Ultrasound radio-frequency time series for finding malignant breast lesions

This project provides an insight into ultrasound-based solutions for breast lesion characterization to reduce the patient recall rate after mammography screening. In this work, ultrasound radio frequency time series analysis is performed for separating benign and malignant breast lesions with similar B-mode appearance. The radio frequency time series method is versatile and requires only a few seconds of imaging with no need for additional instrumentation. This study employs the spectral and fractal features of ultrasound radio frequency time series along with a machine learning framework with leave-one-patient-out cross validation of the classification. Support vector machines, bagged decision trees (random forest), and naive Bayes methods are used and compared in this context. For clinical relevance, cancer probability maps are also produced, by estimating the posterior malignancy probability of regions of size 1 mm in the suspicious lesions. Recorded area under the receiver operating curve is, 0.79 using SVM, 0.74 using random forest, and 0.68 using naive Bayes classification. All classifiers successfully classified 6 out of 7 patients with malignant breast lesions and 4 out of 5 patients with benign lesions, with success defined as correct classification of at least 80% of the 1 mm regions. The above findings suggests that ultrasound radio frequency time series along with the developed machine learning framework can help in differentiating malignant from benign breast lesions.