MAP: Domain Generalization via Meta-Learning on Anatomy-Consistent Pseudo-Modalities

Abstract: Deep models suffer from limited generalization capability to unseen domains, which has severely hindered their clinical applicability. Specifically for the retinal vessel segmentation task, although the model is supposed to learn the anatomy of the target, it can be distracted by confounding factors like intensity and contrast. We propose Meta learning on Anatomy-consistent Pseudo-modalities (MAP), a method that improves model generalizability by learning structural features. We first leverage a feature extraction network to generate three distinct pseudo-modalities that share the vessel structure of the original image. Next, we use the episodic learning paradigm by selecting one of the pseudo-modalities as the meta-train dataset, and perform meta-testing on a continuous augmented image space generated through Dirichlet mixup of the remaining pseudo-modalities. Further, we introduce two loss functions that facilitate the model's focus on shape information by clustering the latent vectors obtained from images featuring identical vasculature. We evaluate our model on seven public datasets of various retinal imaging modalities and we conclude that MAP has substantially better generalizability. Our code is publically available at https://github.com/DeweiHu/MAP.