Research On Key Issues Of Medical Image Retrieval For Lung Cancer Computer-aided Diagnosis Application And System Implemention

A large number of meaningful medical images have been created in the medical institutions, by the development of the medical imaging technology and the widly use of computer-aided diagnosis (CAD) system. How to organize and management the medical images more effectively and even use these images to provede doctors more scientific basis for the diagnosis has become the most urgent problem in medical imgae field. The content-based medical image retrieval (CBMIR) arises rapidly in such circumstances. In order to maximize the value of the CBMIR, this thesis researched and implemented the cntent-based medical image retrieval for lung cancer CAD (CBMIR-lung cancer CAD) based on CBMIR technology research. The technology retrieval the lesions, suspicious lesions which the lung cancer CAD detected or the region of interest (ROI) not detected, and show the images from the database which have the same pathological diagnosed features. This can help the doctor for clinical diagnosis, illness tracking and prognosis research.In this study, the applications of CBMIR in medical image field and key techniques of CBMIR were summarized. The main issues of CBMIR-lung cancer CAD are as follows:(1) Feature extraction. The thesis extracted the features of the pathological region from shape and texture based in lung CT images as comprehensive as possible. And extracted two improved characteristics simultaneously:(a) obtained the contour of the region based on Otsu segmentation and Snake, and expressed the contour moment invariants with the Hu moment invariants,(b) Extracted the Gabor txture features from the filtered images based on local binary pattern (LBP).(2) The optimization of the features in retrieval. Optimized the features extracted in (1) with integrating multi-feature and fusing features by principal component analysis (PCA). And examined the precision of two methods with SPN, GGO, malignant pulmonary nodules and non-nodules example images in retrieval experiments, optimized the methods from the results.(3) Studied and achieved a3-demensional SIFT (Scale Invariant Feature Transform) extension, proposed the3-degree3D SIFT and improved its process to obtain the tilt. In order to resolve the large dimension of the descriptor and complicated algorithm in certifying the rotation invariance, the thesis raised a mothed which combines three demensional ring model and PCA-SIFT method, the descriptor obtained with this method has low dimentions and natural rotation invariance.(4) Connected the retrieval system to the CAD interface, the system has these functions: image preprocessing, image and feature database management, feature extraction, image retrieval with multi-method,2D and3D SIFT image registration and results feedback.