Objective:
Assessing and monitoring chronic wounds is a time-consuming task for clinicians. Subjective measures were accounted for many systematic bias and inconsistency issues. Therefore, the application of deep learning techniques holds significant value in clinical diagnosis and treatment of chronic wounds.

Material and Methods:
There are 1608 images in our wound dataset, collected from plastic and reconstructive surgery clinic of NCKUH (National Cheng Kung University Hospital) and the diabetic foot ulcer open dataset. Preprocessed images were fed to deep learning models. Four deep learning models, including Fully Convolutional Network (FCN), DeeplabV3+, Feature Pyramid Networks (FPN), and LinkNet were employed for wound segmentation and tissue segmentation. Automated wound area calculation were done after segmentation and wound scores were computed according to composition of different wound tissues.

Results
Among four models, FPN demonstrated best accuracy of wound segmentation with highest intersection over union (IoU) and Dice coefficient, 88.13% and 92.72 respectively. The IoU of wound tissue classification among necrotic tissue, and granulation tissue were 81.39% and 63.92% respectively and Dice coefficients for two tissues were 87.22% and 71.73% respectively.

Conclusions
In our study, the proposed method can be used to automatically calculate wound area and percentage of different wound tissues followed by a calculated healing score. It only took roughly 1.42 second to finish the analysis of each wound image. This approach significantly reduces the cost and time on chronic wound assessment and has a promising future in telemedicine.

Download