A practical framework for unsupervised structure preservation medical image enhancement

Low-quality (LQ) images often lead to difficulties in the screening and diagnosis of medical diseases. Unsupervised generative adversarial networks (GAN)-based image enhancement methods offer promising solutions. However, there is a quality-originality trade-off in that they produce visually pleasing results but fail to reserve the originality, especially the structural inputs. Moreover, objectively evaluating structure preservation for unsupervised medical image enhancement tasks (i.e., without reference images) is essential. In this study, we propose (1) Laplacian structural similarity index measure (LaSSIM) - a non-reference objective structure preservation evaluation for unsupervised medical image enhancement methods; and (2) a novel unsupervised GAN-based method called Laplacian medical image enhancement (LaMEGAN) to balance both originality and quality from LQ images. The proposed LaSSIM does not require clean reference images and is superior to SSIM in capturing image structural changes under image degradations, such as strong blurring on various image datasets. Experiments demonstrate that our LaMEGAN effectively balances the quality and originality trade-off. Compared to CycleGAN, which achieves superior quality scores but lacks in structure preservation, LaMEGAN outperforms significantly in structure preservation, scoring 4.05 compared to 3.58 on the mean doctor opinion score (MDOS). Additionally, LaMEGAN produces visually appealing images with quality scores close to CycleGAN in all eight evaluation metrics. The implementation code will be available at https://github.com/AillisInc/USPMIE.