Computer Aided Diagnosis of Clustered Microcalcifications Using Artificial Neural Nets

Abstract

Objective: Development of a fully automated computer application for detection and classification of clustered microcalcifications using neural nets. Material and Methods: Mammographic films with clustered microcalcifications of known histology were digitized. All clusters were rated by two radiologists on a 3 point scale: benign, indeterminate and malignant. Automated detected clustered microcalcifications were clustered. Features derived from those clusters were used as input to 2 artificial neural nets: one was trained to identify the indeterminate clusters, whereas the second ANN classified the remaining clusters in benign or malignant ones. Performance evaluation followed the patient-based receiver operator characteristic analysis. Results: For identification of patients with indeterminate clusters a an Az-value of 0.8741 could be achieved. For the remaining patients their clusters could be classified as benign or malignant at an Az-value of 0.8749, a sensitivity of 0.977 and specificity of 0.471. Conclusions: A fully automated computer system for detection and classification of clustered microcalcifications was developed. The system is able to identify patients with indeterminate clusters, where additional investigations are recommended, and produces a reliable estimation of the biologic dignity for the remaining ones.