Hui Li PhD, Presenter: Nothing to Disclose

Maryellen L. Giger PhD, Abstract Co-Author: Stockholder, Hologic, Inc Shareholder, Quantitative Insights, Inc Royalties, Hologic, Inc Royalties, General Electric Company Royalties, MEDIAN Technologies Royalties, Riverain Technologies, LLC Royalties, Mitsubishi Corporation Royalties, Toshiba Corporation Researcher, Koninklijke Philips Electronics NV Researcher, U-Systems, Inc

Li Lan, Abstract Co-Author: Nothing to Disclose

Sunny Y. Duan, Abstract Co-Author: Nothing to Disclose

Stephan Hu, Abstract Co-Author: Nothing to Disclose

Gillian Maclaine Newstead MD, Abstract Co-Author: Medical Advisory Board, Bayer AG Consultant, Three Palm Software LLC

Hiroyuki Abe MD, Abstract Co-Author: Nothing to Disclose

Michelle Lindgren MD, Abstract Co-Author: Nothing to Disclose

To develop a pre-processing method for automatically identifying mass and non-mass-like lesions on breast MRI for subsequent separate analysis in the task of distinguishing between malignant and benign lesions

Our dataset included 123 biopsy-proven lesions from 103 MRI studies acquired between January 2009 and April 2010, including 35 benign mass, 50 malignant mass, 11 benign non-mass-like and 27 malignant non-mass-like lesions. Each MRI underwent computerized 3D lesion segmentation and feature extraction to extract lesion characteristics of morphology, texture, and kinetics. Output from the system yielded the probability of the lesion being a mass (as opposed to a non-mass-like lesion) from a Bayesian artificial neural network (BANN). Classification performance was evaluated with a leave-one-case-out method using ROC analysis with area under the ROC curve as the figure of merit.

The classifier’s performance gave an AUC value of 0.93 (SE=0.02) in the task of differentiating mass lesions from non-mass-like enhancement breast lesions. The main lesion features incorporated into the classifier included Sphericity (shape), Time to Peak (kinetic), Variance (texture), Sum Entropy (texture), and Average Gray-Level (texture). Also, incorporation of the pre-processing step in our CADx algorithm yielded a statistical significant improvement in the malignant and benign classification task. AUC values of 0.88 (SE=0.03) and 0.95 (SE=0.02) were obtained in the task of distinguishing between malignant and benign lesions on the entire dataset and mass lesions only (p-value=0.04).

Automated quantitative image analyses of breast MRI can yield features sufficient for identifying lesions as mass or non-mass-like lesions prior to application of CADx algorithms.

In order to improve CADx diagnostic accuracy, computer algorithms (as do radiologists) should recognize lesions as mass or non-mass-like lesions prior to assessing likelihood of malignancy.

Li, H, Giger, M, Lan, L, Duan, S, Hu, S, Newstead, G, Abe, H, Lindgren, M, Pre-processing within a Breast MRI CADx System for the Initial Separation of Mass and Non-mass-like Lesions Prior to Computer Characterization.  Radiological Society of North America 2013 Scientific Assembly and Annual Meeting, December 1 - December 6, 2013 ,Chicago IL. http://archive.rsna.org/2013/13023083.html