Research On Large Medical Image Retrieval And Segmentation Algorithms

The better interpretation of images through computer-assisted analysis is a long-standing problem in the field of medical imaging.In image understanding,recent advances in machine learning help identify,retrieve,and segment medical images.The core of deep learning is to learn layered feature representations from data rather than manual features designed based on domain-specific knowledge.In this way,deep learning quickly proved to be the most effective method,and to obtain better performance in various medical applications.This article describes the related technologies involved in “CT lung nodule image retrieval and digital pathological image segmentation”.The content-based medical image retrieval(CBMIR)system can find effective methods for diagnosis and treatment of various diseases from historical cases,and it is also an efficient management tool for processing large amounts of data.Automated image segmentation,which aims at automated extraction of organ boundary features,reduces the search space for improving the efficiency of image retrieval.The research on image retrieval and segmentation in this paper mainly includes the following points:(1)Due to the difficulty in collecting medical images,most researchers focus their research on migration studies,but these methods often overlook some of the inherent characteristics of medical images.In this paper,200+GB CT pulmonary nodules images were collected in a large open challenge.Data was preprocessed using medical domain knowledge,data sets were generated,and models were directly trained on medical images using a deep learning framework.(2)In this paper,through the improvement of FC2 layer in InRes-Net,the model can learn image features and hash function at the same time,which enables the deep learning technology to automatically extract and encode CT pulmonary nodules image features.(3)In this paper,a “coarse+fine” search strategy is used to obtain an efficient CT pulmonary nodule image retrieval system.(4)High-resolution digital pathology images provide rich information on the morphological and functional characteristics of biological systems,but the size of single images exceeds 150,000 x 150,000.This paper uses the Patch Sampling technique to sample the original data,which solves the problem of large training data size and also adds training samples.(5)In view of the complex clinical features of digital pathological images,this paper uses a GICN structure on different feature layers to form a multi-level feature pyramid,which resolves the contradiction between classification and positioning;the global average pooling structure of different sizes constitutes a A multi-scale feature pyramid that facilitates the integration of contextual information of different sizes and regions.This paper proposes corresponding solutions for CT pulmonary nodule image features,and finally obtains an average accuracy of 0.73 on the retrieval system.Because the CT pulmonary nodule images had no mask label and did not meet the conditions for training the segmentation model,the model was finally selected on a set of labeled digital pathological images.The segmentation accuracy was 0.63,which was more than 10% higher than the previous method.The above experiments verify the effectiveness of the proposed algorithm.