Design and Preliminary Evaluation of a Knowledge-Based Artificial Intelligence Framework for Skin Lesion Triage

Abstract

The timely clinical intervention in case of proper detection and evaluation of skin lesions is crucial in case of lacking access to specialized dermatology services. While Artificial Intelligence (AI) and Machine Learning (ML) have shown impressive results in automatic skin lesions assessment, many current methods are mostly developed and evaluated using images obtained with dermoscopy under optimal conditions. Their accuracy might be affected when the same methods are applied to the clinical images obtained by means of conventional or smartphone cameras due to illumination issues, lower quality, background, out-of-focus and other skin artifacts such as hairs. In this paper, we propose a new method – DermAlert – that represents an artificial intelligence-based knowledge-assisted system for automatic skin lesion analysis and clinical decision making. Our method consists of several steps including pre-processing, segmentation, hand-crafted features extraction, machine learning-based classification, and knowledge-based triage of the cases. Pre-processing involves a chain of methods, such as the application of DullRazor approach for digital removal of the hairs from the image and Contrast Limited Adaptive Histogram Equalization (CLAHE) for local contrast enhancement of the smartphone images. Support Vector Machine (SVM) is used as the primary classifier owing to its applicability to feature-based classifications and comparatively low computing power needs. The experimentation involves exploiting the complementary nature of the HAM10000 dermoscopy dataset and the PAD-UFES-20 smartphone-derived clinical images dataset. The subsequent process involves the generation of structured triage recommendations on the basis of the risk factors and the output of the machine learning models by means of a knowledge-based module. Based on a sample of 100 images from the HAM10000 dataset, an initial testing results in an accuracy of 91.2%; however, this figure should not be perceived as clinical validation and is preliminary.