CeMDISC-ViT: Cerebral Microbleed Detection using Inter-Slice Context via Vision Transformers

Cerebral microbleeds (CMBs), chronic microscopic hemorrhages in brain tissue, are prevalent in conditions such as cerebral amyloid angiopathy and traumatic brain injury. Despite advances in Magnetic Resonance Imaging (MRI), manual detection of CMBs remains labor-intensive and prone to high false positive rates, inter-rater variability, and misdiagnosis. This study proposes a novel automated method utilizing Vision Transformers (ViTs) with inter-slice contextualization to enhance CMB detection and segmentation. Unlike traditional approaches focusing on individual slices, this method leverages spatial relationships across slices to improve accuracy. The system comprises a preprocessing pipeline, an interslice region proposal module, and an inter-slice microbleed segmentation module. MRI scans are preprocessed to ensure consistent spatial dimensions and intensity normalization. The region proposal module extracts inter-slice features using custom convolutional layers and a transformer encoder with multi-head self-attention, outputting bounding boxes for potential microbleed regions. Detected regions are refined by the segmentation module, which applies inter-slice attention and generates binary masks to identify true microbleeds. Evaluation on the “Where is Valdo?” Challenge dataset demonstrates the system’s robustness. The proposed interslice ViT approach outperforms traditional single-slice methods, achieving higher recall, precision, and Dice coefficients, with reduced variability. Comparative analysis highlights the method’s potential to reduce false positives and improve diagnostic reliability. By addressing critical limitations in existing methods, this study contributes to advancing automated CMB detection, with implications for broader applications in medical imaging and neurodegenerative disease research.