Junji Shiraishi PhD, Presenter: Nothing to Disclose

Kenji Suzuki PhD, Abstract Co-Author: Nothing to Disclose

Qiang Li PhD, Abstract Co-Author: Nothing to Disclose

Roger Engelmann MS, Abstract Co-Author: Nothing to Disclose

Shigehiko Katsuragawa PhD, Abstract Co-Author: Nothing to Disclose

Kunio Doi PhD, Abstract Co-Author: Nothing to Disclose

We have been developing computerized schemes for lung nodule detection on chest radiographs for the last two decades. In this study, the performance of an advanced computerized scheme was validated by use of a Jackknife test method.

One thousand digitized chest radiographs (500 x 500 matrix size) including 1077 nodules were collected from ten institutions. All images were divided randomly into two sets which were used for training (500 cases) and testing (500 cases). In this computerized scheme, a lung field was segmented with a ribcage detection technique first, and then divided into a 7 x 7 grid of ROIs (each with a matrix size of 64 x 64 pixels). Each ROI was classified into apical, peripheral, hilar, and diaphragm/heart anatomical regions by use of their image features. Initial candidates were identified by use of a multiple difference image technique in which the filter size and shape were varied adaptively according to the location and the anatomical classification of the ROI. Ninety image features extracted from the original, difference and contralateral subtraction images were employed for the rule-based scheme and linear discriminant analysis (LDA) to remove false positive candidates. Template matching technique was applied to remove the remaining false positives. Settings and parameters in the rule-based test, LDA and template matching technique were determined only by use of training cases so that test cases were examined independently. Jackknife tests ware performed with several combinations of training and testing image data sets in the same way.

Preliminary results obtained from 500 testing cases indicated that the average sensitivity in detecting lung nodules was 70.1 % with 5.8 false positives per image, whereas the sensitivity obtained from 500 training cases was 75.0 % with 4.9 false positives per image.

The advanced CAD scheme involving anatomically adaptive difference image filtering, contralateral subtraction, and template matching techniques provided an improved, robust detection of pulmonary nodules.

K.D.: Kunio Doi is a shareholder in Deus Technologies, Inc., Rockville, MD.K.D.,S.K.: Shigehiko Katsuragawa and Kunio Doi are shareholders in R2 Technology, Inc., Sunnyvale, CA.

Shiraishi, J, Suzuki, K, Li, Q, Engelmann, R, Katsuragawa, S, Doi, K, Computer-aided Detection of Lung Nodules on Chest Radiographs: Evaluation with a Large Scale Image Database.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4404308.html