Monica M. S. Matsumoto PhD, Presenter: Nothing to Disclose

Chandra Sehgal PhD, Abstract Co-Author: Nothing to Disclose

Jayaram K. Udupa PhD, Abstract Co-Author: Nothing to Disclose

Ultrasound imaging can assist in the characterization of solid breast lesions. Further, pattern recognition techniques can facilitate the classification of complex lesions as malignant or benign, and help in reducing the number of unnecessary biopsies. Image features, such as sharpness of margins, texture, and shape of the lesion can be utilized in computer-aided diagnosis. The hypothesis of this work was that such features measured over the boundary and in the vicinity of the lesion can differentiate benign from malignant lesions.

The analysis of texture features over the boundary of solid lesions may hold promise for malignancy classification in ultrasound breast images.

The features related to texture have achieved an accuracy rate of 81.0%, with K=13 neighbors and width= ±5 pixels. The results show that the CAD method can collect information relevant to classify a lesion from an annular region that includes the boundary. Further, the analysis of the malignancy score provided an area under the ROC curve of 0.803.

In this study we had breast ultrasound images of 100 patients, 50 benign and 50 malignant. The evaluation of whether the lesion was benign or malignant was done through biopsy and pathology assessment. The benign cases included cysts and complex cysts (7), fibroadenomas (33), and other benign reports (10). The malignant cases included ductal carcinoma in situ (9), infiltrating ductal carcinoma (36), and other malignant reports (5). The boundaries of the masses were hand-drawn by an ultrasound specialist. The features included texture (local binary pattern histograms) and standardized image intensity information. The classifier was based on K-Nearest Neighbor (KNN) and the classifier output was a score of malignancy. Experiments were done with multi-fold sets, by randomly designating 60 patient data sets as a training set and the remaining 40 for testing. Different widths of the region of interest over the boundary were tested, as well as the number of K neighbors.

Matsumoto, M, Sehgal, C, Udupa, J, Breast Ultrasound Lesion Classification Based on Boundary Features.  Radiological Society of North America 2012 Scientific Assembly and Annual Meeting, November 25 - November 30, 2012 ,Chicago IL. http://archive.rsna.org/2012/12025029.html