Enhanced semi-supervised lung CT segmentation with shape-aware CycleGAN synthesis of pathological data

This paper addresses the problem of pathological lung segmentation, a significant challenge in medical image analysis, particularly pronounced in cases of peripheral opacities (severe fibrosis and consolidation) because of the textural similarity between lung tissue and surrounding areas. Fully-supervised AI models require a large annotated training dataset to overcome the high variability between healthy and pathological lungs, which is hardly achievable. To overcome this challenge, this paper emphasizes the use of CycleGAN for unpaired image-to-image translation, in order to provide a data augmentation method able to generate fake pathological images, matching an existing ground truth. Although previous studies have employed CycleGAN, they often neglect the issue of shape deformation, which is crucial for accurate medical image augmentation with constrained ground truth. To address this issue, our work introduces an innovative strategy that incorporates additional loss functions. Specifically, it proposes an L1 loss based on the lung surrounding which shape is constrained to remain unchanged at the transition from the healthy to pathological domains. The lung surrounding is derived based on ground truth lung masks available in the healthy domain. Furthermore, preprocessing steps, such as cropping based on ribs/vertebra locations, are applied to refine the input for the CycleGAN, ensuring that the network focuses on the lung region. The method is applied to train in semi-supervised manner a lung segmentation model using on-the-fly data augmentation generating synthetic pathological tissues, thus requiring a reduced amount of annotated data. The results demonstrate similar performance compared to fully-supervised methods trained on large databases.