Computer aided diagnosis in mammography with content-based image retrieval.

摘要

Computer-aided diagnosis (CAD) for breast cancer, a common form of cancer in women, has been an active research area. This work aims to investigate and develop CAD techniques for clustered microcalcifications (MCCs), which can be an important early sign of breast cancer. The contributions of this work include development of a database of cancer cases and algorithms for detection and classification of MCCs.;First, a database consisting of a large number of cases is built from different sources. To support the merging of cases from different data sources, a feature comparison study is conducted between mammograms from screen film and full field digital mammography (FFDM) systems. It is demonstrated that the features extracted from film and FFDM are highly correlated and there is no adverse effect on a CAD task of classification when used together.;Second, a spatial point process (SPP) approach is proposed to exploit the spatial distribution among different MCs in a mammogram directly during the detection process. This is different from the conventional approach in which detection algorithms are employed to first identify individual MCs in a mammogram, which are subsequently grouped into clusters by a clustering algorithm. The performance of the proposed approach is demonstrated to be superior to an existing method based on the support vector machine (SVM).;Third, in observation of the emerging of large databases from the picture archiving and communication (PAC) systems in the clinics, a retrieval driven approach is proposed for classification of MCCs. In this approach, for a case to be diagnosed (i.e., query), a set of similar cases is retrieved from a database and subsequently is used to train an adaptive classifier specifically for the query case using the technique of logistic regression. The proposed approach is demonstrated to lead to significant improvement in classification accuracy.;Moreover, the proposed adaptive classification approach is further developed using regularization techniques, where a prior is first derived from a baseline classifier and then used to regularize the adaptive classifier trained with the retrieved cases. The regularized adaptive classifier can be more computationally efficient, and is demonstrated to achieve further improvement in performance.