Researchers at the University of Toronto have tested and validated the superiority of a two-stage cascade classifier over a traditional, single-shot classifier when using computer-aided diagnosis (CAD) to differentiate between mass and nonmass breast lesions.
In a retrospective study of 280 histologically proved mass lesions and 129 histologically proved nonmass lesions identified in MR imaging studies, their cascaded classifier decreased the overall misclassification rate by 12 percent (72 of 409) of cases missed with cascade versus 82 of 409 missed with one-shot classifier.
The authors write ahead of print in Radiology that their system holds promise for better screening of women at high risk for breast cancer and for whom early diagnosis may mean the difference between a good outcome and its opposite.
Drawing from radiology reports to categorize mass and nonmass enhancement according to Breast Imaging Reporting and Data System (BI-RADS) classifications, Cristina Gallego-Ortiz, MSc, and Anne L. Martel, PhD, extracted image data from dynamic contrast-enhanced MR.
They analyzed the data using feature selection techniques and binary, multiclass and cascade classifiers.
The classifier that performed the best was a two-stage cascade classifier, which had median areas under the receiver operating characteristics curve (AUCs) of 0.91. By comparison, the AUC for one-shot classifiers was 0.89 (P = .0027).
The cascade classifier was made up of three separately trained binary classifiers, the authors explain. For the cascade evaluation, cases fed to second-stage classifiers—one for masses and one for nonmasses—were based on the prediction of the first stage, which served as the classifier for differentiating between masses and nonmasses.
Cascading classifiers “offer the advantage to further exploit differences among lesion groups in future studies, such as allowing the design of new features intended only for mass and nonmass lesions,” they write. “[T]his study demonstrates that CAD shows superior performance for diagnosing malignant and benign lesions when discriminating among mass and nonmass lesions with a cascade arrangement of classifiers.”