Deep Attention V-Net Architecture for Enhanced Multiple Sclerosis Segmentation

Abstract

The central nervous system is affected by multiple sclerosis (MS) which destroys the neurocommunication. Among the diagnostic imaging systems, magnetic resonance imaging is the most preferred one to track new and enlarged MS lesions. In this paper, we propose a deep-attention V-Net architecture with modified compression and expansion sections to segment MS. The first network performs feature extraction and expansion, thus delivering enhanced feature maps for segmenting the region of interest. The second network performs feature extraction with modified V-Net architecture and performs segmentation using the soft-max function. This model is evaluated on the publicly available MICCAI 16, MSSEG-2, and Brain MRI Dataset of Multiple Sclerosis with Consensus Manual Lesion Segmentation and Patient Meta Information dataset (2022) datasets and compared with the existing models. The proposed deep-attention V-Net model is also compared with sequential models, using V-Net and U-Net in terms of precision, sensitivity, accuracy, loss, mean IOU, F1 Score, and dice score. The suggested approach delivers a dice score of 0.8900 using the MICCAI 16 dataset, 0.9000 using the MSSEG-2 dataset and 0.9638 using the combined MSSEG 2 and Brain MRI Dataset of Multiple Sclerosis with Consensus Manual Lesion Segmentation and Patient Meta Information dataset (2022) datasets. These dice score values are superior to other deep-learning networks.

Keywords

• CNN
• deep learning
• lesion
• MRI
• MICCAI
• multiple sclerosis