Sample Alignment for Image-to-Image Translation Based Medical Domain Adaptation

Abstract

Image-to-image (I2I) translation is a popular paradigm in domain adaptation (DA), and has been frequently used to address the lack of labeled data. However, as a result of the sample bias in medical data caused by the attributes of imaging modality or pathology, the I2I translation based DA always suffers from synthesis artifacts. For boosting the DA in medical scenarios, a sample alignment algorithm is proposed to correct the sample bias in medical data. Specifically, diffeomorphic transformation and symmetric resampling are employed to implement the sample alignment. The topological structure in medical samples is first aligned using diffeomorphic transformation. Then paired image data are collected from the aligned samples by symmetric resampling to train the I2I translation models. In the experiment, the proposed algorithm was applied to boost the DA of cross-modality data and pathological ones. Our algorithm not only improved the quality of synthesized images, but also promoted the DA of diagnosis models learned from synthesized data.