A Feature Regularization Based Meta-Learning Framework for Generalizing Prostate Mri Segmentation

Abstract

Magnetic Resonance Imaging acquired by different operators and devices often vary greatly, causing the domain shift problem, where deep learning models trained from existing data sources perform poorly on other data sources. This paper proposes a novel feature regularization based meta learning framework to address this problem. In particular, we design a domain discriminator module to regularize the encoder to extract domain-invariant features, and an image reconstruction module to regularize the shape compactness of predictions for target domain data. We evaluate our method on three public prostate MRI datasets. Experimental results show that our approach has better segmentation performance and more powerful generalization performance.