Research On The Framework Of Wound Segmentation Based On Deep Convolutional Neural Networks

Chronic wounds are common in clinical surgery,which always takes a long time to heal and is susceptible to infection.Chronic wounds bring great pains to patients and require long-term care and treatment which involves high costs for health services.This brings great burden to the society and the family.With the development of computer technology,the research and application of digital image processing technology in medical field is wider and deeper.The automatic measurement of wound area using these techniques can improve the accuracy of assessment and reduce the cost of medical treatment,which is of great theoretical and practical significance.The core process in measuring wound area is to identify wound boundary using image segmentation algorithms.The existing deep segmentation models based on convolutional neural network have achieved remarkable results in image segmentation task with the ability to learn and extract features automatically.However,these deep segmentation networks are hard to train with the lack of enough wound images.At the same time,the segmentation model trained with a small amount of wound images can hardly learn some potential semantic information.This paper conducts a study of these issues and the main works are as follows:First,the existing deep learning models towards image segmentation are complex with a large number of parameters and it relies on a large amount of annotation data to avoid overfitting during the training process.On the other hand,the large-scale network models with high computational complexity can hardly be applied on resourceconstrained device.To address this problem,we propose a wound segmentation model named WoundSeg.By adjusting the structure to simplify the large scale segmentation network in terms of the number of layers and channels,the lightweight network with learning parameters greatly reduced can be trained efficiently with the small wound dataset.Compared to the previous works,the segmentation accuracy is significantly improved with5-fold cross validation accuracy(mIoU)more than 84%.Second,the wound segmentation is easy to make mistakes because of the residues of tissue fluid and drugs.Meanwhile the segmentation model based on deep CNNs often leads to large error on the boundary due to the structure of the network.To address this problem,semantic information of the chronic wound image is used to post process the segmentation results from the deep segmentation network:(1)the holes in wound regions are filled with morphological operations based on the simple connectedness of wound regions,and some small regions are eliminated selectively;(2)the skin segmentation is used to verify the wound regions predicted by wound segmentation model,and further eliminate some wound regions misjudged from confusing backgrounds;(3)the full connected CRFs method considering the homogeneousness of wound regions is used to improve segmentation boundary,and the features used are extracted from superpixels.As a supplement to the DCNNs based deep segmentation model,the post-processing methods eliminate most of the segmentation mistakes and improve the segmentation boundary effectively.