DNA-Net: Genetic-Inspired Dual-Chain Learning for Medical Image Domain Generalization without Negative Transfer

Yucheng Song, Jincan Wang, HaoKang Ding, Zhiqiang Tian, Zhining Liao, Kangxu Fan, Zhifang Liao

Keywords: DNA-Net, Genetic-Inspired, Medical Image Domain Generalization, Negative Transfer Mitigation, Structure-Constrained Diffusion, Cross-Domain Representation Learning

Abstract: Domain Generalization (DG) in medical image segmentation remains a challenging yet essential problem, particularly when aiming to avoid negative transfer between source and target domains. The prevalent domain shift in clinical datasets often limits deep learning models’ generalization beyond the source domain, and many existing style augmentation methods—typically based on nonlinear transformations—fail to accurately capture the target domain distribution. Moreover, most DG approaches neglect the issue of negative knowledge transfer, leading to degraded performance in the source domain. To address these challenges, we propose a structure-constrained, diffusion-based style divergence augmentation that operates in the frequency domain using a continuous style combination mechanism. This generates diverse samples with broad domain coverage, improving representation robustness. Furthermore, inspired by biological genetics, we introduce DNA-Net, a genetic-inspired dual-chain collaborative learning framework. By jointly optimizing two related tasks—source domain image reconstruction and generalized segmentation—DNA-Net explicitly suppresses negative transfer while enhancing cross-domain segmentation performance. Extensive experiments on two public medical image benchmarks demonstrate that our approach surpasses state-of-the-art DG methods, achieving superior performance on both source and target domains. Our code is available at https://anonymous.4open.science/r/DNA-Net-SESD-5891/.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 7174