U-Net Approach to MRI-Based Segmentation of Multiple Sclerosis Lesions

Globally, millions are affected by Multiple Sclerosis (MS), yet remarkable advancements in lesion segmentation are transforming diagnostic processes and significantly enhancing patient care. For effective real-time lesion segmentation, it is essential to develop neural networks that use fewer convolutional layers and require less training data, ensuring high accuracy is maintained while also reducing processing time. This study presents a modified U-Net architecture with discrete encoder and decoder blocks aimed at increasing segmentation efficiency and accuracy. The proposed model demonstrates promising results, underscoring its effective-ness based on various established quality metrics. The evaluation of model performance, including the Dice Similarity Coefficient (DSC), Precision, Specificity, and Sensitivity for lesion segmentation, was conducted using the MICCAI 2016 (MSSEG-1) dataset. Notably, the achieved DSC score of 88.56% on the test dataset and 93.43% on the training set, along with a Precision value of 83.49% on the test dataset and 86.61% on the training set, as well as a Sensitivity value of 95.81% on the test dataset and 98.38% on the training set, represents the highest performance among cutting-edge techniques, further highlighting the model’s effectiveness. This approach is entirely automated, efficient, and offers impressive accuracy. Implementing additional enhancements to the model may yield a more dependable and cost-effective alternative to standard MS lesion segmentation methods.