Hui-Hua Wen PhD, Presenter: Nothing to Disclose

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

Gillian Maclaine Newstead MD, Abstract Co-Author: Nothing to Disclose

Karla J. Horsch PhD, Abstract Co-Author: Nothing to Disclose

Yading Yuan, Abstract Co-Author: Nothing to Disclose

Charles Edgar Metz PhD, Abstract Co-Author: Nothing to Disclose

Li Lan, Abstract Co-Author: Nothing to Disclose

Karen Drukker PhD, Abstract Co-Author: Nothing to Disclose

We investigated methods with which to correlate lesions seen on multiple breast images of the same patient using computer-extracted lesion features and an artificial neural network. Such automated lesion matching is expected to aid radiologists in comparing and fusing information from multiple images for improving diagnostic accuracy and patient outcome.

We developed a automated lesion matching scheme using the correlation of computer extracted features between lesions seen on two different images. In our method, lesion features are first extracted and then combined to facilitate the automated matching of lesions observed on multiple breast images. Initially, regression is performed using either linear regression or an artificial neural network in order to model pairs of images of the same actual lesion. For a given pair of images, we can classify whether this pair corresponds to the same actual lesion by performing discrimination according to the established model. By computing the ROC curve from a training database and determining an optimal threshold value of the classifier, the likelihood that a pair of images is of the same actual lesion is calculated. In our preliminary study, we obtained 262 biopsy-proven lesions of breast sonography to constitute a corresponding pair training database. In addition, 53 non-corresponding lesion pairs were generated from 35 patients who had two or more lesions in the same breast, and these comprised a non-corresponding training database. An independent testing database included 89 corresponding lesion pairs and 50 non-corresponding lesion pairs.

The correlation coefficients for various combinations of features of the same lesion from two views showed substantial variations, ranging from 0.023 to 0.825. We identified that the posterior acoustic behavior outperformed the other feature pairs, yielding an AUC value of 0.83.

Our study indicates that the correlation between computer-extracted lesion feature(s) between sonographic views of a breast lesion, may be useful for matching images of the same actual lesion. It is expected that the proposed method can also be applied in two views mammograms and multimodality breast imaging.

G.M.N.: shareholders in R2 Technology, Inc.M.L.G.: Stockholder of R2 Technology Inc., & grant support R2 Technology.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.

Wen, H, Giger, M, Newstead, G, Horsch, K, Yuan, Y, Metz, C, Lan, L, Drukker, K, Correlative Matching of Lesions from Multiple Breast Images Using an Artificial Neural Network.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4406820.html