Effect of Reference Image Retrieval on Breast Mass Classification Performance: ROC Analysis

Retrieval of similar cases with the diagnostic and therapeutic results as a reference may be useful in differential diagnosis of abnormalities. Image retrieval method for breast masses on mammograms has been investigated in our previous study, and the result indicated the potential advantage of a machine learning technique using sample cases with experts‘ subjective similarity data. In this study, the effect of presenting reference images to observers‘ ablity to distinguish between benign and malignant masses was investigated. Eleven physicians and 11 radiological technologists evaluated 98 masses and recorded their confidence of a lesion being malignant without and with reference images. The areas under the receiver operating characteristic curves improved from 0.926 to 0.938 (p=0.17) and from 0.895 to 0.928 (p=0.004) for the physician and technologist groups, respectively. The results indicate that reference images may be useful for diagnosis of breast masses.