Multi-Pathway 3D CNN With Conditional Random Field for Automated Segmentation of Multiple Sclerosis Lesions in MRI

• Published in: IEEE access Vol. 13; pp. 62154 - 62164
• Main Authors: Saeed, Reeda, Ansari, Shahab U., Hanif, Muhammad, Javed, Kamran, Haider, Usman, Maab, Iffat, Mian Qaisar, Saeed, Plawiak, Pawel
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Autoimmune diseases; Automation; Central nervous system; conditional random field; Conditional random fields; Convolutional neural networks; Feature extraction; Filters; Image segmentation; Lesions; Magnetic resonance imaging; MS lesion segmentation; Multiple sclerosis; Solid modeling; Three-dimensional displays; Training
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3556885

Multiple Sclerosis (MS) is a chronic and autoimmune disease that causes lesions in the central nervous system. It is diagnosed based on accurate identification and segmentation of lesions in magnetic resonance imaging (MRI). The structure and dimension of the lesions provide useful information about the course and status of the disease. Manual detection of lesions is labor intensive, highly time-consuming, and prone to error. One of the challenges in automatic MS lesion segmentation is the high variability of the lesion's size and shape. In this work, a novel hybridization of the multi-scale features extraction, multi-pathway 3D convolutional neural network (CNN), and Conditional Random Field (CRF) is employed for an automated MS lesion detection and segmentation. To capture regions of interest of various shapes and sizes, we extracted multi-scale features using multi-resolution 3D input images for accurate MS lesion segmentation. To reduce over-segmentation, we employed the CRF as a post-processing step to refine the MS lesion segmentation by minimizing false positives. The CNN model is trained with 5 subjects with a mean of 4.4 time points taken from the ISBI 2015 MS lesion segmentation challenge. The model is tested on 14 subjects with a mean of 4.4 time points in the ISBI 2015 dataset. The results showed that the devised model obtained a Total Weighted Score of 91.1%, which is higher than the human rater Score of 89.4%.