Various computerized features extracted from breast ultrasound images are useful in assessing the malignancy of breast tumors. However, the underlying relationship between the computerized features and tumor malignancy may not be linear in nature. We use the decision tree ensemble trained by the cost-sensitive boosting algorithm to approximate the target function for malignancy assessment and to reflect this relationship qualitatively. Partial dependence plots are employed to explore and visualize the effect of features on the output of the decision tree ensemble. In the experiments, 31 image features are extracted to quantify the sonographic characteristics of breast tumors. Patient age is used as an external feature because of its high clinical importance. The area under the receiver-operating characteristic curve of the tree ensembles can reach 0.95 with sensitivity of 0.95 (61/64) at the associated specificity 0.74 (77/104). The partial dependence plots of the four most important features are demonstrated to show the influence of the features on malignancy, and they are in accord with the empirical observations. The results can provide visual and qualitative references on the computerized image features for physicians, and can be useful for enhancing the interpretability of computer-aided diagnosis systems for breast ultrasound.
The shapes of malignant breast tumors are more complex than the benign lesions due to their nature of infiltration into
surrounding tissues. We investigated the efficacy of shape features and presented a method using polygon shape
complexity to improve the discrimination of benign and malignant breast lesions on ultrasound. First, 63 lesions (32
benign and 31 malignant) were segmented by K-way normalized cut with the priori rules on the ultrasound images.
Then, the shape measures were computed from the automatically extracted lesion contours. A polygon shape complexity
measure (SCM) was introduced to characterize the complexity of breast lesion contour, which was calculated from the
polygonal model of lesion contour. Three new statistical parameters were derived from the local integral invariant
signatures to quantify the local property of the lesion contour. Receiver operating characteristic (ROC) analysis was
carried on to evaluate the performance of each individual shape feature. SCM outperformed the other shape measures,
the area under ROC curve (AUC) of SCM was 0.91, and the sensitivity of SCM could reach 0.97 with the specificity
0.66. The measures of shape feature and margin feature were combined in a linear discriminant classifier. The
resubstitution and leave-one-out AUC of the linear discriminant classifier were 0.94 and 0.92, respectively. The
distinguishing ability of SCM showed that it could be a useful index for the clinical diagnosis and computer-aided
diagnosis to reduce the number of unnecessary biopsies.