Multi-Scale Boundary Perception Cascaded Segmentation Network for Pancreatic Lesione

The segmentation of pancreatic lesions in CT images is crucial for clinical diagnosis and treatment. Nevertheless, accurately segmenting pancreatic lesions presents a challenge owing to the low contrast between the pancreas and surrounding tissues, the diverse anatomical structures of the pancreas and its lesions, and the inherent issue of blurred boundaries. This paper proposes a multi-scale boundary perception cascaded network (MBPC) for the segmentation of pancreatic lesions in CT images. To address the challenge of low contrast between the pancreas and surrounding tissues, we have developed an effective preprocessing scheme. The encoder features a multi-scale fusion module, facilitating the extraction of semantic information at different scales and mitigating the impact of shape transformations. Then, We introduced the Boundary-Aware Module (BAM) and the Boundary Fusion Module (BFM) to extract boundary information for the pancreas and its lesions, thereby improving the model's performance in addressing edge-related tasks within the pancreatic and lesion regions. Finally, the proposed method is experimentally validated on public datasets, demonstrating competitive performance compared to some state-of-the-art pancreatic segmentation algorithms.