基于 Swin Transformer 网络与 Adapt-RandAugment数据增强方法的小肠胶囊内镜图像分类方法研究OACSTPCD

Objective To propose a method for classifying small bowel capsule endoscopy images by combining the Swin Transformer network with an improved Adapt-RandAugment data augmentation approach,aiming to enhance the accuracy and efficiency of small bowel lesion classification and recognition.Methods An Adapt-RandAugment data augmentation approach was formulated based on the RandAugment data enhancement sub-strategy and the principles of no feature loss and no distortion when enhancing small bowel capsule endoscopy images.In the publicly available Kvasir-Capsule dataset of small bowel capsule endoscopic images,the Adapt-RandAugment data augmentation approach was trained based on the Swin Transformer network,and the convolutional neural networks ResNet152 and DenseNet161 were used as the benchmarks to validate the combined Swin Transformer network and Adapt-RandAugment data augmentation approach for small bowel capsule endoscopy image classification.Results The proposed algorithm gained advantages over ResNet152 and DenseNet161 networks in the indicators,which had the macro average precision(MAC-PRE),macro average recall(MAC-REC),macro average F1 score(MAC-Fi-S)being 0.383 2,0.314 8 and 0.290 5 respectively,the micro average precision(MIC-PRE),micro average recall(MIC-REC)and micro average F1 score(MIC-Fi-S)all being 0.755 3,and the Matthews correlation coe-fficient(MCC)being 0.452 3.Conclusion The proposed small bowel capsule endoscopy image classification method based on Swin Transformer network and Adapt-RandAugment data augmentation approach behaves well in classified recognition efficiency and accuracy.[Chinese Medical Equipment Journal,2024,45(6):9-16]