Semi-Supervised Pseudo-Healthy Image Synthesis Via Confidence Augmentation

Pseudo-healthy image synthesis, which computationally synthesizes a pathology-free image from a pathological one, has been proved valuable in many downstream medical image analysis tasks, from lesion detection, data augmentation to clinical surgery suggestion. Thanks to the advancement of generative adversarial networks (GANs), recent studies have made steady progress to synthesize realistic-looking pseudo-healthy images with the perseverance of the structure identity as well as the healthy-looking appearance. Nevertheless, it is challenging to generate high-quality pseudo-healthy images in the absence of the lesion segmentation mask. In this paper, we aim to alleviate the needs of a large amount of lesion segmentation labeled data when synthesizing pseudo-healthy images. We propose a semi-supervised pseudo-healthy image synthesis framework which leverages unlabeled pathological image data for efficient pseudo-healthy image synthesis based on a novel confidence augmentation trick. Furthermore, we re-design the network architecture which takes advantage of previous studies and allows for more flexible applications. Both quantitative and qualitative experiments show that our model outperforms the best state-of-the-art model by up to 3\% in generating high-quality images and 6\% for data augmentation in a supervised learning setting. Moreover, the proposed semi-supervised learning achieves comparable pseudo-healthy image synthesis quality with supervised learning models, using only 50\% of lesion segmentation labeled data.