Blood Cell Imaging Analysis for Multiple Sclerosis Classification Using Leave-One-Subject-Out Study with Deep Neural Networks

Abstract

Multiple sclerosis (MS) is a chronic neurological disease whose diagnosis typically relies on resource-intensive and partly invasive procedures such as MRI and spinal fluid analysis. This study explores whether deep learning can identify distinguishing patterns in RBC images associated with MS, aiming to determine whether blood-cell imaging may provide information relevant to understanding the disease. The initial dataset comprised 121 images from 13 independent subjects. To expand the dataset to 1944 images, image augmentation techniques were applied to the training data to increase image variability. Due to the small dataset, training was carried out using a Leave-One-Subject-Out (LOSO) strategy to assess model performance on unseen subjects. Four deep learning architectures, VGG16, DenseNet201, MobileNetV2, and the Swin Transformer, were evaluated using subject-level Leave-One-Out Cross-Validation (LOOCV). The Swin Transformer achieved the highest mean accuracy (73.53%), followed by the fully fine-tuned VGG16 (71.02%). Explainable AI techniques (Grad-CAM and SWTformer-v1) applied to the given dataset showed that VGG16 focused on broader structural patterns in cell groupings, while SWTformer-v1 attended to more localized regions. These findings suggest that deep-learning models can identify distinguishing image patterns that contribute to classifying RBC images from MS and healthy controls. The XAI analyses further provide insight into the image regions the models use, although their biological relevance requires validation in larger cohorts.