Improved Skin Disease Classification with Mask R-CNN and Augmented Dataset

Abstract

Skin diseases are a significant global health concern, impacting millions worldwide. Severe diseases like psoriasis and dermatitis can coexist with more benign skin issues like acne and eczema. Primary care physicians in tropical areas often treat patients with skin issues in locations where onchocerciasis and tinea imbricate are prominent, such infections might even take center stage. Usually, skin illnesses are disregarded medically and considered cosmetic, but they can have serious psychosocial effects, especially at an early age, and very few global studies have attempted to quantify the frequency of skin diseases. Nevertheless, the ability to make an accurate diagnosis at an early stage is crucial for successful treatment of complex diseases. However, Skin disease identification is a complex process. We introduce a state-of-the-art approach that uses Mask R-CNN in conjunction with an augmented dataset from the HAM10000 from the Harvard University Open Data Repository to achieve close to 80% accuracy in skin disease classification. We provide an in-depth analysis of our approach, covering data preprocessing, model architecture, training, and evaluation, along with detailed tables presenting training and testing results and associated hyperparameters.