Multiple sclerosis diagnosis with deep learning and explainable AI

Diagnosing multiple sclerosis (MS) presents significant challenges due to its complex clinical presentation and the subjective interpretation of imaging findings. Machine learning (ML) and deep learning (DL) models, despite their potential, often exacerbate these challenges with their opaque decision-making processes, hindering clinical integration. This study addresses these limitations by employing eXplainable Artificial Intelligence (XAI) techniques, specifically integrating Grad-CAM within a Convolutional Neural Network (CNN) framework, EfficientNetB1, for the diagnosis of MS. The primary objective is to enhance the transparency and reliability of MS diagnosis by providing clear visual insights into the model’s decision-making process, while also identifying and mitigating potential biases and irrelevant features. Using a dataset comprising FLAIR axial and sagittal MRI images of MS patients and healthy individuals, the CNN model is trained and integrated with Grad-CAM. Post-integration observations revealed potential biases and irrelevant features, particularly in the erroneous highlighting of certain regions by the model. Subsequent adjustments and re-training using 10-fold cross-validation led to an improved model with accuracy rates of 99.82% for axial, 99.76% for sagittal images, and 99.36% overall. Furthermore, testing on a separate dataset confirmed the model’s ability to generalize and perform well across various clinical contexts. In conclusion, this study underscores the critical role of transparent and interpretable models in medical diagnostics, demonstrating that the integration of XAI techniques can significantly enhance the reliability and clinical applicability of models.