Research On Sign Recognition Method Of Lung Nodules Based On Image Retrieval With Supervised Hashing

Medical signs are important for the diagnosis of lung diseases,and there is a strong correlation between medical signs and lung lesions,which can help physicians to determine the degree of benign and malignant lung cancers.With the wide application of medical CT(Computed Tomography),lung CT imaging data is growing explosively,greatly increasing the workload of radiologists,and easily leading to misdiagnosis in the process of disease diagnosis.Computer-aided diagnostic technology can effectively reduce the workload of physicians,and assist them completing the medical imaging-based disease detection and diagnosis,and improve the diagnostic stability and efficiency.In the vast number of confirmed medical imaging data,images similar to the query one can be obtained by medical image retrieval technology.The medical signs and diagnostic schemes of these diagnosed historical lesions provide a strong reference for physicians.This is an effective method for computer-aided diagnosis of lung diseases.Because there are a wide variety of medical signs in lung,and the performances of them are complex and diverse,physicians have difficulty in diagnosing lung diseases accurately and stably when screening a large number of medical images.In view of this problem,the paper proposes a medical sign recognition method of lung nodules based on image similarity retrieval to help physicians to correctly identify the signs of lung nodules and provide references for benign and malignant diagnosis of lung nodules.The method uses a supervised hashing-based image retrieval technology to obtain similar lesion images quickly from a lung nodule CT image library,and then recognizes medical signs on the basis of the retrieval results.The specific research contents are as follows:(1)Feature extraction of lung nodules based on convolution neural network.Feature extraction is the basis for the realization of lung nodule CT image retrieval.Because several categories of imaging signs(mixed signs)could appear in one nodule,handcrafted features are difficult to effectively express these sign information.Aiming at this problem,the paper presents a convolution neural network(CNN)-based feature extraction method for lung nodules.The method first uses the parameter-shared CNNs to train the single-sign data,and adjust the network parameters to effectively identify the single-signs.Then,the trained parameters are transferred to the network of training mixed-sign data(lung nodules images),and these parameters are fine-tuned by a loss function and an error back propagation mechanism.Thus,the semantic features of lung nodules are obtained from the fully connected layer.Finally,in order to evaluate the extracted image features of lung nodules,support vector machine and extreme learning machine are used to classify these features according to the type of medical signs.And the classification effects of the texture features are also compared.The experimental results show that the image features extracted with our method can more effectively express the sign information of lung nodules.(2)Supervised hashing-based lung nodule image retrieval and sign recognition.In order to retrieve the similar images quickly and accurately from the lung nodule CT image library,this paper presents a retrieval method for lung nodule images based on supervised hashing.Principal component analysis algorithm is first used to process the extracted semantic features of lung nodules,and the image features are mapped into short hash codes by the constructed hash functions with supervised information to preserve the semantic similarity in the original feature space.Then,by designing adaptive weight vectors,weighted Hamming distance is used to measure the similarity of lung nodule images.Finally,the k-nearest neighbor algorithm is used to identify the sign categories from the retrieved results.The experimental results on the public dataset show that the retrieval precision of our method can reach 87.29% when the length of hash code is 48,which is helpful to improve the recognition rate of lung nodule medical signs.In addition,experiments in clinical diagnostic data also illustrate that our method can effectively identify the signs of lung nodules.