WE‐D‐304A‐04: Performance of a Method to Aid Breast Ultrasound Interpretation Using Image Processing and Case‐Based Reasoning
Michael P. André, Michael Y. Galperin, Haydee Ojeda‐Fournier, Linda K.M. Olson, Ashley Berry, Christopher Comstock, Mary K. O'Boyle · Medical Physics · 2009
Purpose: To examine factors that may impact accuracy, reproducibility of Breast Ultrasound CADx software. Goal for CADx is standardized reporting of findings using well‐defined descriptors and to aid accuracy and reproducibility of interpretation by radiologists. Method and Materials: Breast Companion® (BC, Almen Laboratories and UCSD) provides tools to analyze breast masses. It computes 9 features of the mass, compares these to images in Reference Library of masses with known findings. Using Relative Similarity it retrieves instantaneously a cluster of the most similar cases and outputs numerical data including Computerized Lesion Assessment (CLA) following the ACR BI‐RADS assessment category (1 through 5). Results: On 596 cases with known findings BC achieved ROC Area 0.98±0.02, Sensitivity 99.7%, Specificity 96.1%, significantly higher than four experienced radiologists who read the same cases (ROC Areas 0.88 to 0.90±0.02). Increasing the number of cases in the Reference Library from 41–331 and the number Test Set cases from 20–924 keeping mix same, ROC varied from 0.90±0.07 to 0.98±0.02 (NS). In 55 cases (40% malignant) with confirmed findings BC calculated CLAs for radial and anti‐radial images that were highly correlated, not significantly different by two‐tailed t‐test (p>.5). CLA for simple cysts was 2.0, mean for complicated cysts was 2.5, 3.0 for solid benign and 4.5 for malignants. 28 subjects (40% malignant) were analyzed who had exams of the mass on two different occasions within 3 weeks on GE and Siemens (NS difference). Simulated masses of were imaged over 3 months at 9, 10, 13 MHz with varying gain, dynamic range, focal depth post‐processing. CLA was stable and independent of or linearly scaled to machine parameters. Conclusion: The BC system accuracy is high, stable over range of conditions. Measured parameters are dimensionless, relative to their backgrounds so it appears normalization achieves mitigation in machine effects.