基于改进的Faster R-CNN的息肉目标检测和分类方法

Object detection and classification of polyps based on improved Faster R-CNN

为了解决结肠镜下腺瘤性息肉和增生性息肉不易分型的问题,提出一种基于改进的Faster R-CNN的目标检测及息肉分类模型.在数据预处理阶段,对原有的2426张息肉图像(1582张腺瘤性息肉图像,844张增生性息肉图像)通过2种方式进行图像增强,并且通过改进的特征提取、边界框回归以及非极大值抑制的网络,用602张图像(386张腺瘤性图像,216张增生性息肉图像)进行测试.通过实验证明,在交并比(IoU)取0.5时,获得了86.8%的平均精度均值,相较于改进之前提升了2.3%.实验结果验证了该模型的潜在临床应用价值.

In order to solve the problem of difficult classification of adenomatous polyps and hyperplastic polyps under colorectoscopy,an object detection and polyp classification model based on improved Faster R-CNN was proposed.Polyp images(1582 adenomatous polyp images,844 hyperplastic polyp images)were image-enhanced in two ways,and 602 images(386 adenomatous images,216 hyperplastic polyp images)were tested.It is proved by experiments that when the intersection-union ratio(Io U)is set to be 0.5,an average precision of 86.8%is obtained,which is 2.3%higher than that before the improvement.The experimental results verified the potential clinical application value of this model.