Convolutional neural network based system for fully automatic FLAIR MRI segmentation in multiple sclerosis diagnosis

This study presents an automated system using Convolutional Neural Networks (CNNs) for segmenting FLAIR Magnetic Resonance Imaging (MRI) images to aid in the diagnosis of Multiple Sclerosis (MS). The dataset included 103 patients from Imam Khomeini Hospital, Tehran and an additional 10 patients from an external center. Key preprocessing steps included skull stripping, normalization, resizing, segmentation mask processing, entropy-based exclusion, and data augmentation. The nnU-Net architecture tailored for 2D slices was employed and trained using a fivefold cross-validation approach. In the slice-level classification approach, the model achieved 83% accuracy, 100% sensitivity, 75% positive predictive value (PPV), and 99% negative predictive value (NPV) on the internal test set. For the external test set, the accuracy was 76%, sensitivity 100%, PPV 68%, and NPV 100%. Voxel-level segmentation showed a Dice Similarity Coefficient (DSC) of 70% for the internal set and 75% for the external set. The CNN-based system with nnU-Net architecture demonstrated high accuracy and reliability in segmenting MS lesions, highlighting its potential for enhancing clinical decision-making.