The aim of this project is to develop a computer-assisted consultation system for clinical radiologists in the classification of benign and malignant (cancer) calcifications on mammograms. An automated screening of digital mammograms can be further developed when this system is fully tested in various clinical settings. The success of this project will inspire revolutionary improvements in other cancer diagnostic procedures through further research and development. We have demonstrated that the newly developed artificial visual neural networks and training methods (by presenting microcalcification patterns in the training session) can detect microcalcifications on mammograms. A similar method can be used for the detection of lung nodules on chest radiographs. The proposed feature extraction methods and artificial visual neural networks simulate radiologists' routine practices in reading mammograms. Radiologists' viewing patterns and decision making processes will be modeled and converted to computer readable form. Preliminary studies have shown the promise of this approach. The Phase I study will address issues related to the differentiation of malignant from benign microcalcifications based on radiographs taken from breast tissue specimens. Based on Phase I study, the plan of the Phase II study is to (i) conduct the research involving classification of microcalcifications on clinical mammograms, (ii) analyze various features of benign and malignant microcalcifications and seek potential correlation with the learning patterns of the artificial visual neural network, and (iii) integrate the research outcome and implement a prototype consultation system for clinical use.
