Enhanced pancreas segmentation via self-supervised pretraining and attention-guided u-net ensemble

Accurate pancreas segmentation is vital for early disease detection and treatment planning, but remains challenging due to the organ’s variable shape, low contrast and unclear boundaries. To address these issues, we propose a three-stage deep learning framework that performs effectively with limited annotated data. First, a YOLO-based model localises the pancreas region in abdominal CT scans. Second, the U-Net with attention and contrast enhancement modules is trained as a baseline segmentation model. Third, to improve robustness and accuracy, we pretrain the segmentation encoder using a Masked Autoencoder (MAE) on both full CT images and pancreas-localised patches. This self-supervised pretraining enables the model to learn general anatomical context and pancreas-specific features from unlabelled data. Finally, an ensemble of the baseline and MAE-pretrained models integrates complementary strengths. The method was evaluated on a private dataset of 144 contrast-enhanced abdominal CT scans, with 57 scans used for four-fold cross-validation and 87 as a hold-out test set. It was also assessed using the publicly available NIH Pancreas-CT dataset with 82 scans. Our model achieved a Dice Similarity Coefficient (DSC) of 0.873 on the private test set and 0.878 on the NIH dataset, along with high precision and recall. These results highlight the effectiveness and generalisability of the proposed framework in data-limited medical imaging scenarios.