A Smart Platform for Management and Diagnosis of Multiple Sclerosis

Abstract

Multiple Sclerosis is a serious disease that affects the central nervous system and is not always easy to diagnose. MRI plays a big role in detecting the lesions caused by this disease, but reading these images takes time and depends a lot on the experience of the doctor. The goal of this work was to use deep learning to help neurologists in this task, and to make the diagnosis and the follow-up of patients easier. To reach this goal, we first studied the medical side of the disease and the role of MRI in its detection. Then we presented the main ideas of machine learning and deep learning, and especially convolutional neural networks, which are well suited for image analysis. After that, we designed and built our platform, MyelineCare, with its AI engine MyelineVision. This engine is a CNN that we trained from scratch to classify FLAIR images into two classes, MS and Healthy, and to show the suspicious regions using Grad-CAM. The model gave very good results on the test set. Still, we tried to stay honest about these results: the images came from different sources, which can make the task easier than it really is, so the results should be confirmed later on a larger and more controlled datasets. It is also important to say that our platform is only a tool to help the doctor, and not to replace the doctor, because the final decision always belongs to the neurologist. In the end, this project allowed us to bring together the medical field and artificial intelligence, and to build a complete and usable platform. We learned a lot during this work, on the technical side as well as on the medical side. The platform can still be improved in the future, for example by working on 3D images, by adding lesion segmentation, or by testing it in real conditions with doctors. We hope that this work can be a small step toward a better use of artificial intelligence in the medical field. 80 SUMMERY,CONTRIBUTIONS AND LIMITATION,CONCLUSION Contributions and Limitations: This work contributes a complete, interpretable and deployable decision-support platform for Multiple Sclerosis, rather than an isolated classification model. It integrates a compact convolutional neural network of about 1.2 million parameters trained entirely from scratch on FLAIR slices into a usable clinical environment built on a React frontend, a FastAPI backend and a Supabase database. Around this engine, the platform provides the functionalities of a realistic clinical workflow: role-based authentication, patient management, MRI upload and analysis, Grad-CAM visual explanations, a probability score presented alongside McDonald_criteria context, report generation, longitudinal follow-up, a voice assistant and an administration module. By adding a lightweight, explainable classifier with a full management layer and framing the tool as a decision aid rather than a diagnostic replacement, MyLineCare is distinguished from the more specialised platforms reviewed in the state of the art. Its main contribution is therefore the demonstration of the technical feasibility of such an integrated, interpretable system, together with a working prototype that can serve as a foundation for clinically validated future development. The project also has clear limitations that must be stated honestly, the most important concerning the data. There is currently no large, public, unified FLAIR dataset containing both MS and healthy brains acquired under a common protocol, so the dataset had to be assembled from several heterogeneous sources the MS images from one collection and the healthy controls from others. As a result, the two classes differ not only in the presence of disease but also in scanner, intensity, resolution and preprocessing, which means the near-perfect results obtained (accuracy 99.94%, AUC 1.00) most likely reflect the model's ability to distinguish the origin of the images rather than true MS-specific features. These metrics should therefore be read as evidence that the training pipeline functions correctly, not as a measure of real clinical accuracy. Related limitations follow from the same constraint: the data consists of 2D slices rather than 3D volumes, the labels are image-level rather than lesion-level, the train/test split is performed at the slice level (risking leakage if slices of the same patient fall in both subsets), and the system was never validated on prospective clinical cases. These limitations define the natural direction of future work: validating the model on single- source or domain-harmonised data, enforcing a strict patient-level train/test separation, testing on external and prospectively collected cases with clinicians, and extending the analysis toward volumetric, lesion-annotated data. In summary, the value of this project lies in the design and implementation of a coherent, explainable and fully integrated MS support platform, and in the transparent analysis of its current diagnostic limitations an honest assessment that is itself an essential step toward responsible medical artificial intelligence