Application Of Background Balance Optimization Method In Fine Grained Medical Image Recognition

Fine-grained image recognition is a popular research direction in the field of computer vision.Since medical image datasets such as skin diseases and gastrointestinal diseases have typical fine-grained features such as small inter-class differences and large intraclass differences,fine-grained medical image classification is challenging.sex.In recent years,computer-aided diagnosis has developed rapidly,which also puts forward new requirements for medical image classification.Starting from the idea of finegrained image classification,this paper makes the following work:1.Analysis of background balance optimization mode: In the experiment,it is found that the foreground-background imbalance problem exists widely in finegrained image classification research.The root cause of the foreground-background imbalance problem is that the number of foreground samples is not equal to the number of background samples.number of cases.For example,using Vi T as the basic classification network for the CUB bird dataset can achieve an accuracy rate of 91.03%,and when we use the corresponding algorithm to delete part of the background,the accuracy rate can be increased to 91.94%.All experiments in this paper are based on the foreground-background imbalance problem.2.Propose an optimization method for fine-grained natural image background balance: Due to the lack of fine-grained research in the field of medical image classification,this paper first verifies the proposed ideas on fine-grained natural datasets,and then applies them to fine-grained medical image datasets.Based on the problem of unbalanced foreground and background,this paper proposes three background balance optimization methods: fixed cropping,cropping based on scale normalization,and Background Balance Optimization(BBO)algorithm based on image region expansion Foreground and background imbalance problem.For the CUB bird data set and the Stanford Dogs dog data set,the BBO algorithm proposed in this paper has achieved good results on the basis of Res Net and Vi T networks,which proves that the method in this paper has generalization on data sets and network models.Using the BBO-based Vi T network basic classification accuracy in the CUB dataset can reach 91.94%,and the performance is significantly improved and is better than the current mainstream classification model.3.Propose a fine-grained medical image classification method based on background balance optimization: For the particularity of medical image datasets,this paper proposes three background balance optimization methods,including fixed cropping,cropping based on mask dilation and BBO algorithm based on mask dilation,to improve the problem of front-background imbalance in medical image classification tasks.For the ISIC skin disease image data set and the Kvasir gastrointestinal disease data set,after fine-tuning the basic networks such as Res Net and Vi T,the use of the BBO algorithm has a certain improvement.At the same time,this paper also conducts generalization experiments on the existing mainstream medical classification models.Using BBO based on the ISIC2018 model by Nils et al.can increase the accuracy rate to 86.67%.Using BBO based on the ISIC2019 model based on Alxiang et al.can increase the accuracy rate to 74.17%.4.Propose a fine-grained medical image segmentation method based on background balance optimization: Aiming at the mainstream segmentation task of current medical image research,this paper also conducts relevant experiments,and conducts generalization experiments on the existing mainstream medical segmentation models.For the context encoder network(CE-Net),this paper uses BBO to convert the Jacquard index(JA)increased to 79.1%.For the new boundaryaware convolutional network(BA-Net),this paper uses BBO to increase the Jacquard index(JA)to 81.2%,which proves the front-background imbalance theory and background balance optimization proposed in this paper.The algorithm is also applicable to segmentation tasks,and it also has generalization ability for segmentation models.