The Research On Classification Model And Quantitative Method Of Diabetic Foot Based On Detection Of Wound

Diabetic foot(DF)is one of the most common and serious complications of diabetes.The prognosis of DF is very poor,and the awareness rate and the visiting rate of DF are low in China.In the treatment of DF,the healing cycle of diabetic foot ulcers(DFU)is longer.It is difficult for doctors to quantify the progress of diabetic foot ulcer healing in the short term treatment with the existing grading standard,so as to evaluate the effectiveness of the treatment plan.Therefore,early detection,early diagnosis and early treatment of diabetic foot is particularly important.In recent years,computer-aided diagnosis of DF and analysis of DFU based on deep learning are widely studied.The former used the thermal imaging technique to diagnose DFU.The latter used neural network to analyze DFU.Clinicians in China use the Wagner grading standard to assess the condition of DF patients.Although the existing methods can help doctors diagnose and quantify DF,they still face four problems.First,the diagnostic method of DF based on thermography is sensitive to the surrounding environment and is prone to cause errors.Second,the analysis method of DFU based on deep learning is only for the analysis of DFU wounds,and can not distinguish between DFU wounds and non-DF chronic wounds.Third,the classification of DF images by using neural network directly is not as effective as expected.Fourth,the Wagner grading standard used widely in China lacks the subdivision,which is difficult to quantify the healing progress of DFU wound in the short term.For the above analysis,the research work of this paper mainly includes the following three points.First,in order to solve the problem of inaccurate classification of DF and non-DF images,the Fusion Network(Fusion Net)fusing object detection information is proposed.The model detects ulcer wounds from a chronic wound image through a object detection module.It extracts global foot features and local wound features by using Res Net and fuses these two features to classify DF and non-DF images.In this paper,the object detection model Cascade Attention Det Net(CA-Det Net)is proposed to further improve the accuracy of the object detection module.The experimental results show that this model can accurately classify DF and non-DF images,and its Area Under Curve(AUC)score is 94.87%.In addition,the importance of global foot features and local wound features in the classification of DF and non-DF images is analysed in detail.It is proved that it is not easy to diagnose DF only according to the wound.Therefore,it is suggested that the diagnosis of DF should be based on the whole foot regional condition and the local wound condition.Second,in order to improve the performance of classification model fusing object detection information,the Fusion Segmentation Network(Fusion Seg Net)fusing segmentation information is proposed.The model segments the wound pixels from a chronic wound image by a segmentation module.It uses the pixel information to force the network to accurately extract local wound features and finally integrates local wound features and global foot features to classify DF and non-DF images.In addition,this paper also introduces the foot ulcer segmentation model called Ensemble Segmentation Network,which further improves the accuracy of the segmentation module.The experimental results show that this model can greatly improve the classification accuracy,and the AUC score is 98.95%.Third,in order to solve the problem of quantification of DFU,a method based on foot ulcer segmentation for DFU quantification is proposed.The rough quantization method converts the wound pixel area obtained by the segmentation model into the actual area through the marker(ruler)as the rough quantization index.The fine quantization method segments the pixel area of the ulcer wound and the skin around the wound from the image and transforms them into a ratio as the fine quantization index.The experimental results show that the fine quantization method using the Link Net can achieve excellent segmentation performance,with an average Dice score of 86.38% and an average Intersection over Union(IOU)score of 76.04%.In addition,this paper shows the diabetes foot intelligent medical platform based on We Chat.Our team cooperate with clinicians in the Shanghai Eighth People’s Hospital to apply the classification model of DF and wound quantification method in clinical treatment.The results of the classification model and the quantitative methods can assist clinicians to diagnose DF and analyse patients’ ulcer wound healing progress,which can help clinicians adjust the treatment plan in time.