Co-Registered Volumetric MRI-Based Synthetic Lesion Transplantation for Multiple Sclerosis Classification Using Cutoff Augmentation

Multiple sclerosis (MS) is a chronic, immune-mediated disorder characterized by demyelinating lesions visible on MRI. Limited volumetric datasets and class imbalance hinder automated deep-learning approaches for MS classification. In this study, we extend the cutoff augmentation technique to three-dimensional (3D) MRI volumes and evaluate it using two publicly available pathological MS datasets and a large healthy cohort. Specifically, we used the 3D-MR-MS dataset (30 MS patients; modalities: FLAIR, T1-w, contrast-enhanced T1-w, and T2-w with multi-rater lesion segmentations), the long-MR-MS dataset (20 patients imaged longitudinally, two sessions per patient, with lesion-change masks), and the IXI healthy cohort (≈600 subjects; we selected T1-w volumes and used 50 T1-w scans for fold-specific pairing and background templates). After co-registering anatomical images, lesion masks, and brain masks into a shared healthy reference space, and performing skull stripping and intensity normalization, lesion voxels were transplanted into healthy volumes to generate synthetic pathological examples. A 3D DenseNet-169 trained with five-fold cross-validation demonstrated that 3D cutoff augmentation increased the mean accuracy from 61.2% to 72.7%, doubled the MS recall from 22.3% to 45.4%, and improved the F1 score from 36.3% to 61.7% while preserving precision. These results indicate that co-registered 3D cutoff augmentation effectively mitigates data scarcity and class imbalance for volumetric MS classification.