Partial dependence of breast tumor malignancy on ultrasound image features derived from boosted trees

摘要

Various computerized features extracted from breast ul-ntrasound images are useful in assessing the malignancy of breastntumors. However, the underlying relationship between the comput-nerized features and tumor malignancy may not be linear in nature.nWe use the decision tree ensemble trained by the cost-sensitivenboosting algorithm to approximate the target function for malignancynassessment and to reflect this relationship qualitatively. Partial de-npendence plots are employed to explore and visualize the effect ofnfeatures on the output of the decision tree ensemble. In the experi-nments, 31 image features are extracted to quantify the sonographicncharacteristics of breast tumors. Patient age is used as an external feature because of its high clinical importance. The area under thenreceiver-operating characteristic curve of the tree ensembles cannreach 0.95 with sensitivity of 0.95 u000161/64u0002 at the associated speci-nficity 0.74 u000177/104u0002. The partial dependence plots of the four mostnimportant features are demonstrated to show the influence of thenfeatures on malignancy, and they are in accord with the empiricalnobservations. The results can provide visual and qualitative refer-nences on the computerized image features for physicians, and cannbe useful for enhancing the interpretability of computer-aided diag-nnosis systems for breast ultrasound.