A machine learning approach for multiple sclerosis diagnosis through Detecron Architecture

Multiple sclerosis is a prevalent inflammatory disease affecting the central nervous system, leading to demyelination. Neuroradiology relies on accurate analysis of white matter lesions for diagnosis and prognosis. Automated methods for segmenting lesions in MRI scans offer crucial benefits of objectivity and efficiency, making them particularly valuable for analyzing large datasets. In contrast, manual delineation of MRI lesions is both time-consuming and prone to subjective bias. To overcome these issues, this paper proposes and develops an automated diagnosis approach using the Detecron-2 architecture. The method utilizes a fully modified Convolutional Neural Network on 3D FLAIR-weighted Magnetic Resonance Images.The approach is trained and validated on a dataset of 3000 images acquired from a Siemens 3Tesla MRI machine at the National Institute of Neurology Mongi Ben Hmida in Tunisia, using technical metrics. Comparisons with recent achievements demonstrate promising results. By addressing challenges in data augmentation and deep learning configurations, the proposed model effectively mitigates issues as overfitting. Notably, it achieves an impressive average detection accuracy of 87%, specificity (= 80,19%), precision (= 80%), sensitivity (= 76,1%) and intersection over Union (= 87,9%) when assessing healthy and pathological images. Additionally, the study recognizes the manual monitoring of multiple sclerosis plaques as a time-consuming and challenging task for clinicians. It highlights the importance of lesion segmentation for quantitative analysis of disease progression. As a second focus, the research aims to develop an automated segmentation to enhance the accuracy and efficiency of lesion identification, addressing the inconsistencies and variations observed among different observers.