Component Dimensionality Reduction in Dermoscopic Images using a Machine Learning Technique with Validation By Similarity Metrics

Skin cancer is one of the fastest growing diseases globally, where treatments are generally subject to systems with high implementation and development costs. This paper presents the dimensionality reduction of components in dermoscopic images using the Principal Components Analysis technique based on Machine Learning, and complementing the processing by feature enhancement, as an initial stage of a computer-aided diagnosis system. A digital microscope and a Raspberry Pi 3B+ embedded board with Python programming language and open source-based libraries are used. Images reduced to 10, 25, 50, 75 and 100 components are obtained, validated by similarity metrics, as well as by color preservation and execution response time. The best results were obtained with images between 50 and 100 components where the response time does not exceed 3 seconds with a similarity of more than 90 % in images with defined sections, and more than 70 % when the images do not present defined sections. The system is replicable on a large scale due to its high rate of feature preservation and low implementation cost.