Classification Of DFU Images Based On Deep Semi-Supervised Learning

Diabetic foot ulcers(DFU)are a serious complication of diabetes,with about 1/3 of people with diabetes having a foot ulcer.The global rise in the number of people with diabetes has led to an increase in the number of diabetic foot ulcers.In addition,the high rate of recurrence and mortality of DFU has put tremendous pressure on the healthcare system.In general,condition classification is the first step in carrying out treatment.Achieving DFU image classification through technological means,so as to help patients self-monitor,can greatly reduce medical stress and facilitate follow-up treatment.Currently,most DFU image classification methods are based on deep supervised learning.However,deep supervised learning relies on labeled data,which requires manual labeling by professionals and consumes a lot of resources such as time and manpower.Compared with labeled data,unlabeled data is larger and easier to obtain.Therefore,in this paper,deep semi-supervised learning is used to make full use of unlabeled data to improve DFU image classification accuracy.Meanwhile,deep semi-supervised learning needs to set a reasonable threshold to decide whether the unlabeled data can be used to optimize the model,and a new flexible threshold method using a logarithmic function is proposed in this paper.The experimental results show that by introducing unlabeled data into the model,the model over-fitting problem is effectively solved and the classification performance of the model is improved;the proposed threshold method outperforms the fixed threshold method when dealing with similar and complex samples,and outperforms other flexible threshold methods when classifying unlabeled imbalanced data.In addition,data augmentation is also an effective means to improve the classification performance of the model.Existing data augmentation techniques rely on the underlying transformation operation to change the image for the purpose of enhancing the data.However,the transform operations suitable for natural images or pervasive object images are not always applicable to DFU medical images.In this paper,we propose a new adaptive augmentation pool construction method.First,we find some suitable base transformation operations as the initial augmentation pool and then rely on extracting image features to build the adaptive augmentation pool based on the initial one.The experimental results verify the effectiveness of the proposed construction method,which can be used as a reference for other kinds of medical image augmentation operations.Finally,this paper uses ensemble learning to further improve the model classification performance.In plurality voting,only the number of votes is considered,ignoring the fact that different output results do differ in confidence.Therefore,in this paper,we propose a new ensemble learning method based on voting with expertise,which views high-certainty output results as expert opinions.In case of disagreement in model classification,expert opinions are given priority,followed by the number of votes.The proposed method can effectively overcome the effect of low-confidence but multiple-vote classes and the experimental results verify the effectiveness of the method.In summary,this paper improves the existing work in three aspects: deep semi-supervised learning,data augmentation,and ensemble learning,and proposes three new methods: flexible threshold adjustment of logarithmic functions,adaptive augmentation pool construction,and expert opinion voting integration.Experiments on DFU medical image datasets show that the three methods proposed in this paper can effectively improve the classification accuracy of DFU images compared with previous research work.