Few-shot decision tree for diagnosis of ultrasound breast tumor using BI-RADS features

This paper proposes an ultrasound breast tumor CAD system based on BI-RADS features scoring and decision tree algorithm. Because of the difficulty of biopsy label collection, the proposed system adopts a few-shot learning method. The SVM classifier is employed to preliminarily mark the unlabeled cases firstly. Then these unlabeled cases with the pseudo labels are combined with the few real-labeled cases to train the decision tree. To test the performance of the proposed method, 1208 ultrasound breast images were collected, and three well-experienced clinicians and three interns evaluated these images according to the BI-RADS scoring scheme. All of the images are transformed into vectors such that the algorithm can process. The experimental results show that the system performance improves significantly with the help of pseudo-labeled data. Compared to the decision tree trained by the real-labeled cases only, when the number of real-labeled cases was 40, the accuracy, specificity, sensitivity of the proposed system were increased by 2.05%, 2.47% and 1.81%, respectively; the positive predictive value (PPV) and the negative predictive value (NVP) were increased by 1.29% and 3.05%, respectively. Meanwhile, the performance of the proposed method was the same as the method using sufficient samples. When the number of the labeled cases reached 100, the accuracy, specificity, sensitivity, PPV and NVP of the proposed method were 90.03%, 87.02%, 91.68%, 93.07%, and 85.03%, respectively. The results demonstrate that our method can efficiently distinguish the breast tumor although the labeled data is not sufficient.