Wound image segmentation using deep convolutional neural network

Abstract

Traditional methods of wound diagnosis have been diagnosed and prescribed by the naked eye of an expert. If the wound segmentation algorithm is applied to the wound diagnosis, the area of wound can be quantitated and used as an auxiliary means of treatment. Even with dramatic development of Deep learning technology in recent years, However, a lack of datasets generally occurs overfitting problem of deep learning model, which leads to poor performance for external datasets. Therefore, we trained the wound segmentation model by adding a new wound dataset in addition to the existing Open dataset, the Diabetic Foot Ulcer Challenge Dataset. Machine learning based methods are used when producing new dataset, ground truth images. Thus, in addition to the manual methods, Gradient Vector Flow machine learning techniques is used for ground-truth image production to reduce the time consumed in vain. The wound segmentation model used in this study is a U-net with residual block combined with cross entropy loss and Dice loss. As a result of the experiment, the wound segmentation accuracy was about 90% for Dice coefficient