Berkman Sahiner PhD, Presenter: Nothing to Disclose

Heang-Ping Chan PhD, Abstract Co-Author: Nothing to Disclose

Lubomir M. Hadjiiski PhD, Abstract Co-Author: Nothing to Disclose

Marilyn A. Roubidoux MD, Abstract Co-Author: Nothing to Disclose

Chintana P. Paramagul MD, Abstract Co-Author: Nothing to Disclose

Mark Alan Helvie MD, Abstract Co-Author: Nothing to Disclose

Caroline Elizabeth Blane MD, Abstract Co-Author: Nothing to Disclose

Alexis Virginia Nees MD, Abstract Co-Author: Nothing to Disclose

Janet E. Bailey MD, Abstract Co-Author: Nothing to Disclose

et al, Abstract Co-Author: Nothing to Disclose

Computer-aided diagnosis (CAD) methods have previously been developed to assist radiologists in characterizing breast masses on mammograms and ultrasound (US) images. In this study, we developed a classifier that merged information from both modalities, and assessed its effect on radiologists' accuracy.

The data set consisted of images from 67 patients containing biopsy-proven solid masses (32 benign and 35 malignant). An experienced radiologist identified the region of interest (ROI) containing the lesion on both modalities. The 3D US volumetric data were collected as cine-clips when the transducer was translated across the lesion. US and mammographic features were automatically extracted based on the margin, spiculation, shadowing, and shape characteristics of the masses. The features were combined into a malignancy score using a computer classifier designed with a leave-one-case-out method. Five MQSA radiologists participated in the ROC study. First, the radiologist read the mammogram ROIs, and provided a BIRADS score and a malignancy rating. Second, the US images were displayed along with the mammogram ROIs, the radiologist provided a second malignancy rating, and recommended: (i) 1-year follow-up; (ii) short-term follow-up; or (iii) biopsy. Third, the computer score was displayed, and the radiologist provided a third malignancy rating and revised the recommended action. The classification accuracy was quantified using the area under ROC curve, Az.

The computer classifier achieved a test Az value of 0.91. When reading mammograms alone, the radiologists had an average Az of 0.88 (range: 0.82-0.93). When the mammograms were supplemented by US images, the average Az increased to 0.92 (range:0.86-0.96). With CAD, the average Az increased significantly (p=0.03) to 0.95 (range:0.90-0.98). The average sensitivity for biopsy recommendation also improved from 0.96 to 0.98, and average specificity improved from 0.37 to 0.39.

The radiologists were more accurate in characterizing masses when both mammograms and volumetric US images were available. A well-trained computer algorithm can improve radiologists' accuracy even in this multi-modality reading condition.

Sahiner, B, Chan, H, Hadjiiski, L, Roubidoux, M, Paramagul, C, Helvie, M, Blane, C, Nees, A, Bailey, J, et al, , The Effect of a Multi-Modality Computer Classifier on Radiologists' Accuracy in Characterizing Breast Masses Using Mammograms and Volumetric Ultrasound Images: An ROC Study.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4412549.html