A framework of wound segmentation based on deep convolutional networks

作者：Xiaohui Liu, Changjian Wang, Fangzhao Li, Xiang Zhao, En Zhu, Yuxing Peng · 年份：2017 · DOI：10.1109/cisp-bmei.2017.8302184 · 被引用次数：38 · 研究领域：Pressure Ulcer Prevention and Management、Diabetic Foot Ulcer Assessment and Management、Wound Healing and Treatments

Quantification of the chronic wound size is important in clinical wound treatment as the care for each individual patient is largely based on this assessment. Recently, clinicians usually measure wound area using standardized scale mainly based on visual inspection, which is often subjective and potentially inaccurate. Unfortunatly, automated segmentation of wound area is a challenge task owing to the diverse wound characteristics and the haziness over where wound boundaries lie. To address this problem, we propose a deep convolutional networks based framework named WoundSeg to locate and segment wound areas automatically. Firstly, we establish a newly annotated dataset containing 950 digital images of different chronic wounds, and propose an annotation tool based on watershed algorithm to mask wound regions conveniently for expert clinicians. Then, a light-weight network is built specifically for wound segmentation with limited training data. Finally, morphology operation and skin detection based verification are used to postprocess the predictions of the network. We estimate the generalization error of WoundSeg with 5-fold cross-validation on the above dataset and also on various latest clinic images taken of patients at the wound clinic. Extensive experiments demonstrate the efficiency and effectiveness of the proposed framework which can be considered as a promising approach to replace empirical and imprecise manual measurement for wound areas.