layers followed by max pooling. Pre-trained models and a custom CNN were employed, classifying images into early, acute, and normal categories. Of 7,800 images, 76.9% (6,000) were used for training and 23.1% for testing. The Adam optimizer was selected for its efficiency, low memory needs, and adapt- ability to large datasets. Training ran for 35 epochs with a batch size of 32, shuffling training and validation data between epochs. All images were RGB. ReLU activa- tion was applied to capture complex biomedical image features, while categorical cross-entropy loss optimized multiclass classification accuracy. Class mode was set to “categorical,” predicting enhancement severity. Regularization was implemented using dropout to reduce overfitting. The soft- max output layer produced probability distributions for classification. Training ran on Google Colab GPUs (Tesla K80, T4, P4), with hardware details accessible via “!nvidia-smi. The training process is examined with reference to the optimizer’s repetition as a deep learning technique. In general, the ”Adam” [ 5 ] optimizer function is described by Eq. (1) ω t +1 = ω t − αm t ( 1 ) m t = aggregate of gradients at time t α = learning rate at time t ω t = weights at time t ω t +1 = weights at time t + 1 Positive values are retained, while negative values are converted to zero using the Rectified Linear Unit (ReLU) activation function. f ( x ) = max(0, x ) ( 2 ) x represents the input to the ReLU function. f ( x ) denotes the output of the ReLU function. The function f ( x ) returns x if x is positive or zero; otherwise, it returns zero. Table 2 Parameters Used for the Pre-trained Models and the 13-Layer CNN Model Parameter Pre-trained models 13-layered CNN Models Train Data 70% 70% Test Data 30% 30% Target Size (128,128) (128,128) Batch Size 32 32 Epoch 35 35 Environment of Execution GPU GPU Optimizer Adam Adam Loss Function Categorical CrossEntropy Categorical CrossEntropy Activation Function Softmax Softmax Class Mode Categorical Categorical Colour Mode RGB RGB y t is the true label distribution vector in a one-hot encoded form for the three classes. y p is the predicted probability distribution vector across all three classes. y t,i , y p,i represent the true and predicted probabilities for each of the three classes. Training is initiated by clicking the ”model. fit” in the code. Below the contents of the cell is a progress indicator showing the stages and epoch of the entire procedure. The pre-trained model with all of the parameters and matrices selected indicates that the primary validation parameter has been completed when the designated time is completed. 3.2 Performance Matrices Here, the calculation of accuracy, precision, recall, and AUC performance of each model is stated [ 1 ]. The equations are described below: AUC: A measure of a classifier’s capacity to distinguish between classes is referred to as the Area Under the Curve (AUC) [1]. It just expresses the degree of in dependence or the scale by which it is assessed. Given that AUC is scale and classification-threshold-invariant, it provides an aggregated performance statis- tic across all thresholds. How well the model distinguishes between positive and nega tive categories is shown by the AUC rate. An increased AUC rate indicates improved performance for each specific model. In the following graphs, the x-axis defines the number of epochs, and the y-axis defines accuracy rate 3.3 Performance of CNN Model We selected 7800 images among them, we used 6000 images for training and 1800 images for testing, separated into three portions for normal brain, early (pre), and acute (post). Each class’s training data accounts for 76.9% of the total data, while testing data accounts for 23.1%. Finally, our proposed model was found to be 91% correct. 3.4 Performance of Pre-Trained Models A total of 7,800 images were used, divided into three classes—normal, early (pre), and acute (post), with 1,800 images (600 per class) allocated for testing and the remainder for training. Four pre-trained models, MobileNet, InceptionV3, VGG16, and VGG19 were evaluated. MobileNet achieved the highest performance among them, with 95% training accuracy and 89% testing accuracy. VGG16 achieved 86% accuracy, VGG19 reached 84%, and InceptionV3 achieved 80%. Model performance was assessed through training/testing accuracy and loss curves, as well as confusion matrices, which illustrated classification effectiveness across the three categories. 3.5 Performance of Handcrafted Models We evaluated handcrafted features (HOG, MPV, PO) using SVM and XGBoost classifiers. SVM : HOG 77.5%, MPV 70.4%, PO 66.3% XGBoost : HOG 70.3%, MPV 63.8%, PO 60.1% HOG with SVM achieved the highest accuracy among handcrafted feature methods. 4 Result Pre-trained models achieved the following accuracies: VGG16 86%, VGG19 84%, InceptionV3 80%, and MobileNet 89%. An ensemble of these models reached 88%, while our custom 13-layer CNN achieved 91%. Handcrafted features (HOG, MPV, PO) with SVM yielded 80%, 70%, and 66%, respectively; with XGBoost, 71%, 66%, and 62%. All alternatives performed below our lightweight custom CNN, confirming its superior accuracy on this dataset. 5 Conclusion This study presents a custom 13-layer CNN model for the automatic classification and detection of traumatic meningeal injuries, achieving 91% accuracy on a dataset of 7,800 MRI images categorized as normal, early (pre), and acute (post). The proposed model consistently outperformed well-known transfer learning architectures (VGG16, VGG19, MobileNet, InceptionV3) and handcrafted feature-based methods (HOG, MPV, PO with SVM and XGBoost), demonstrating its robustness and adaptability. By enabling accurate and timely detection, the model has the potential to support early diagnosis, guide treatment decisions, and improve patient outcomes. The results highlight the advantages of a lightweight yet effective deep learning architecture in medical image analysis. Future research will aim to optimize the model’s efficiency, expand dataset diversity, and explore integration into clinical workflows for real-world application. Declarations • Funding - Not Applicable • Conflict of interest/Competing interests - The authors declare that they have no competing interests. • Ethics approval and consent to participate - This study used publicly available, anonymized MRI datasets. As such, ethical approval was not required according to institutional policies and national regulations. All methods were carried out in accordance with relevant guidelines and regulations, and in compliance with the Declaration of Helsinki. The requirement for informed consent was waived. • Consent for publication - Not Applicable • Data availability - This study used publicly available, anonymized MRI datasets. • Materials availability - Not Applicable • Code availability - The source code used in this study is available from the corresponding author on reasonable request. • Author contribution - PKR designed the study, implemented the methodology, and performed experiments. MSU contributed to data preprocessing, analysis, and manuscript preparation. Both authors read and approved the final manuscript. • Acknowledgements - Not Applicable References Davis TS, Nathan JE, Tinoco Martinez AS, De Vis JB, Turtzo LC, Latour LL. Comparison of T1-Post and FLAIR-Post MRI for identification of trau- matic meningeal enhancement in traumatic brain injury patients. PLoS One . 2020;15(7):e0234881. Roozpeykar S, Azizian M, Zamani Z, Farzan MR, Veshnavei HA, Tavoosi N, Toghyani A, Sadeghian A, Afzali M. Contrast-enhanced weighted-T1 and FLAIR sequences in MRI of meningeal lesions. Am J Nucl Med Mol Imaging . 2022;12(2):63–72. Schweitzer AD, Niogi SN, Whitlow CT, Tsiouris AJ. Traumatic brain injury: imaging patterns and complications. Radiographics . 2019;39(6):1571–1595. Kim SC, Park SW, Ryoo I, Jung SC, Yun TJ, Choi S, Kim JH, Sohn CH. Contrast-enhanced FLAIR for evaluating mild traumatic brain injury. PLoS One . 2014;9(7):e102229. Ren X, Guo H, Li S, Wang S, Li J. A novel image classification method with CNN-XGBoost model. In: Digital Forensics and Watermarking: IWDW 2017, Magdeburg, Germany . Springer; 2017. p. 378–390. T. S. Davis, J. E. Nathan, A. S. T. Martinez, J. D. Vis, L. C. Turtzo, and L. Latour, Mri dataset supporting “comparison of t1-post and flair-post mri for identification of traumatic meningeal enhancement in traumatic brain injury patients”, Jul. 2020. [Online]. Available: https://nih.figshare.com/articles/dataset/MRI_dataset_supporting_Comparison_of_T1-Post_and_FLAIR-Post_MRI_for_identification_of_traumatic_meningeal_enhancement_in_traumatic_brain_injury_patients/12386102?file=23626349 Additional Declarations No competing interests reported. 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-7383303","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Short Report","associatedPublications":[],"authors":[{"id":505935185,"identity":"dbf19749-fdd1-4c09-8997-18afb2d30337","order_by":0,"name":"Proloy Kanti Roy","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8UlEQVRIiWNgGAWjYHACxgNAwgDEOsBQASSZmRsI6oFqYQYyzoC0MJKghYGxDWwtfi3m0ocPHPjYxmDMP7v/4OHKebXR/O1ALT8qtuHUYtmXlnBwZhuDmcSdwwwHz247njvjMGMDY8+Z2zi1GJzhMTjM28Zgw3AjmeFg47ZjuQ1ALcyMbURokQdrmXMsdz6xWswMwFoaanI3ENJi2cOWcHDGOQljwxvJBgcbjh3I3QjUchCfX8x5mA8++FBmYzjvRuLjjw01dbnzzh8++OBHBR6HgQhGNgkY/zCYPIBTPUwLwx84vw6f4lEwCkbBKBihAADBuF/7hH7B/QAAAABJRU5ErkJggg==","orcid":"","institution":"BRAC University","correspondingAuthor":true,"prefix":"","firstName":"Proloy","middleName":"Kanti","lastName":"Roy","suffix":""},{"id":505935186,"identity":"e17838f8-a6c8-4a7a-b111-7b62d7403901","order_by":1,"name":"Mohammad Sakib Uddin","email":"","orcid":"","institution":"BRAC University","correspondingAuthor":false,"prefix":"","firstName":"Mohammad","middleName":"Sakib","lastName":"Uddin","suffix":""}],"badges":[],"createdAt":"2025-08-15 17:53:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7383303/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7383303/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90002295,"identity":"5d30e1b7-5e5d-4a7b-a9b0-5043683e695a","added_by":"auto","created_at":"2025-08-27 09:05:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":24774,"visible":true,"origin":"","legend":"\u003cp\u003eCustom CNN Model Accuracy\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7383303/v1/33763ed414a4d5be2f5e04eb.png"},{"id":90002296,"identity":"a86746b7-bfa5-4ba6-a496-3994aad385dc","added_by":"auto","created_at":"2025-08-27 09:05:49","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":26189,"visible":true,"origin":"","legend":"\u003cp\u003eCustom CNN Model Loss\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7383303/v1/a933b6870bc8a330c60e4543.png"},{"id":90003690,"identity":"082cbb31-b198-4b3b-afee-42cdfd53497c","added_by":"auto","created_at":"2025-08-27 09:13:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":72229,"visible":true,"origin":"","legend":"\u003cp\u003eCustom CNN Model Confusion Matrix\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7383303/v1/596f2559940176da2c897e9e.png"},{"id":90002298,"identity":"5c229901-2518-453c-83c3-c4b7cb7a9deb","added_by":"auto","created_at":"2025-08-27 09:05:49","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":43386,"visible":true,"origin":"","legend":"\u003cp\u003eAccuracy comparison of implemented models\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7383303/v1/600eb004ced39c21beee59c1.png"},{"id":92184180,"identity":"6ca40dbb-897e-40fd-b81c-1599df6ec375","added_by":"auto","created_at":"2025-09-25 14:02:16","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":792408,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7383303/v1/d5e400bf-2624-47f9-ab3a-1b62730cc4a5.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Traumatic Meningeal Enhancement Detection by Deep Learning-based Biomedical Image Analysis and Handcrafted Features Extraction","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eDetecting Traumatic Meningeal Enhancement (TME) remains a significant challenge in medical imaging. Current diagnosis depends heavily on subjective radiologist inter- pretation, which is time-consuming and prone to human error and inter-operator variability. TME often presents subtly, appearing as a biomarker on FLAIR MRI after post-contrast administration, even without other cerebral abnormalities. Prior studies [\u003cb\u003e1\u003c/b\u003e] [\u003cb\u003e2\u003c/b\u003e] have highlighted differences in diagnostic performance among MRI sequences.\u003c/p\u003e\u003cp\u003eConvolutional Neural Networks (CNNs) are well-suited for complex brain imag- ing tasks, capable of automatically learning hierarchical patterns from raw MRI data, including FLAIR and contrast-enhanced images [\u003cb\u003e3\u003c/b\u003e] [\u003cb\u003e4\u003c/b\u003e]. CNN-based segmentation offers advantages over manual feature extraction and traditional machine learning by reduc- ing dependence on handcrafted features, adapting to diverse datasets, and enabling end-to-end learning. Depth in CNNs enhances abstract feature extraction but also increases risks of overfitting and computational cost.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e\u003ch2\u003e1.1 Limitations in Current Research\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eTME detection research faces several constraints:\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eScarcity of annotated datasets limits model robustness and generalizability.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eComplex lesion patterns\u0026mdash;subtle enhancements and irregular shapes\u0026mdash;complicate\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003edetection and segmentation.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eInterpretability of CNNs is low, which poses barriers to clinical adoption where transparency is vital.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eHigh computational demand for deep architectures makes the deployment challeng-\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eing.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eInconsistent image quality across datasets reduces reliability.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eTo address binary classification challenges, prior work has applied deep neural networks such as VGG16, showing improved accuracy but requiring substantial com- puting resources. This study builds on such approaches with a customized CNN tailored for TME detection.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e1.2 Research Motivation\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAdvances in AI have expanded opportunities for computer-based clinical decision support, enabling more precise diagnoses and personalized treatment planning. Our motivation is to develop an automated, accurate, and reliable method for detecting TME that supports clinicians by:\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eUnderstanding and applying image processing principles.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eUsing preprocessing techniques such as augmentation and reshaping to enhance\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003edatasets.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eDeveloping a CNN-based model to classify MRI images into normal, early (pre), and acute (post) TME categories.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eLeveraging deep learning to improve diagnostic performance.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eAssisting neurologists in faster, more precise detection through automated analysis.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e1.3 Research Objective\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThe primary aim is to implement a CNN-based detection and classification system for TME that integrates transfer learning with deep learning. Our proposed 13-layer CNN is designed to identify subtle features in MRI images and categorize them into normal, early (pre), and acute (post) stages. By automating detection, we aim to reduce subjectivity, increase diagnostic speed, and improve accuracy compared with manual interpretation.\u003c/p\u003e\u003cp\u003eTME\u0026rsquo;s subtle and multifaceted presentation often delays intervention, potentially worsening patient outcomes. Manual assessment by radiologists is not only subjective but also limited by individual expertise and workload constraints. An automated, objective, and precise approach could transform the diagnostic process, improving both timeliness and reliability.\u003c/p\u003e\u003cp\u003eOur system processes MRI scans through multiple stages: image acquisition, preprocessing, and classification using the proposed CNN model. This approach is expecvarious local hospitals that employ contrast-enhanced MRI to assess TME in patients with suspected traumatic brain injury (TBI). Ultimately, the system is intended to enhance clinical decision-making, reduce diagnostic variability, and improve outcomes for patients with traumatic brain injury.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"2 Methodology","content":"\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThis study applies deep learning and handcrafted feature extraction to detect trau- matic meningeal enhancement (TME) in biomedical images. Using a dataset from different local hospitals, which employed contrast-enhanced MRI to assess TME in suspected traumatic brain injury (TBI) patients, we developed a hybrid detection approach. Our method combines learned representations from deep learning with domain-specific features to improve accuracy and efficiency in TME detection.\u003c/p\u003e\u003cp\u003eWe implemented a 13-layer custom CNN and evaluated four pre-trained CNN models (InceptionV3, VGG16, VGG19, MobileNetV2) using transfer learning and ensemble techniques. Additionally, handcrafted features\u0026mdash;Histogram of Oriented Gra- dients (HOG), Mean Pixel Value (MPV), and Positional Orientation (PO)\u0026mdash;were extracted and used to train Support Vector Machine (SVM) and XGBoost models. Performance was assessed using accuracy, precision, recall, and F1-score. The cus- tomized CNN outperformed both pre-trained models and models using handcrafted features, underscoring the value of deep learning-based biomedical image analysis for TME detection.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Dataset and Pre-Processing\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThe dataset, originally in DICOM format, obtained from publicly available sources [\u003cb\u003e6\u003c/b\u003e], was converted to JPG using the Pydicom and Pillow libraries. A total of 7,800 images were prepared, with 6,000 used for training and 1,800 for testing.\u003c/p\u003e\u003cp\u003ePre-processing aimed to reduce inconsistencies and prepare the data for optimal model performance. Key steps included:\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eImage Conversion: DICOM files were read, pixel arrays extracted, rescaled to 0\u0026ndash;255 (uint8), and saved as JPG images.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eResizing: All images were scaled to 128\u0026times;128 pixels to match the input requirements\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eof both the transfer learning models and the 13-layer CNN.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eAugmentation: To address the limited dataset size, various augmentation techniques were applied, including rescaling, shearing, zooming, shifting (width/height), and\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eflipping (vertical/horizontal), thereby increasing training diversity and improving model generalization.\u003c/p\u003e\u003cp\u003eThese pre-processing steps ensured consistent image dimensions, expanded train- ing samples, and enhanced the dataset\u0026rsquo;s suitability for deep learning-based TME detection.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Architecture of Proposed Model\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eWe used the Keras neural network toolbox to build a sequential CNN model for the given system. The correctness of the model was examined. Thirteen layers make up the model. Additionally, there were three max-pooling layers, three 2D convolutional layers, and an equal number of batch normalization layers. Two layers of dense paint, then one layer of flattening and a dropout layer.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eInput Layer\u003c/b\u003e: This is the first convolutional layer, and it uses the Rectified Linear Unit (ReLU) activation function with 32 filters of size (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). The input shape is supplied as (img height, img width, 3), giving the input images\u0026rsquo; height, width, and colour channels\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003ePooling and Normalization Layers\u003c/b\u003e: The code repeats a similar pattern twice\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003efor additional convolutional blocks. The second convolutional block uses 32 filters, and the third one uses 64 filters. Each block is followed by max pooling and batch normalization.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eFlatten Layer\u003c/b\u003e: This layer flattens the output from the preceding convolutional\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eblocks, transforming it into a one-dimensional vector. It provides the data for the completely interconnected layers.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eFully Connected Layers\u003c/b\u003e: The fully linked layer features ReLU activation function\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eand 256 neurons. With a rate of 0.5 for dropout, half of the neurons in this layer will be arbitrarily set to zero at each update during training. This aids in avoiding overfitting. Three neurons in the last layer, which represent the three classes in the classification job, are part of a densely linked output layer. The probability distributions over the classes are obtained using the softmax activation function.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003e13-Layer Custom CNN Model\u003c/p\u003e\u003c/div\u003e\u003c/caption\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=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLayers\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eOutput Shape\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eParameter\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003econv2d (Conv2D)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(None, 126, 126, 32)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e896\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003emax pooling2d (MaxPooling2D)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(None, 63, 63, 32)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003econv2d 1 (Conv2D)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(None, 61, 61, 32)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e9248\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003emax pooling2d 1 (MaxPooling2D)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(None, 30, 30, 32)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003econv2d 2 (Conv2D)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(None, 28, 28, 64)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e18496\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003emax pooling2d 2 (MaxPooling2D)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(None, 14, 14, 64)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eflatten (Flatten)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(None, 12544)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003edense (Dense)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(None, 256)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e3211520\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003edropout (Dropout)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(None, 256)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003edense 1 (Dense)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(None, 3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e771\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=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Models Architecture\u003c/h2\u003e\u003cp\u003eWe implemented a custom 13-layer CNN alongside transfer learning using VGG16, VGG19, InceptionV3, and MobileNetV2.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eVGG16 and VGG19\u003c/b\u003e: Deep networks (16 and 19 layers) using 3\u0026times;3 convolutions, max pooling, and ReLU activations, taking 128\u0026times;128 RGB inputs.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eInceptionV3\u003c/b\u003e: Optimized with smaller convolutions, auxiliary classifiers, and\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eefficient grid-size reduction to lower parameter count and improve speed.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eMobileNetV2\u003c/b\u003e: Lightweight architecture using depthwise separable convolutions for computational efficiency.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eAn ensemble model averaged predictions from the four pre-trained networks to enhance accuracy.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Handcrafted Feature Models\u003c/h2\u003e\u003cp\u003eHistogram of Oriented Gradients (HOG), Mean Pixel Value (MPV), and Positional Orientation (PO) features were extracted and used to train:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eXGBoost\u003c/b\u003e: Gradient boosting trees with regularization and subsampling for robustness.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eSVM\u003c/b\u003e: Support Vector Classifier tuned for multi-class prediction.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eModel performance was evaluated using accuracy, precision, recall, and F1-score, with the custom CNN outperforming pre-trained models on handcrafted feature datasets.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"3 Performance Study","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 Implementation\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eModel training was regulated by parameters such as size, epochs, layers, groups, trans- former layers, filters, and matrix settings. After preprocessing, training began using convolutional layers followed by max pooling. Pre-trained models and a custom CNN were employed, classifying images into early, acute, and normal categories. Of 7,800 images, 76.9% (6,000) were used for training and 23.1% for testing.\u003c/p\u003e\n \u003cp\u003eThe Adam optimizer was selected for its efficiency, low memory needs, and adapt- ability to large datasets. Training ran for 35 epochs with a batch size of 32, shuffling training and validation data between epochs. All images were RGB. ReLU activa- tion was applied to capture complex biomedical image features, while categorical cross-entropy loss optimized multiclass classification accuracy. Class mode was set to \u0026ldquo;categorical,\u0026rdquo; predicting enhancement severity.\u003c/p\u003e\n \u003cp\u003eRegularization was implemented using dropout to reduce overfitting. The soft- max output layer produced probability distributions for classification. Training ran on Google Colab GPUs (Tesla K80, T4, P4), with hardware details accessible via \u0026ldquo;!nvidia-smi.\u003c/p\u003e\n \u003cp\u003eThe training process is examined with reference to the optimizer\u0026rsquo;s repetition as a deep learning technique. In general, the \u0026rdquo;Adam\u0026rdquo; [\u003cstrong\u003e5\u003c/strong\u003e] optimizer function is described by Eq.\u0026nbsp;(1)\u003c/p\u003e\n \u003cp\u003e\u003cem\u003e\u0026omega;\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e+1\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;\u003cem\u003e\u0026omega;\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e\u0026thinsp;\u003cem\u003e\u0026minus;\u0026thinsp;\u0026alpha;m\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e (\u003cspan class=\"CitationRef\"\u003e1\u003c/span\u003e)\u003c/p\u003e\n \u003cp\u003e\u003cem\u003em\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = aggregate of gradients at time \u003cem\u003et \u0026alpha;\u003c/em\u003e\u0026thinsp;=\u0026thinsp;learning rate at time \u003cem\u003et\u003c/em\u003e\u003c/p\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cem\u003e\u0026omega;\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;weights at time \u003cem\u003et\u003c/em\u003e\u003c/p\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026omega;\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e+1\u003c/sub\u003e\u0026thinsp;=\u0026thinsp;weights at time \u003cem\u003et\u003c/em\u003e\u0026thinsp;+\u0026thinsp;1\u003c/p\u003e\n \u003cp\u003ePositive values are retained, while negative values are converted to zero using the Rectified Linear Unit (ReLU) activation function.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003ef\u003c/em\u003e (\u003cem\u003ex\u003c/em\u003e)\u0026thinsp;=\u0026thinsp;max(0, \u003cem\u003ex\u003c/em\u003e) (\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e)\u003c/p\u003e\n \u003cp\u003e\u003cem\u003ex\u003c/em\u003e represents the input to the ReLU function.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003ef\u003c/em\u003e (\u003cem\u003ex\u003c/em\u003e) denotes the output of the ReLU function.\u003c/p\u003e\n \u003cp\u003eThe function \u003cem\u003ef\u003c/em\u003e (\u003cem\u003ex\u003c/em\u003e) returns \u003cem\u003ex\u003c/em\u003e if \u003cem\u003ex\u003c/em\u003e is positive or zero; otherwise, it returns zero.\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eParameters Used for the Pre-trained Models and the 13-Layer CNN Model\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eParameter\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePre-trained models\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e13-layered CNN Models\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTrain Data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e70%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e70%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTest Data\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e30%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e30%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTarget Size\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(128,128)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e(128,128)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBatch Size\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEpoch\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e35\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEnvironment of Execution\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGPU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGPU\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOptimizer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAdam\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAdam\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLoss Function\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCategorical CrossEntropy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCategorical CrossEntropy\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eActivation Function\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSoftmax\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSoftmax\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eClass Mode\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eColour Mode\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRGB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRGB\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cem\u003ey\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e is the true label distribution vector in a one-hot encoded form for the three classes.\u003c/p\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003e\u003cem\u003ey\u003c/em\u003e\u003csub\u003e\u003cem\u003ep\u003c/em\u003e\u003c/sub\u003e is the predicted probability distribution vector across all three classes.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003ey\u003c/em\u003e\u003csub\u003e\u003cem\u003et,i\u003c/em\u003e\u003c/sub\u003e, \u003cem\u003ey\u003c/em\u003e\u003csub\u003e\u003cem\u003ep,i\u003c/em\u003e\u003c/sub\u003e represent the true and predicted probabilities for each of the three classes.\u003c/p\u003e\n \u003cp\u003eTraining is initiated by clicking the \u0026rdquo;model. fit\u0026rdquo; in the code. Below the contents of the cell is a progress indicator showing the stages and epoch of the entire procedure. The pre-trained model with all of the parameters and matrices selected indicates that the primary validation parameter has been completed when the designated time is completed.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2 Performance Matrices\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eHere, the calculation of accuracy, precision, recall, and AUC performance of each model is stated [\u003cstrong\u003e1\u003c/strong\u003e]. The equations are described below:\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003ch3\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/h3\u003e\n\u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eAUC: A measure of a classifier\u0026rsquo;s capacity to distinguish between classes is referred to as the Area Under the Curve (AUC) [1]. It just expresses the degree of in dependence or the scale by which it is assessed. Given that AUC is scale and classification-threshold-invariant, it provides an aggregated performance statis- tic across all thresholds. How well the model distinguishes between positive and nega tive categories is shown by the AUC rate. An increased AUC rate indicates improved performance for each specific model. In the following graphs, the x-axis defines the number of epochs, and the y-axis defines accuracy rate\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n \u003ch2\u003e3.3 Performance of CNN Model\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eWe selected 7800 images among them, we used 6000 images for training and 1800 images for testing, separated into three portions for normal brain, early (pre), and acute (post). Each class\u0026rsquo;s training data accounts for 76.9% of the total data, while testing data accounts for 23.1%. Finally, our proposed model was found to be 91% correct.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\n \u003ch2\u003e3.4 Performance of Pre-Trained Models\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eA total of 7,800 images were used, divided into three classes\u0026mdash;normal, early (pre), and acute (post), with 1,800 images (600 per class) allocated for testing and the remainder for training. Four pre-trained models, MobileNet, InceptionV3, VGG16, and VGG19 were evaluated. MobileNet achieved the highest performance among them, with 95% training accuracy and 89% testing accuracy. VGG16 achieved 86% accuracy, VGG19 reached 84%, and InceptionV3 achieved 80%. Model performance was assessed through training/testing accuracy and loss curves, as well as confusion matrices, which illustrated classification effectiveness across the three categories.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\n \u003ch2\u003e3.5 Performance of Handcrafted Models\u003c/h2\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eWe evaluated handcrafted features (HOG, MPV, PO) using SVM and XGBoost classifiers.\u003c/p\u003e\n \u003c/div\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eSVM\u003c/strong\u003e: HOG 77.5%, MPV 70.4%, PO 66.3%\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003e\u003cstrong\u003eXGBoost\u003c/strong\u003e: HOG 70.3%, MPV 63.8%, PO 60.1%\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003cdiv class=\"BlockQuote\"\u003e\n \u003cp\u003eHOG with SVM achieved the highest accuracy among handcrafted feature methods.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"4 Result","content":"\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003ePre-trained models achieved the following accuracies: VGG16 86%, VGG19 84%, InceptionV3 80%, and MobileNet 89%. An ensemble of these models reached 88%, while our custom 13-layer CNN achieved 91%. Handcrafted features (HOG, MPV, PO) with SVM yielded 80%, 70%, and 66%, respectively; with XGBoost, 71%, 66%, and 62%. All alternatives performed below our lightweight custom CNN, confirming its superior accuracy on this dataset.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThis study presents a custom 13-layer CNN model for the automatic classification and detection of traumatic meningeal injuries, achieving 91% accuracy on a dataset of 7,800 MRI images categorized as normal, early (pre), and acute (post). The proposed model consistently outperformed well-known transfer learning architectures (VGG16, VGG19, MobileNet, InceptionV3) and handcrafted feature-based methods (HOG, MPV, PO with SVM and XGBoost), demonstrating its robustness and adaptability. By enabling accurate and timely detection, the model has the potential to support early diagnosis, guide treatment decisions, and improve patient outcomes. The results highlight the advantages of a lightweight yet effective deep learning architecture in medical image analysis. Future research will aim to optimize the model\u0026rsquo;s efficiency, expand dataset diversity, and explore integration into clinical workflows for real-world application.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e• Funding - Not Applicable\u003c/p\u003e\n\u003cp\u003e• Conflict of interest/Competing interests - The authors declare that they have no\u003c/p\u003e\n\u003cp\u003ecompeting interests.\u003c/p\u003e\n\u003cp\u003e• Ethics approval and consent to participate - This study used publicly available, anonymized MRI datasets. As such, ethical approval was not required according to institutional policies and national regulations. All methods were carried out in accordance with relevant guidelines and regulations, and in compliance with the Declaration of Helsinki. The requirement for informed consent was waived.\u003c/p\u003e\n\u003cp\u003e• Consent for publication - Not Applicable\u003c/p\u003e\n\u003cp\u003e• Data availability - This study used publicly available, anonymized MRI datasets.\u003c/p\u003e\n\u003cp\u003e• Materials availability - Not Applicable\u003c/p\u003e\n\u003cp\u003e• Code availability - The source code used in this study is available from the\u003c/p\u003e\n\u003cp\u003ecorresponding\u0026nbsp;author\u0026nbsp;on\u0026nbsp;reasonable\u0026nbsp;request.\u003c/p\u003e\n\u003cp\u003e• Author contribution -\u0026nbsp;PKR designed the study, implemented the methodology, and performed experiments. MSU contributed to data preprocessing, analysis, and\u003c/p\u003e\n\u003cp\u003emanuscript\u0026nbsp;preparation.\u0026nbsp;Both\u0026nbsp;authors\u0026nbsp;read\u0026nbsp;and\u0026nbsp;approved\u0026nbsp;the\u0026nbsp;final manuscript.\u003c/p\u003e\n\u003cp\u003e• Acknowledgements - Not Applicable\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eDavis TS, Nathan JE, Tinoco Martinez AS, De Vis JB, Turtzo LC, Latour LL. Comparison of T1-Post and FLAIR-Post MRI for identification of trau- matic meningeal enhancement in traumatic brain injury patients. \u003cem\u003ePLoS One\u003c/em\u003e. 2020;15(7):e0234881.\u003c/li\u003e\n \u003cli\u003eRoozpeykar S, Azizian M, Zamani Z, Farzan MR, Veshnavei HA, Tavoosi N, Toghyani A, Sadeghian A, Afzali M. Contrast-enhanced weighted-T1 and FLAIR sequences in MRI of meningeal lesions. \u003cem\u003eAm\u0026nbsp;\u003c/em\u003e\u003cem\u003eJ\u0026nbsp;\u003c/em\u003e\u003cem\u003eNucl Med Mol Imaging\u003c/em\u003e. 2022;12(2):63\u0026ndash;72.\u003c/li\u003e\n \u003cli\u003eSchweitzer AD, Niogi SN, Whitlow CT, Tsiouris AJ. Traumatic brain injury: imaging patterns and complications. \u003cem\u003eRadiographics\u003c/em\u003e. 2019;39(6):1571\u0026ndash;1595.\u003c/li\u003e\n \u003cli\u003eKim SC, Park SW, Ryoo I, Jung SC, Yun TJ, Choi S, Kim JH, Sohn CH. Contrast-enhanced FLAIR for evaluating mild traumatic brain injury. \u003cem\u003ePLoS One\u003c/em\u003e. 2014;9(7):e102229.\u003c/li\u003e\n \u003cli\u003eRen X, Guo H, Li S, Wang S, Li J. A novel image classification method with CNN-XGBoost model. In: \u003cem\u003eDigital Forensics and Watermarking: IWDW 2017, Magdeburg, Germany\u003c/em\u003e. Springer; 2017. p. 378\u0026ndash;390.\u003c/li\u003e\n \u003cli\u003eT. S. Davis, J. E. Nathan, A. S. T. Martinez, J. D. Vis, L. C. Turtzo, and L. Latour, Mri dataset supporting \u0026ldquo;comparison of t1-post and flair-post mri for identification of traumatic meningeal enhancement in traumatic brain injury patients\u0026rdquo;, Jul. 2020. [Online]. Available: https://nih.figshare.com/articles/dataset/MRI_dataset_supporting_Comparison_of_T1-Post_and_FLAIR-Post_MRI_for_identification_of_traumatic_meningeal_enhancement_in_traumatic_brain_injury_patients/12386102?file=23626349\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","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":"TME, CNN, Pre-trained, feature extraction, Image Processing, Transfer Learning","lastPublishedDoi":"10.21203/rs.3.rs-7383303/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7383303/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTraumatic Meningeal Enhancement (TME) is characterized by abnormal enhancement of the meninges following head trauma, detectable on MRI even without other cerebral abnormalities. Causes include falls, motor vehicle acci- dents, sports injuries, and assaults. This study investigates advanced imaging for TME detection using a customized 13-layer convolutional neural network (CNN) compared to transfer learning and handcrafted feature extraction methods. A data set of 7,800 MRI images was enhanced and classified into normal, early (pre), and acute (post) categories. Seventy percent of the data was used for training and 30% for testing. Four transfer learning models, VGG19, VGG16, InceptionV3, and MobileNet, achieved accuracies of 84%, 86%, 80%, and 89%, respectively; their ensemble achieved 88.83%. Handcrafted feature extraction using positional orientation (PO), histogram of oriented gradients (HOG), and mean pixel value (MPV) was tested with Support Vector Machine (SVM) and XGBoost, yield- ing lower accuracies (SVM: 65\u0026ndash;81%; XGBoost: 57\u0026ndash;72%). The customized CNN achieved the highest accuracy at 91%, outperforming all other models. These results demonstrate the potential of the proposed CNN for accurate and reliable identification of TME in medical imaging.\u003c/p\u003e","manuscriptTitle":"Traumatic Meningeal Enhancement Detection by Deep Learning-based Biomedical Image Analysis and Handcrafted Features Extraction","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-27 09:05:44","doi":"10.21203/rs.3.rs-7383303/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":"8b9c6788-9559-4f4b-be9a-f519261a8877","owner":[],"postedDate":"August 27th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-25T13:54:08+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-27 09:05:44","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7383303","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7383303","identity":"rs-7383303","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}