Transformer-based Self-supervised Learning for Automated Detection of Rare Pathologies in High-resolution 3D Medical Imaging

Computational diagnosis of rare pathologies in high-resolution 3D medical images is a challenging task, as it suffers from scarce labelled data in the presence of subtle disease presentation and demand for effectual volumetric representations. Towards this goal, this work introduces a transformer-based self-supervised learning framework that uses abundant unlabelled MRI and CT scans to learn universal anatomical representations without requiring manual label annotation. The model features masked volume modelling, in which randomly occluded 3D patches are modelled to deal with long-range spatial dependencies. Following pre-training, the model is fine-tuned with a few labelled samples from rare brain and lung pathologies. Experimental results on the BraTS 2021 and LIDC-IDRI datasets show that its performance surpasses supervised U-Net and ResNet-3D baselines with higher Dice and AUC-ROC scores. Attention maps offer interpretability by highlighting the clinically relevant areas that affect model predictions. The findings suggest self-supervised transformer architectures as a scalable and data-efficient approach to rare pathology detection in 3D medical imaging.