Semi-Supervised Domain Generalization for Medical Image Analysis

Despite of their recent progress in medical image analysis, the deep learning methods typically fail to generalize well in the case of domain shift (e.g. introduced by different imaging devices). To relieve the above problem, this paper introduces a general regularization-based semi-supervised domain generalization method. Specifically, the stability and orthogonality of the learned features are introduced as two regularization factors of the learning objective. The former encourages the model to learn to represent as many domain-invariant features as possible, while the latter constrains the features to be orthogonal to each other so that a more complete set of features is represented. Implemented as a feature correlation matrix, both regularization factors can be applied to both labelled and unlabelled data. Experimental results on two benchmark datasets, Fundus Image Segmentation and Chest X-ray Diagnosis, have shown the promise of the proposed method.