Deep Learning driven prediction of Multiple Sclerosis using structural MRI data

Document Type

Conference Proceeding

Publication Date

9-9-2025

Publication Title

Mult Scler J

Keywords

Neurosciences & Neurology

Abstract

Introduction: Accurate and early diagnosis of MS is crucial for effective disease management. Magnetic Resonance Imaging (MRI) is a pivotal diagnostic modality due to its sensitivity in detecting pathological lesions. However, conventional diagnostic approaches heavily depend on manual interpretation, introducing variability and diagnostic delays.Objectives/Aims: In this study, we propose a deep learning framework employing Convolutional Neural Networks (CNN) for automated prediction of MS using MRI scans. Methods: The study setting was the MS clinic of a large integrated health care system serving south-eastern Michigan. Electronic data base was searched and patients with confirmed MS diagnosis and healthy controls of 150 subjects used from a dataset. MRI images were pre-processed to enhance lesion visibility. This was followed by training CNN models to classify images into MS and healthy control categories. The CNN architecture is optimized to detect intricate patterns specific to MS lesions, enhancing prediction reliability and diagnostic consistency. Data augmentation techniques was implemented to strengthen model robustness, addressing issues related to limited training data. Model performance is evaluated using metrics including accuracy, precision, recall, and F1-score, providing comprehensive insights into predictive capabilities. Results: Of the 100 MS patients, 72 were female, 27 male, 42 were White, 57 were Black, and the mean age was 45.65 years. A custom convolutional neural network (CNN) was designed with three convolutional layers (32, 64, 128 filters), ReLU activation, and max pooling, followed by flattening and two fully connected dense layers (128 units and a final output neuron with sigmoid activation). The model was trained on 100 preprocessed MRI images using binary cross-entropy loss and the Adam optimizer, over 50 epochs with 100 iterations. Data augmentation and dropout (rate 0.3) were employed to reduce overfitting. The model achieved a validation accuracy of approximately 80%, demonstrating potential for MRI-based MS prediction on limited data. Conclusion: Our preliminary findings demonstrate that CNN models effectively identify MRI features indicative of MS, achieving high diagnostic accuracy and precision. This study highlights the potential of CNN-driven approaches to significantly enhance diagnostic accuracy and clinical efficiency. Future research will involve validation using larger datasets and exploration of CNN explainability techniques to improve clinician trust and encourage adoption in MS diagnosis.

Volume

31

Issue

3

First Page

1059

Last Page

1060

Find It @ Sladen

Share

COinS