Rie Tachibana MS, Presenter: Nothing to Disclose

Shoji Kido MD, Abstract Co-Author: Nothing to Disclose

Nobuyuki Tanaka MD, Abstract Co-Author: Nothing to Disclose

Tsuneo Matsumoto MD, Abstract Co-Author: Nothing to Disclose

The diagnosis of idiopathic interstitial pneumonias (IIPs) is a difficult problem for radiologists because they have a variety of patterns. The purpose in this study was to develop the computerized classification for the image features on high-resolution CT (HRCT) images of three high-incidence groups of IIPs and normals by use of power spectrum analysis.

We have developed a computerized classification system for IIPs with 57 HRCT images (13 idiopathic pulmonary fibrosis (IPF) / usual interstitial pneumonia (UIP), 19 nonspecific interstitial pneumonia (NSIP), 10 cryptogenic organizing pneumonia (COP) / bronchiolitis obliterans organizing pneumonia (BOOP), and 15 normals). The acquisition parameters of the HRCT images were: 512 x 512 pixels, 0.352 mm pixel size, and 2 mm slice thickness. In our system, we first selected regions of interest (ROIs) with 31 x 31 pixels on HRCT images. Next, the Fourier transform was performed on an ROI image for calculation of the power spectrum. Then, three texture features (the route mean square, the first moment, and the average of the wedge sampling geometry) were calculated from the power spectrum. Finally, we used Mahalanobis distance in analyzing each ROI image for discriminant analysis. A leave-one-out method was used for evaluation of the performance of our system.

In the experiment using 57 ROI images, a correct classification rate of 84.2 % (48/57) was achieved. In the COP/BOOP group, the classification rate was 70.0% (7/10), but 3 cases were mistaken for NSIP. In the NSIP group, the classification rate was 84.2% (16/19), but one case was mistaken for BOOP, and 2 cases were mistaken for IPF/UIP. In the IPF/UIP group, the classification rate was 84.6% (11/13), but 2 cases were mistaken for NSIP. In the normal group, the classification rate was 93.3% (14/15), but one case was mistaken for NSIP.

Our computerized classification system of IIPs achieved a high correct classification rate for three IIPs groups (COP/BOOP, NSIP, and IPF/UIP) and normals. This system would be useful in assisting radiologists in the diagnosis of IIPs.

Tachibana, R, Kido, S, Tanaka, N, Matsumoto, T, Computerized Classification of Idiopathic Interstitial Pneumonia on High-Resolution CT Images Using Power Spectrum Analysis.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4414009.html