Resolution preservation over architectural complexity: high-resolution patch-based semi-supervised learning for pancreas segmentation from CT images

Precise segmentation of the pancreas from computed tomography (CT) scans is critical for diagnosis and surgical planning but remains challenging due to the organ’s small volume (<0.5% of the abdomen), high anatomical variability, and low contrast with surround­ing tissues. Conventional deep learning methods typically downsample inputs to fit GPU memory, destroying fine-grained boundary information essential for small-organ delinea­tion. This study proposes a high-resolution patch-based framework that preserves native CT resolution (512 × 512) through 256 × 256 patch extraction with domain-specific Hounsfield unit (HU) windowing. A lightweight 2D U-Net (7.8M parameters) achieves 0.809±0.015 Dice (peak 0.849) on the Medical Segmentation Decathlon (MSD) Task07 Pancreas dataset, com­petitive with nnU-Net (0.823 Dice, 88M parameters) and far above SAM (0.705) and Med­SAM (0.439). Five-fold cross-validation across all 281 volumes yields a population-level Dice of 0.630±0.025. Integrating uncertainty-aware mean teacher (UA-MT) semi-supervised learning recovers 99.3% of supervised performance using only 50% of labelled data, with reproducible results across three seeds (σ = 0.012). On 80 external TCIA volumes, UA-MT (0.603 Dice) out­performs the supervised model (0.544), demonstrating superior cross-domain generalisation.