| With the explosive growth of CT imaging data,it becomes an extremely challenging task for physicians to diagnose lung diseases through imaging.Faced with huge amounts of CT images,a sufficient number,experienced,and steadystate physicians are required to complete the diagnosis.Otherwise,misdiagnosis and missed diagnosis are inevitable.With the aid of computer,computer-aided diagnosis technology can achieve a similar lesion search method similar to human diagnosis process,which provides primary screening and secondary selection for physicians with confirmed historical lung cases,thus greatly reduced missed diagnosis rate and misdiagnosis rate.Now,it has become a new approach for physicians to diagnose lung lesions.In the process of similarity diagnosis of lung cancer,the similarity retrieval method based on low-level CT image features play a positive role in improving the diagnostic accuracy and reducing the missed diagnosis rate,but there is a big difference between low-level features obtained through computers and high-level semantic features described by physicians,the semantic features of lung nodules are more high-level semantic description of the signs of pulmonary nodules by physicians.The high-level features of images are objective reflection of the inherent coherence.Deep learning forms a more abstract high-level representation through layer-by-layer feature combinations,rather than a series of complex image preprocessing and artificial feature engineering.This research mainly studies the similarity retrieval method of pulmonary nodules using the deep learning technology.The main research contents and innovations are as follows:(1)There exsists semantic inconsistency between the the low-level features based on hand-crafted and the high-level semantic features described by physicians.To tackle this problem,a retrieval method of lung nodules based on medical signs and convolutional neural networks(CNN)is proposed in this study.The nine medical signs values annotated by radiologists in LIDC database are firstly used to construct an accurate hash code for the training set.Then,the highlevel semantic features of pulmonary nodule images are extracted through the CNN,and the high-dimensional features are reduced with the principal component analysis(PCA)method.Therefore,the hash function can be inversely learned with the hash code obtained in the previous stage.Finally,by calculating the weighted Hamming distance between the query one and the database images,the results that are most similar to the query nodule image are returned.The experimental results show that the method in this study achieves higher accuracy and retrieval accuracy in the retrieval process of pulmonary nodule images.(2)Due to the existing two-step retrieval strategy are not suitable for the similarity retrieval parts,a retrieval method for pulmonary nodule with deep supervised hashing is proposed in this study.The method designs an end-to-end coding model based on CNN,which can directly converts the raw input image into compact binary codes.When constructing the hash function,the similarity between pair-wise images and sign labels of pulmonary nodule are simultaneously used,so that the learned binary code can not only maintain the semantic similarity of pulmonary nodule images but also be consistent with the label information.Finally,a two-step search strategy is proposed in the similarity retrieval process,namely,a series of candidate sets are firstly determined by rough search between different sign classes,and then the weighted Hamming distance is used to achieve fine-grained level ranking.Extensive experiments show that the hash function based on the supervised information with the tripletlabels is more discriminative.Moreover,the hash function constructed with deep network can well preserve the semantic similarity between the raw pulmonary nodule images.Compared with the mainstream similarity retrieval methods based on deep hashing,the method in our study obtains the highest accuracy of 93.54%. |
PDF Full Text Request