Research On Detection Of Colonic Polyps Based On Multiple Instance Learning

Colorectal cancer is a health problem that affects the length of human life today.It has a high morbidity and mortality rate.Colorectal cancer is mostly caused by malignant polyps,so to a large extent,colorectal cancer can be prevented by the early detection and removal of polyps.With the continuous development of medical technology,CT examination has important applicationge in the diagnosis of various clinical diseases.Compared with traditional colonoscopy,virtual colonoscopy based on CT data is a non-invasive detection method,it also can be observed many times.For the virtual endoscopy of the colon,the computer-aided detection of colon polyps has attracted more and more researchers’ attention,but the existing colon polyp detection algorithm is usually based on the accurate segmentation of the colon.The algorithm relies too much on the results of colon segmentation.We apply the multi-instance learning framework to colon polyp detection and propose a polyp detection algorithm based on multi-instance learning.Multi-instance learning is considered to be four machine learning frameworks juxtaposed with supervised learning,unsupervised learning,and reinforcement learning,and has become a research hotspot in the field of machine learning.Compared with the single correspondence of the sample labels of other methods,multi-instance learning is more suitable for the logical structure in the real scene because of the hierarchical representation structure of the training examples,which also makes it have superiority on inexact supervision.The unique advantages of Multi-instance learning have been widely used,especially in the field of computer-assisted detection.Multi-instance learning framework has become a great help for medical image-assisted detection methods.We convert the CT image polyp detection problem into a multi-instance learning problem by cutting the image into multiple regions,and use the convolutional neural network model pre-trained on a large data set as a feature extractor to extract the features of the example,then using pooling operation to obtain the feature expression of packet by aggregating the features of instances,and finally the detection result is obtained through the classification algorithm.In addition,considering the relationship between CT adjacent sequence images,a three-frame difference method is introduced to calculate the difference image,and the features of the difference image are extracted for feature fusion to obtain the new features of the instance.Finally,experiments prove that the colon polyp recognition algorithm and its improved algorithm proposed in this paper have better classification performance.