Using Machine Learning Techniques to Improve the Automated Diagnosis of Pigmented Skin Lesions

Abstract

The scientific community’s interest in predicting skin cancers through computer-aided
diagnosis has grown in recent years, especially regarding melanomas. This phenomenon is
mainly due to the continuous and progressive improvement of computer vision technologies
and algorithms that can be used more effectively to recognize skin features and pigmented
lesions that may indicate potential cancer risk. Despite this, the pre-triage instrumentation and
services on the market today appear to be excessively expensive, both because of the cost of
the equipment itself and because of the necessary employment of teams of specialist
dermatologists to analyze the images of skin lesions and pigmented lesions. On the other hand,
applications based on deep learning techniques follow a black-box approach, thus making it
impossible to identify which specific lesion or pigmented lesion feature might suggest a
potential cancer risk. This paper presents the development of a melanoma classifier based on
verifying diagnostic criteria and a specific medically approved seven-point checklist using a
logistic regression classifier. The results, also obtained from comparing clinical and
dermoscopic images, show great potential and room for improvement.