HDNeXt: Hybrid Dynamic MedNeXt with Level Set Regularization for Medical Image Segmentation

Haoyu Cao, Tianyi Han, Yunyun Yang; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 597-613

Deep learning has been extensively employed in the field of medical image segmentation, demonstrating its robustness and efficacy. However, the pursuit of consistent segmentation performance across diverse instrumental conditions and the challenge of achieving precise boundary delineation in segmented images remain significant hurdles. In this paper, we aim to develop a model capable of achieving consistent, high-quality segmentation of identical regions of interest across varying instrumental conditions, with precise boundary delineation. Toward this end, we introduce our Hybrid Dynamic MedNeXt (HDNeXt) model, an advanced framework capable of dynamically generating weights across diverse medical images to maintain consistently high segmentation performance. HDNeXt builds on the robust segmentation framework of MedNeXt by incorporating dynamic convolution techniques, which endow the model with the capability for dynamic weight adjustment, significantly enhancing its segmentation performance. To tackle the second challenge, we devised a novel loss function, L_ CR , formulated on the Curvature of the segmentation boundary and Region-Fitting energy derived from level set methods, which significantly enhances boundary precision during training and optimizes overall segmentation performance. Experiments were conducted on the abdominal CT datasets Synapse and the cardiac MRI datasets ACDC to demonstrate the efficiency and effectiveness of our method. Our method achieved an average Dice coefficient of 84.38 on the Synapse datasets and 93.59 on the ACDC datasets, surpassing other 2D state-of-the-art segmentation models and achieving optimal performance for 2D medical image segmentation. Codes are available at https://github.com/HaoyuCao/HDNeXt