Subject Skin Tone Classification with Implications in Wound Imaging using Deep Learning

Chronic wound healing is inconsistent on an individual basis, leading to large treatment costs. The effectiveness of any treatment approach is typically assessed by visual inspection of the wound. Optical imaging technologies have recently been developed to objectively assess wound physiology to complement the subjective visual assessment. One such device is a low-cost SmartPhone Oxygenation Tool (SPOT), which can measure the tissue oxygenation of the wounds via non-contact imaging and assessing healing status. The varying skin tones impact tissue oxygenation measurements due to the different melanin concentrations in the epidermis of the skin. Hence, it is essential to consider melanin-related attenuation in the epidermis and account for it during tissue oxygenation measurements. This study aims to implement a machine learning algorithm to classify the skin tones using in-vivo measurements from control subjects towards a future correction for these skin tones during imaging studies using SPOT. In this study, we developed a benchmark dataset of 75,348 samples of 28 × 28 RGB images of human subjects' hands. The images were then converted to 28 × 28 grayscale images and were flattened to attain 784 pixels or features for each sample. We also developed a deep learning-based pipeline to classify the FST skin types, producing high accuracies (> 98%). The deep learning model can be incorporated into the SPOT device as an additional feature to verify or correct the melanin concentration during near-infrared (NIR) imaging of wound regions.