ABSTRACT
More than 11 million cases of burn injuries, resulting to over 450,000 deaths, are reported annually. This was associated with lack of experienced medical personnel and diagnostic errors, such as underestimation/overestimation of burn wounds that leads to inappropriate intervention. The use of machine learning (ML) has recently shown that more useful information can be extracted from clinical images and processed accurately, and above all, more reliably than humans. Towards this end, this work entails the use of an ML algorithm to automatically discriminate healthy skin, superficial burn, and full-thickness burn. Specifically, we leverage the filters of a fully trained neural network for image feature extraction, after which the features are used to train a multiclass support vector machine (SVM). Using 1,980 images, evaluation of the algorithm is performed via a tenfold cross-validation technique. Eventually, classification accuracy of 99.9% was achieved. To the best of our knowledge, this is the first time that 2D images were used to extensively categorize superficial burn and full-thickness burn images. We believe that the proposed pipeline can be deployed so that a remote patient only has to upload an image taken via his/her mobile phone to know the severity of the situation.
