A new approach for identifying skin diseases from dermatological RGB images using source separation

TL;DR

This paper introduces a novel two-step method using source separation and chromophore analysis to improve skin disease identification from RGB dermatological images, demonstrating promising results on open access data.

Contribution

It presents a new BSS approach with specialized initialization for NMF and a disease identification criterion based on chromophore contributions, advancing dermatological image analysis.

Findings

Effective separation of chromophores in RGB images.

Improved disease identification accuracy over existing methods.

Validation on open access melanoma and nevus images.

Abstract

In this article, we propose a new BSS approach for identifying skin diseases from RGB images that proceeds in two steps. We begin by separating the three main chromophores (oxyhemoglobin, deoxyhemoglobin and melanin) using Non-negative Matrix Factorization (NMF). For this purpose, we propose a special initialization of the solution matrices based on the sparsity of the chromophores, instead of initializing them with random matrices as is the case for basic versions of NMF. We then propose a new disease identification criterion that exploits the three contributions of each chromophore on the three spectral bands of our RGB dermatological image. To validate our approach, we used an open access database containing RGB images of melanoma and neavus. The results obtained showed good performance for our approach in terms of chromophore separation, compared to the most commonly used method in the literature, as well as disease identification compared to identification based on the most popular criterion.