Super-Resolution Image Reconstruction using a GAN-based approach: application in Dermatology

Cancer is a leading cause of global death, with skin cancer being a particularly deadly subtype. Malignant melanoma, an aggressive form of skin cancer, has been increasing worldwide. Early detection through screenings is crucial for improving survival rates. Dermoscopy, a non-invasive diagnostic method using a specialized instrument called a dermatoscope, has been widely adopted by dermatologists due to its ease of use and versatility, for evaluating suspicious skin lesions and avoiding unnecessary biopsies of stable lesions. Artificial intelligence (AI) is increasingly being used in biomedical research and clinical practice for automating diagnoses and assisting physicians. The accuracy of skin disease diagnosis is often linked to the experience of the dermatologist, making AI a useful tool for supplementary opinions and screening benign lesions. With the rise of mobile devices and digital healthcare, smartphone-based cameras and dermatoscopes are improving access to care and aiding treatment and screening. Teledermatology, a subspecialty of telemedicine that uses digital images for remote dermatological consultations, is becoming a convenient way for patients to receive direct physician evaluations through photograph submissions. However, image quality can be poor due to lighting, blurriness, and focus issues, hindering accurate diagnoses. Improving image quality is crucial, as high-resolution images are critical to precise diagnoses. To address this issue, software applications that incorporate automated computer vision and image processing tools, along with patient education on proper image capture and transmission techniques, can help restore and enhance the quality of images. The present study aims to propose a Generative Adversarial Network (GAN)-based methodology for the reconstruction of high-resolution skin lesion images from low-quality counterparts. This is achieved by selecting and processing images from the open public ISIC dataset through image processing techniques to eliminate artifacts and prepare the data for input into the GAN model. Data augmentation was also applied to increase the size of the dataset. Subsequently, the final paired dataset of low-resolution and high-resolution image pairs was generated through the application of degradation and enhancement techniques, and the Real-ESRGAN architecture was employed in the training phase. The resulting model was then used to generate super-resolution output images from available test sets, which were evaluated through the use of Image Quality Assessment (IQA) metrics (e.g., SSIM, FSIM, and NIQE) in addition to visual assessment. The dissimilarity between restoring the original, unaltered image (GT) and its low-resolution (LR) counterpart obtained through the degradation pipeline was analyzed to comprehend the potential efficacy of the degradation and restoration processes developed. Finally, to further validate the restoration performance of the trained GAN model, both the ground truth images and the synthetic restored images from the ISIC test set were input into a deep network for skin lesion classification, permitting a comparison between the original-quality images and the reconstructed, super-resolution images with regards to their ability to accurately classify skin lesions. In conclusion, the proposed approach demonstrates potential in the reconstitution of high-resolution images of skin lesions from low-quality ones, serving as a favorable starting point for further research and optimization.