PMDC: Pancreas Segmentation with Mamba Block and Deformable 3D Convolution

Pancreas segmentation is a task traditionally hindered by the morphological complexity and variability of the pancreas. Transformers have excelled in capturing context information in medical images but are limited by computational resources. Recently, state space models (SSMs), especially Mamba, have demonstrated the ability to linearly scale with input sequences, surpassing transformers in long-sequence data analysis tasks and showing potential in visual tasks as well. Inspired by this, we propose a novel approach for pancreas segmentation using Mamba Block and deformable 3D convolution (PMDC). PMDC integrates Convolutional Neural Networks (CNN) and SSM, capturing both local fine-grained features and long-range dependencies through a Mamba-based U-Net architecture and leveraging deformable 3D convolutions (3DCN) to dynamically adapt to pancreatic shape variations. PMDC also ensures scalability and generalizability by self-configuring across different datasets. Experiments on NIH, MSD, and PancCT datasets demonstrate that PMDC outperforms existing CNNs and transformers, offering a robust solution for pancreas segmentation.