Interpretable Semantic-aware Unsupervised Segmentation of Skin Lesions

Segmenting skin lesion images presents significant challenges due to the irregular morphology of lesions, com-plex backgrounds, and blurred boundaries. While traditional manual pixel-level annotations are precise, they are also labor-intensive and costly. This necessitates the development of efficient unsupervised lesion segmentation techniques. In this study, we introduce a simple yet effective framework that employs readily available generative adversarial networks (GANs) to leverage interpretable semantic directions within the latent space for unsu-pervised lesion segmentation. By reframing the disentanglement of semantic directions as a semantic space reconstruction task, our approach eliminates the reliance on supervised information. This shift removes the need for extensive manual annotations during training, without compromising segmentation accuracy. Extensive experiments on a large dataset of dermoscopic images demonstrate that our framework significantly improves both lesion synthesis and unsupervised segmentation. Remarkably, our approach achieves a 6-point Dice score improvement over the supervised U-Net model. Additionally, the proposed framework is adaptable to various GAN backbone networks, producing segmentation outputs that exhibit enhanced realism, sharper contours, and superior boundary delineation accuracy. This study paves the way for more cost-efficient and accurate skin lesion segmentation, with potential applications in automated dermatological diagnostics and beyond.