A novel CNN framework for skin disease classification using adaptive percentage filter for image binarization and fast-marching inpainting method.

Artificial intelligence assists medical practitioners with disease diagnosis from acute to chronic illnesses and has brought significant breakthrough in image-oriented analysis of diseases. Skin cancer is one of the deadliest diseases, posing a great threat to human health. Skin cancer warrants early and precise examination of skin lesions to prevent mortality. Skin lesions can be examined by biopsy, an invasive procedure to remove piece of tissue, tested in a laboratory. Biopsy is a painful procedure. The predictive models in artificial intelligence are found to be useful in analyzing skin lesions-based images. Hence, in this paper, we have analyzed 10,015 skin lesion-based images and classified seven types of skin diseases among which, one of them is melanoma, a form indicative of skin cancer. This work proposes, a novel APFB (Adaptive Percentage Filger for Binarization) to remove the hairline noise from the skin images. The performance of the APFB filter is compared with the state-of-art hairline noise removal methods such as Contrast Limited Adaptive Histogram Equalization (CLAHE) and Bilateral methods and proven to be best in terms of image quality analysis metrics such as PSNR, SSIM, BRISQUE and NIQE. The original images and hairline noise removed images are then passed into custom built CNN models for skin disease classification. The model examines the lesions features like pigment, texture, color, lines, shapes and symmetry to classify the different types of skin diseases. Among the four CNN models we created, the CNN model that takes images of APFB hair line noise removal method obtains 99.98% training accuracy and 100% testing accuracy than the other models. Thus, the model may also be used to predict skin cancer without a biopsy. [ABSTRACT FROM AUTHOR]