PiMPiC: An Overlap-Aware Contrastive Learning Framework for 3D Patch-Based Medical Image Segmentation

Deep learning models for 3D medical image segmentation typically require large annotated datasets. Contrastive learning enables representation learning without labels, but standard approaches struggle with structural similarity in medical images and 3D patch generalization. We propose two novel modules — Patch Intersection Measurement (PiM) and Patch Intersection Contrast (PiC) — improving contrastive learning by measuring overlap and aligning overlapping patch embeddings, leading to improved segmentation performance on pancreas and kidney cancer datasets. :contentReference[oaicite:1]{index=1}

Data Engineering in Medical Imaging