Abnormity Detection And Analysis Of Mammography Based On Hashing

With the development of the society and living standards, the incidence of breast cancer increases year by year. Breast cancer has been the second deadly cancer of women. The computer aided detection(CAD) systems aim at improving the survival rate and the life quality of patients and assisting doctors to make the final diagnosis via detecting abnormities automatically. As one of the main symptoms of breast cancer, mass detection is of paramount importance in CAD systems. Considering the variety of mass shapes and the huge size of mammograms, it is necessary to propose a real-time mass detection system which can adjust to the large-scale database. In this paper, we try to detect suspicious regions automatically by analyzing the pathologic characteristics of mass and applying artificial intelligence and machine learning methods. Therefore, we could assist radiologists in a more scientific way.Based on hashing, content-based image retrieval(CBIR) and semantic features, this paper goes into suspicious lesion detection methods for mammography. Main works are summarized as follows.First of all, a fast hashing-based mass detection method is proposed. After the research on various hashing algorithms, we integrate histogram of oriented gradient(HOG) and kernel-based supervised hashing(KSH) to design a universal detection method suitable for different mass types and large-scale data. The proposed method could conquer the nonlinear characteristic of mammographic classification. Besides, the supervised information what we do have are fully utilized in our proposed method. Due to the efficiency of hashing, the proposed method make the real-time detection system come true and improve the practicability of detection systems.In order to represent masses in a more accurate way, this paper proposed two novel feature extraction methods named pairwise Bo W and hierarchy-weigh Gist. Breast masses perform a unique growth distribution characteristic. Taking this character into consideration, the proposed pairwise Bo W model expends traditional Bo W to a multi-dimensional structure and could combine spatial information and local feature information. The proposed hierarchy-weigh Gist feature could express the distribution of mass and refine the global Gist to contain more texture information. Two proposed features could achieve the effectiveness of mass description for breast images.Finally, this paper proposed the multi-feature fusion retrieval and analysis algorithm based on hashing to predict the mass probability and have a further analysis. Thus, we could assist radiologists to make the final diagnose in a more reasonable way. To fully describe breast image, different specific features of every suspicious region are extracted. The graph model is applied to the multi-feature fusion retrieval algorithm based on hashing. Consequently, the probability of suspicious regions could be predicted from the retrieval results and reduce the number of false positive. The overall performance of our detection system will be improved by this way.The experimental results show that the hashing based abnormity detection and analysis method proposed in this paper can distinguish masses from normal tissues more precisely and improve the mass detection accuracy while maintaining a low false positive rate at the same time.