Statistical Methods for Breast Mass Classification by Ultrasound Imaging
C Sehgal*, L Sultan, B Levenback, S Venkatesh, Univ Pennsylvania, Philadelphia, PASU-D-134-6 Sunday 2:05PM - 3:00PM Room: 134
Purpose: The overall goal of this study is to develop a computer-based image analysis system for breast ultrasound that may aid physicians in differentiating benign and malignant masses with higher confidence and thus help in reducing unnecessary biopsies. Towards this goal we describe an approach that combines two independent probabilistic classifiers to improve diagnosis of breast masses.
Methods: B-scan images of 266 patients with biopsy proven breast masses were analyzed for margin grayscale and shape features. These features were used with two statistical methods, logistic regression and naive Bayes, to classify the lesions as malignant and benign. The diagnostic performance of the ultrasound image features was evaluated by using them alone and by combining them with mammographic BI-RADS categories and patient age. The probability predictions of the two classifiers were compared to assess consensus between them. The performances of the classifiers were evaluated using area under the curve (AUC) of the ROC Receiver Operating Characteristic. Training and testing were performed using leave-one-out validation.
Results: The results showed that combined features outperformed the individual features. Logistic regression performed slightly better than naive Bayes (AUC: 0.902 ± 0.023 vs. 0.865 ± 0.027, p < 0.03). The agreement between the two models at different decision thresholds was stable at ~88 %. The reaming 12% were treated as the cases that needed further diagnostic investigation. Treating each algorithm as an independent observer and using the consensus between the two models as the criterion for mass differentiation demonstrated a reduction in biopsy by 48% could be achieved at the cost of missed malignancies in 2% cases.
Conclusion: Computer-based image analysis of breast ultrasound images can aid the differentiation of benign and malignant masses and help in improving biopsy yields.
Funding Support, Disclosures, and Conflict of Interest: NIH grant R01 CA130946