Karen Drukker PhD, Presenter: Nothing to Disclose

Maryellen Lissak Giger PhD, Abstract Co-Author: Nothing to Disclose

To evaluate performance robustness of a computer-aided detection and diagnosis method for breast ultrasound images collected with ultrasound scanners from two different manufacturers.

The computerized scheme first detects potential lesions based on expected lesion shape and margin characteristics. Subsequently, image features of all candidate lesions are calculated and employed to classify the candidates into different categories. Two separate classification tasks were performed and evaluated at the classification stage: The first classification task was the distinction between all actual lesions and false-positive detections, and the second classification task was the distinction between cancererous lesions and all other detected lesion candidates (including false-positive detections). The computerized method was trained on a database of 458 cases (1740 images including malignant, benign solid, cystic, and normal pathologies) collected with ATL 3000 equipment. The computerized method was then applied, without any adjustment of processing parameters, to images of an independent database of 151 cases (a single image per case) collected with Accuson equipment.

In the distinction between all actual lesions and false-positive detections, Az values of 0.95 and 0.87 were obtained with the training and testing data sets, respectively. A sensitivity by image of 80% was achieved at 0.58 and 0.28 false-positives per image for training and testing, respectively. In the distinction of cancer from all other detections (false-positives plus all benign lesions) the Az values for training and testing sets were identical at 0.84. A sensitivity by image of 80% was achieved at 0.50 false-positive cancers per image for the training set, and at 0.93 false-positive cancers per image for the testing data set.

The results demonstrate promising performance of our fully automated computerized lesion detection and classification method, and robustness with respect to the different ultrasound equipment used in this study.

M.L.G.: Shareholder: R2 Technology Inc, Sunnyvale, CA.K.D.: Work supported in parts by: USPHS grants CA89452 & T32 CA09649 and the US Army Medical Research & Materiel Command DAMD 97-2445. Thanks to Michael Stern & Siemens Medical Solutions USA, Inc., Ultrasound Div for their help in collectig the sonographic databases.

Drukker, K, Giger, M, Robustness of a Computerized Lesion Detection and Classification Scheme for Breast Ultrasound across Acquisition Platforms.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4406116.html