SWS-NET:An Image Segmentation Framework For Chronic Wounds Based On Self-Supervised Learning

Chronic wounds are the wound that cannot be cured in a short period of time through normal treatment processes.During the wound treatment process,clinical doctors need to continuously measure and evaluate the wound to monitor the wound healing process and treatment effect.The use of supervised deep learning methods for auxiliary diagnosis and evaluation of wound surfaces requires a large amount of labeled data,and chronic wound related datasets are currently difficult to meet this condition.In order to improve the segmentation effect of the wound image with a small number of labeled samples for training,this paper proposes a self-supervised image segmentation framework,and uses active learning in the downstream data filtering phase.The experimental results show that after the contrastive learning pre training of full amount of unlabeled data,the segmentation framework can achieve significant improvements in precision,recall and MIo U(up to 9%)by using only a small amount of labeled data filtered by active learning in the downstream.It is proved that the idea of self supervised learning is effective in improving the segmentation effect when the labeled data of the wound image is scarce.The main research content and work of this paper are as follows:1)This paper proposes a network framework for chronic wound image segmentation based on self supervised learning,which named SWS-NET.The network structure has been optimized and the convolutional layer has been modified to enhance feature extraction ability.Through ablation experiments,it has been proven that the modification of the network structure is effective in improving the segmentation effect of chronic wound image.2)A loss function suitable for chronic wound segmentation was constructed.This improves the effect of segmentation task.Discussed the influence of the weight of different components in the loss function on the segmentation effect,mainly the binary cross entropy loss and Dice coefficient loss,and selected the loss function combination with the most stable training process and the best effect.3)Before conducting training of supervised segmentation,use unlabeled data for self supervised pre training.By comparing and learning the feature information of distorted views,minimize the redundancy between the two views,and learn useful knowledge that is beneficial for wound image segmentation.In this process,we also focused on exploring the impact of different distortion methods on segmentation tasks,and ultimately summarized an optimal distortion strategy.4)Active learning is integrated into the data set processing phase,and a two-stage data processing strategy is proposed according to the representativeness and uncertainty of the data.The most potential samples are selected through the combination of Bi K-Means clustering representativeness algorithm and wound edge threshold adaptive uncertainty algorithm,so as to reduce the labor cost in the data labeling phase and Improve the segmentation task in various evaluation indicators(especially recall).5)In order to facilitate the collection and verification of wound data,and to apply the latest technology to actual wound measurement,this paper combines the proposed segmentation model with the actual needs of chronic wound management to develop a chronic wound information management system.Four functional modules are designed: permission management,patient information,wound management,and wound diagnosis.Developed using django as the backend framework,We Chat mini program as the front-end interface,deployed and tested in production environment using nginx and uwsgi,ultimately achieving a good combination of algorithm model and engineering practice.