Semi-supervised segmentation of hyperspectral pathological imagery based on shape priors and contrastive learning

Deep learning-based methods have made significant progress in the field of hyperspectral pathological image segmentation. However, these methods heavily rely on a large amount of labeled samples, which are time-consuming and require specialized expertise to obtain. This poses a challenge for cholangiocarcinoma microscopy hyperspectral pathological images segmentation tasks. To address this issue, this paper proposes a semi-supervised segmentation method of cholangiocarcinoma pathological images based on shape priors and contrastive learning (SPCL), leveraging both labeled and unlabeled data. SPCL incorporates morphological opening operation as a data perturbation strategy to handle the complex noise patterns in pathological images. Additionally, the method incorporates shape priors of cholangiocarcinoma to optimize uncertainty estimation, thereby enhancing the confidence of model predictions. Moreover, SPCL introduce separate feature vector predictions for each class of samples and utilize image-level contrastive learning to improve the extraction of informative representations by the model. The proposed method is tested on cholangiocarcinoma microscopic hyperspectral data sets. Remarkably, by training with only 10% labeled data, SPCL achieves an overall accuracy of 89.25%. With 20% labeled data, the overall accuracy further improves to 91.17%. The approach outperforms existing methods, demonstrating the best segmentation performance and providing an effective solution to alleviate the challenges posed by limited labeled data, thus enhancing bile duct tumor pathology image segmentation performance.