Robust color medical image segmentation on unseen domain by randomized illumination enhancement

Owing to the data distribution shifts generated by collecting images using various imaging protocols and device vendors, the generalization capability of deep models is crucial for medical image analysis when applied to test datasets in clinical environments. Domain generalization (DG) methods have shown promising generalization performance in the field of medical image segmentation. In contrast to conventional DG, which has strict requirements regarding the availability of multiple source domains, we consider a more challenging problem, that is, single-domain generalization (SDG), where only a single source is available during network training. In this scenario, the augmentation of the entire image to improve the model generalization ability may cause alteration of hue values, resulting in the wrong segmentation of tissues in color medical images. To resolve this problem, we first present a novel illumination-randomized SDG framework to improve the model generalization power for color medical image segmentation by synthesizing randomized illumination maps. Specifically, we devise unsupervised retinex-based image decomposition neural networks (ID-Nets) to decompose color medical images into reflectance and illumination maps. Illumination maps are augmented by performing illumination randomization to generate medical color images under diverse illumination conditions. Second, to measure the quality of retinex-based image decomposition, we devise a novel metric, the transport gradient consistency index, by modeling physical illumination. Extensive experiments are performed to evaluate our proposed framework on two retinal fundus image segmentation tasks: optic cup and disc segmentation. The experimental results demonstrate that our framework outperforms other SDG and image enhancement methods, surpassing the state-of-the-art SDG methods by up to 9.6% with respect to the Dice coefficient.