Research On Medical Image Retrieval Using Deep Learning Technology

With the widespread application of imaging modalities such as X-ray,computed tomography(CT)and nuclear magnetic resonance(MR),the retrieval and utilization of medical image big data has become a hot research topic.Medical images can provide important imaging information for clinical diagnosis.How to assist clinicians to quickly and accurately retrieve the required images from the massive image library has become an important research topic.In the traditional content-based image retrieval method,It mainly uses the color,texture and shape of the image as the retrieval basis.However,there is a "semantic gap" between these features and the high-level semantic features that people need,so the retrieval performance has been limited.In recent years,with deep learning technology in major computer vision competitions have made breakthroughs.Its good feature expression ability has also shown excellent performance in the field of natural image retrieval.Therefore,deep learning technology brings great opportunities for medical image retrieval.Unlike natural images,medical image retrieval is a challenging task.First,most medical images are grayscale images with blurred boundaries,noise,and poor contrast.Therefore,traditional feature representation methods are difficult to capture the nuances of their features.Second,the same organization has different images for different patients,different modalities,and different imaging devices,and even there may be differences between different frames of the same modality;Second,the same organization has different differences between different patients,different modalities,and different devices,and there are even many differences between different frames of the same modality.Third,the distribution of medical image categories is uneven,and there is a problem of category imbalance.The most important thing is that data tag acquisition requires professional labeling,which is time consuming and laborious,It is difficult to get enough medical image retrieval annotation data sets.Since the traditional content-based medical image retrieval method is limited by the feature representation ability,medical image retrieval accuracy and efficiency are difficult to further improve.The paper proposes a medical image retrieval idea based on deep learning technology,This paper proposes a medical image retrieval idea based on deep learning technology,and designs and trains a medical image retrieval model based on convolutional neural network.Its main purpose is to improve the accuracy andefficiency of medical image retrieval.The main research contents of this thesis are as follows:(1)The advantages and disadvantages of the traditional Content-Based Medical Image Retrieval(CBMIR)algorithm are studied.A multi-feature fusion rule is proposed to realize a medical image retrieval algorithm based on multi-feature fusion.The accuracy and effectiveness of the CBMIR algorithm depends on the feature representation method and the similarity matching index method.The paper analyzes color features,local texture features and HOG feature representation methods.A single feature representation method does not have good robustness when images with complex backgrounds,noise,or blurred boundaries.Therefore,according to the characteristics of medical images,a multi-feature weighting fusion rule that combines gray and texture is proposed to improve the representation ability and robustness of features.Experiments show that the feature representation ability through multi-feature fusion is stronger than single feature representation ability,and it has better accuracy in medical image retrieval.(2)A medical image retrieval method based on deep convolutional neural network is proposed and implemented.Traditional CBMIR methods are easily limited by the ability of feature representation,which leads to low retrieval accuracy.With the breakthrough progress of deep learning technology in many fields,convolutional neural networks have strong semantic feature representation ability.Therefore,we propose a medical image retrieval framework based on Deep Convolutional Neural Network(DCNN).A deep convolutional neural network classification model was established and trained.The model was used to automatically extract medical image features,and then the performance of different layer features was compared and analyzed.Contrastive experiments show that the method based on deep convolutional neural network improves the accuracy of medical image retrieval from 0.86 to 0.93,which proves that the proposed method is more effective.(3)An efficient method for medical image retrieval based on deep hash convolutional neural network is proposed and implemented.Since the feature dimensions of the traditional CBMIR method and the DCNN-based medical image retrieval method are usually relatively high,it is time-consuming to calculate these high-dimensional features represented by real numbers,and a large memory space is required.Therefore,based on the above research,we propose a medical image retrieval algorithm based on deep hashing for the problem of low computational efficiency and the inability to express similar semantic information.The similaritymatching is performed according to the Hamming distance of these binary codes,thereby reducing the time complexity of the retrieval and ensuring the retrieval precision.Experiments show that the newly proposed method only needs 0.037 seconds to retrieve an image on average,which proves that the proposed method effectively improves the retrieval efficiency of medical images.