Progressively evolutionary deep learning network for 3D medical image segmentation

Motivated by the goals of improving segmentation of challenging organ cases containing local irregular shapes and highly varied shapes between subjects, a novel dual-branch framework is presented for 3D medical image segmentation. The image branch utilizes nested skip connections and a cross-slice feature modeling module to capture rich semantic context. The point cloud branch iteratively refines the surface mesh through an evolution module and adaptive topology optimization, ensuring smooth and consistent shapes. Anatomical plausibility is further enforced through a composite loss function, combining spatial consistency loss, surface distance loss, normal consistency loss, and Laplacian loss. The effectiveness of our method has been validated on two abdominal segmentation challenge datasets (BTCV and FLARE2021). Additionally, we evaluated the generalization ability of our method on the MM-WHS challenge dataset. Our method ranks fourth in overall score on the FLARE2021 dataset and achieves one of the most competitive results in abdominal segmentation. The code is available at: https://github.com/secret353/code.