Transformer-based Characterization of Breast Lesions in Handheld Ultrasound Images WITH Classification Inconsistency Measure

Grigorios M. Karageorgos, John Karigiannis, Cynthia Davis, Soumya Ghose, Jhimli Mitra

Abstract: Ultrasound imaging combined with computer aided diagnosis methods has shown promise in improving the availability of breast cancer screening in low-and-middle- income countries. In this study, we employ a vision transformer-based architecture to discriminate between benign and malignant lesions in breast ultrasound images obtained through a handheld device. We further introduce a metric, the classification inconsistency rate (CIR), that quantifies the uncertainty of the model for a queried breast ultrasound image. The trained model provided a test accuracy and area under the ROC curve (AUC) equal to 89.3 % and 0.95, respectively. These metrics increased to 94.4% and 0.96, respectively, by excluding uncertain predictions based on the CIR values. Fisher exact test indicated that false predictions had significantly higher probability of providing CIR>0, compared to true predictions (p-value<0.00001). The trained model with the uncertainty measure, are expected to assist in breast cancer diagnosis using portable ultrasound scanners in remote settings.