A multi-scale tumor detection algorithm in whole breast sonography incorporating breast anatomy and tissue morphological information

A computerized whole breast ultrasound tumor detection method is proposed in this study. This algorithm can assist physicians to locate potential tumors when reading numerous 2D slices of ultrasound images. The proposed algorithm is composed of five steps. First, the modified fuzzy C-means bias correction method is applied to facilitate the second step of blob-like structrue detection. At the third step, the anomotically-impossible blob structures are eliminated with the aid of plate-like structure identification in the sonographic volume. At the fourth step, we conduct feature selection to find out useful features to profile the breast lesions. With the five selected useful features, at the last step, we adopt the logistic regression classifier to identify the real lesions from the pool of blob structure candidates. The proposed method is tested on 45 volume images acquired from 27 patients with totally 86 lesions. The area under the receiver operating characteristic curve (ROC) is used as the assessment metric to compare our method with the state-of-the-art detection method. The experimental results suggest that the proposed method has better detection performance for the 45 volume images.