Encoder-decoder semantic segmentation models for pressure wound images

Abstract

Segmentation of wound images is important for efficient wound treatment so that appropriate treatment methods can be recommended quickly. Wound measurement, is subjective for an overall assessment. The establishment of a high-performance automatic segmentation system is of great importance for wound care. The use of machine learning methods will make performing wound segmentation with high performance possible. Great success can be achieved with deep learning, which is a sub-branch of machine learning and has been used in the analysis of images recently (classification, segmentation, etc.). In this study, pressure wound segmentation was discussed with different encoder-decoder based segmentation models. All methods are implemented on the Medetec pressure wound image dataset. In the experiments, FCN, PSP, UNet, SegNet and DeepLabV3 segmentation architectures were used on a five-fold cross-validation. Performances of the models were measured in the experiments and it was demonstrated that the most successful architecture was MobileNet-UNet with 99.67% accuracy.