基于卷积神经网络的结肠镜图像息肉识别

Polyp Recognition in Colonoscopy Image Based on Convolutional Neural Network

肠息肉是一种结直肠表面突起的病变,并有可能进一步发展为结直肠癌,目前结直肠癌是我国第三大癌症.提前发现肠息肉并进行摘除是防止结直肠癌的有效方法,当前人工筛查仍存在较大程度的漏检测.对VGG16、ResNet18、GoogLeNet、EfficientNet和SeNet五种卷积神经网络模型进行肠息肉检测适应性研究,实验结果表明GoogLeNet模型性能较优.在2100张图像规模的数据集中,达到了准确率99.84%、精确率100%、召回率99.7%,可为结肠镜辅助息肉筛查提供支撑.

Intestinal polyp,a prominent lesion on the surface of the colorectal,may further develop into colorectal cancer,the third largest cancer now in China.Early detection and removal of intestinal polyps is an effective way to prevent colorectal cancer.Currently,there is still a large degree of missed detection in manual screening.In this paper,the adaptability of five convolutional neural network models of VGG16,ResNet18,GoogLeNet,EfficientNet,and SeNet was studied for intestinal polyp detection.The experimental results show that the GoogLeNet model performs the best.It achieves 99.84%in a dataset with 2,100 images when employed to judge the metric of accuracy.In addition,the precision and recall metrics achieve 100%and 99.7%,respectively.This work proves that GoogLeNet can support colonoscopy-assisted polyp screening.