Computer Aided Detection System Based Bag Of Words

Traditional Computer-Aided Detection usually classifies the breast X-ray images according to low-level features which prone to miss detection for some lesions. In order to describe the latent semantic features of images, we present a computer-aided detection system based on the bag of words model to finish the classification. First, removing the background blocks which surrounding the organs of interest. Then, blocks belong to the training samples will be clustered to get a visual vocabulary to express the latent semantic of all images, ultimately using the visual vocabulary as features input to the computer-aided detection system. Also we have proposed the method which combining genetic algorithm and the bag of words model to find the best parameter of the visual vocabulary. In the end, we further determined the coefficient of images corresponding to the visual words according to the sparse representation method of the biometric identification technology, resulting in the most sparsely linear combination of the medical images, thereby improved the performance and the speed of the detection.Specifically, the main contents of this thesis are:(1) We present the segmentation algorithm for X-ray mammography images as well as extraction process of low-level features.(2) Visual vocabulary learning has been finished via K-means clustering;Then the optimal parameters of the clustering are calculated by genetic algorithm.(3) Latent semantic features of images are acquired from the visual vocabulary histogram vectors.(4) The classification algorithm is designed based on latent semantics,Then the performances of the proposed method and the traditional algorithms are compared and analyzed.(5) Sparse vocabulary representation of the entire images has been studied and employed for computer aided detectionExperimental results show that proposed computer-aided detection method based on the bag of words model can effectively improve the accuracy of detection while maintaining the sensitivity and the performance.