Automated quantification of 3D wound morphology by machine learning and optical coherence tomography in type 2 diabetes.

Abstract

Driven by increased prevalence of type 2 diabetes and ageing populations, wounds affect millions of people each year, but monitoring and treatment remain limited. Glucocorticoid (stress hormones) activation by the enzyme 11β-hydroxysteroid dehydrogenase type 1 (11β-HSD1) also impairs healing. We recently reported that 11β-HSD1 inhibition with oral AZD4017 improves acute wound healing by manual 2D optical coherence tomography (OCT), although this method is subjective and labour-intensive. Here, we aimed to develop an automated method of 3D OCT for rapid identification and quantification of multiple wound morphologies. We analysed 204 3D OCT scans of 3mm punch biopsies representing 24 480 2D wound image frames. A u-net method was used for image segmentation into 4 key wound morphologies: early granulation tissue, late granulation tissue, neo-epidermis, and blood clot. U-net training was conducted with 0.2% of available frames, with a mini-batch accuracy of 86%. The trained model was applied to compare segment area (per frame) and volume (per scan) at days 2 and 7 post-wounding and in AZD4017 compared to placebo. Automated OCT distinguished wound tissue morphologies, quantifying their volumetric transition during healing, and correlating with corresponding manual measurements. Further, AZD4017 improved epidermal re-epithelialisation (by manual OCT) with a corresponding trend towards increased neo-epidermis volume (by automated OCT). Machine learning and OCT can quantify wound healing for automated, non-invasive monitoring in real-time. This sensitive and reproducible new approach offers a step-change in wound healing research, paving the way for further development in chronic wounds.