Position-aware representation learning with anatomical priors for enhanced pancreas tumor segmentation

Abstract

Accurate pancreatic tumor segmentation in CT images is crucial but challenging due to the complex anatomy and varied tumor appearance. Previous methods predominantly adopt two-stage segmentation approaches to identify and localize tumors and rely heavily on CNN-extracted texture features. In this study, we propose a tumor position-aware branch to learn pancreatic anatomical priors and integrate them into a standard 3D U-Net segmentation network. The tumor position-aware branch consists of three innovative components. Firstly, the proposed method utilizes discrete information bottleneck theory to extract compact and informative segmentation features with pancreatic anatomical priors. Secondly, we propose a coordinate position encoding transformer that encodes the spatial coordinates of each patch within the CT volume. This encoding provides the model with a global positional context, allowing it to effectively model the spatial relationships between anatomical structures. Thirdly, a probability margin regularization loss is proposed to further eliminate the interference of background patches on the learning of pancreatic anatomical positions. Our model is trained and validated our model on the public Medical Segmentation Decathlon (MSD) dataset and a private clinical dataset. Experimental results demonstrate that our approach achieves competitive performance compared to state-of-the-art (SOTA) methods in both pancreas and tumor segmentation, with Dice scores of 82.11% for the pancreas and 55.56% for the tumor on the MSD dataset. The proposed framework offers an effective solution to leverage anatomical priors and enhance representation learning for improved pancreatic tumor segmentation.