Classification of Multiple Sclerosis (MS) and Healthy-Control blood samples using Transfer Learning

Multiple sclerosis (MS) is a severe inflammatory condition of the central nervous system
(CNS) affecting about 2.5 million people globally. It is more common in females, usually
diagnosed in their 30s and 40s, and can shorten life expectancy by 5 to 10 years. While
MS is rarely fatal; its effects on a person's life can be profound, which signifies
comprehensive management and support. Most studies regarding MS focus on how
lymphocytes and other immune cells are involved in the disease. However, little attention
has been given to red blood cells (erythrocytes), which might also be important in
developing MS. Artificial intelligence (AI) has shown significant potential in medical
imaging for analyzing blood cells, enabling accurate and efficient diagnosis of various
conditions through automated image analysis. The project aims to implement an AI
pipeline based on Deep Learning (DL) algorithms (e.g., Transfer Learning approach) to
classify MS and Healthy Blood cells.

Eindrapportage

The project "Classification of Multiple Sclerosis (MS) and Healthy-Control blood samples using Transfer Learning" investigated whether peripheral blood cell morphology contains discriminative features that could support the classification or early diagnosis of Multiple Sclerosis (MS). While MS diagnosis traditionally relies on modalities such as MRI, PET imaging, and proteomics, we explored a minimally invasive alternative by analyzing blood cell images from MS patients and healthy controls.
Due to a limited dataset, we applied image–processing–based data augmentation techniques to enhance variability and improve model robustness. We then trained deep learning classifiers using both convolutional neural networks (CNNs) and transformer-based architectures. To ensure reliable evaluation despite the small cohort size, we adopted a Leave-One-Patient-Out (LOPO) cross-validation strategy, which provides strict patient-level separation between training and test data.
Among the tested models, the Swin Transformer achieved the most consistent and highest performance, with F1-scores predominantly in the 80–90% range across validation folds. These results suggest that blood cell images may encode subtle morphological patterns associated with MS.
An important and unexpected finding emerged during LOPO evaluation: one specific fold produced a notably lower F1-score. This observation raises the possibility of sex-specific morphological variations influencing model generalization. It also highlights the limitations of small, demographically imbalanced datasets in biomedical AI research.
Overall, our findings demonstrate the feasibility of transformer-based approaches for MS classification from blood cell imagery while emphasizing the critical need for larger, sex-balanced cohorts to validate and generalize these results at scale.