Ctfnet: toward high generalization medical image segmentation via coarse-to-fine structures for multi-center datasets

Dong Sui, Sitong Bao, Donghui Lei, Yacong Li, Maozu Guo, Xiangyu Li, Kuanquan Wang, Gongning Luo

Abstract: Over the past decade, deep learning has revolutionized medical image segmentation. However, models trained on data from one medical center often fail to generalize effectively to others due to dataset heterogeneity. To address this crucial challenge of multi-center generalization, we propose CTFNet, a novel model built on a Coarse-to-Fine structure. CTFNet introduces a Multi-scale Channel Reduction Module (MCRM) to suppress dataset-specific features and a refiner module for progressive boundary calibration, enabling robust segmentation across diverse clinical settings. Extensive experiments on eight medical image segmentation datasets demonstrate that CTFNet significantly outperforms current state-of-the-art models in both segmentation accuracy and multi-center datasets generalization performance. Remarkably, when trained on the ETIS dataset, CTFNet demonstrated exceptional generalization performance on the CVC-ClinicDB dataset, achieving Dice, IoU, and MAE scores of 78.11, 68.65, and 3.66, respectively, surpassing the best-performing baseline scores of 20.21, 23.44, and 4.94. The code has been released at: https://github.com/baositong/CTFNet.

External IDs:dblp:journals/vc/SuiBLLGLWL26