Breast ultrasonography: diagnostic efficacy of a computer-aided diagnostic system using fuzzy inference

We have developed a computer-aided diagnostic (CAD) system using fuzzy inference for breast sonography and have evaluated the performance of the system. Our CAD system is not an automated image processing method, but requires the observer's subjective "scoring." Seven radiologists interpreted 54 breast mass lesions (24 malignant, 30 benign). Six criteria (shape, border, halo, internal echoes, posterior echoes, and edge shadows) were scored using a five-point rating scale. The output was described as a real number from 0.0 to 1.0. For cancer diagnosis, the sensitivity of the radiologists, a six-criteria CAD version, and a four-criteria version (excluding posterior echoes and edge shadows) were 63.1%, 82.1%, and 78.0%, respectively. Specificity was 71.0%, 42.9%, and 51.0%, respectively, while accuracy was 67.5%, 60.3%, and 63.0%, respectively. No significant differences in the areas under ROC curve (Az) were found between the observers and CAD system. Several cases assessed as false-negative by observers were correctly diagnosed by the CAD system, and diagnostic efficacy could be improved, especially among novice sonographers. Although the system's high false-positive rate needs to be reduced, it will assist radiologists in generating a level of suspicion for malignancy.