Convolution Neural Network with Coordinate Attention for Real-Time Wound Segmentation and Automatic Wound Assessment
Open Access
- 23 April 2023
- journal article
- research article
- Published by MDPI AG in Healthcare
- Vol. 11 (9) , 1205
Abstract
Background: Wound treatment in emergency care requires the rapid assessment of wound size by medical staff. Limited medical resources and the empirical assessment of wounds can delay the treatment of patients, and manual contact measurement methods are often inaccurate and susceptible to wound infection. This study aimed to prepare an Automatic Wound Segmentation Assessment (AWSA) framework for real-time wound segmentation and automatic wound region estimation. Methods: This method comprised a short-term dense concatenate classification network (STDC-Net) as the backbone, realizing a segmentation accuracy–prediction speed trade-off. A coordinated attention mechanism was introduced to further improve the network segmentation performance. A functional relationship model between prior graphics pixels and shooting heights was constructed to achieve wound area measurement. Finally, extensive experiments on two types of wound datasets were conducted. Results: The experimental results showed that real-time AWSA outperformed state-of-the-art methods such as mAP, mIoU, recall, and dice score. The AUC value, which reflected the comprehensive segmentation ability, also reached the highest level of about 99.5%. The FPS values of our proposed segmentation method in the two datasets were 100.08 and 102.11, respectively, which were about 42% higher than those of the second-ranked method, reflecting better real-time performance. Moreover, real-time AWSA could automatically estimate the wound area in square centimeters with a relative error of only about 3.1%. Conclusion: The real-time AWSA method used the STDC-Net classification network as its backbone and improved the network processing speed while accurately segmenting the wound, realizing a segmentation accuracy–prediction speed trade-off.Keywords
Funding Information
- Natural Science Foundation of Chongqing Municipality (2020ZX 1200048)
This publication has 30 references indexed in Scilit:
- Automatic Skin Lesion Segmentation Using Deep Fully Convolutional Networks With Jaccard DistanceIEEE Transactions on Medical Imaging, 2017
- SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image SegmentationIEEE Transactions on Pattern Analysis and Machine Intelligence, 2017
- Automated Melanoma Recognition in Dermoscopy Images via Very Deep Residual NetworksIEEE Transactions on Medical Imaging, 2016
- A Label-Scaled Similarity Measure for Content-Based Image Retrieval2016
- Fully Convolutional Networks for Semantic SegmentationIEEE Transactions on Pattern Analysis and Machine Intelligence, 2016
- Learning Deconvolution Network for Semantic Segmentation2015
- U-Net: Convolutional Networks for Biomedical Image Segmentation2015
- Pattern Recognition of Lower Member Skin Ulcers in Medical Images with Machine Learning Algorithms2015
- A novel and accurate technique of photographic wound measurementIndian Journal of Plastic Surgery, 2012
- An overview of techniques used to measure wound area and volumeJournal of Wound Care, 2009