AI-powered Wound Tissue Classification and Segmentation

Background: Accurate and consistent wound documentation has long been a challenge in the adoption of best practice wound care. With few wound specialists, minimal formal wound education for most care providers, and the abundance of wound types, reliably assessing and treating wounds can be difficult - especially at scale. 

A critical part of treating a wound is determining the presence and proportions of tissue types in the wound. Even among expert wound care physicians, our analysis indicates very poor inter-rater agreement in determining the presence and regions of various tissue types, as low as 0.014 (Krippendorff alpha value) for epithelial tissue. This variability in wound assessments can greatly compromise diagnosis and treatment decisions, thus negatively impacting patient and healing outcomes.    

To solve this challenge, we’ve developed a new deep learning, AI-based method which is capable of reliably predicting 4 tissue types (epithelial, granulation, slough and eschar) in the wound bed, along with automatically computing their relative proportions and confidence in the estimations, all within milliseconds on an off-the-shelf mobile phone. 

Methods: Our deep learning model architecture is a mobile-optimized, encoder-decoder, convolutional neural network trained using 7000 labeled wound images and were validated using 200 ground-truth images, taken in diverse imaging conditions as is the case in actual clinical settings. Our model predictions (for tissue type presence and quantification) were jointly evaluated by 5 wound care physicians, and to mitigate the effects of low inter-rater agreement, we enforced the overall tissue prediction quality rating to be made by consensus by all expert-raters.

Results: Using this approach, 91% of our model's predictions were jointly rated as being "very good" to "fair", and only 9% of the predictions were rated as being "poor". In our presentation, we will provide a detailed analysis of the results obtained by our model as well as the clinical, patient and health system benefits of rapid, autonomous tissue classification. DiscussionOur deep-learning model presents a significant advancement in objective, accurate and repeatable wound measurements which leads to more accurate prognostics of wound healing, improved treatment regimes and elevates the abilities of various caregivers to document and treat wounds.