Masayuki Kuwahara BS, Abstract Co-Author: Nothing to Disclose

Hayaru Shouno PhD, Abstract Co-Author: Nothing to Disclose

Shoji Kido MD, PhD, Presenter: Nothing to Disclose

In a diagnosis of idiopathic interstitial pneumonias (IIPs), interpretation of features on high-resolution computed tomography (HRCT) image is effective for classification of those diseases. However, image patterns of IIPs on HRCT images have so much variety, which causes difficulty of classification. The purpose of our study is to develop a diagnosis support system for classification of those HRCT images using a boosting method.

We developed a computerized classification system for IIP images, and validated the system with 48 HRCT images (10: cryptogenic organizing pneumonia (COP), 18 nonspecific interstitial pneumonia (NSIP), 13: usual interstitial pneumonia (UIP), and 7: normal). The acquisition parameters of those HRCT images were: 512 x 512 pixels, 0.352 mm pixel size, and 2 mm slice thickness. In our system, we assumed an input image was selected as region of interest (ROI) with 32 x 32 pixels on a HRCT image. The input image was decomposed by several Gabor filters in the pre-processing. The boosting method was a kind of meta-learning algorithm. It required plural classifier and integrated each classifier's result based on the confidences of the classifier. We applied "Ada-Boost" method for the classifying problem, and adopted multi-layered perceptron (MLP) as each classifier. In the boosting method, the Gabor filtered images were provided to a MLP, and the MLP was trained by back-propagation (BP). The confidence rate of the MLP was calculated, and misclassified patterns were focused. Then, the next MLP was trained by BP with the focused patterns selectively.

In the experiment using 48 ROI images, a classification error rate of 4.2 % (2/48) was obtained. In the COP group, UIP group, and normal group, whole examples were classified with correctness. In the NSIP group, however, 2 cases were mistaken, one was classified as COP group, and the other was classified as UIP group. So that the error rate of the NSIP group was 11% (2/18).

Our computerized classification system for IIPs achieved low error classification rate for three IIP groups (COP, NSIP, and UIP) and normals. This system would be useful for assisting radiologists in the diagnosis of IIPs.

In a diagnosis of idiopathic interstitial pneumonias (IIPs), interpretation of features on high-resolution computed tomography (HRCT) image is effective for classification of those diseases. However, image patterns of IIPs on HRCT images have so much variety, which causes difficulty of classification. The purpose of our study is to develop a diagnosis support system for classification of those HRCT images using a boosting method.

We developed a computerized classification system for IIP images, and validated the system with 48 HRCT images (10: cryptogenic organizing pneumonia (COP), 18 nonspecific interstitial pneumonia (NSIP), 13: usual interstitial pneumonia (UIP), and 7: normal). The acquisition parameters of those HRCT images were: 512 x 512 pixels, 0.352 mm pixel size, and 2 mm slice thickness. In our system, we assumed an input image was selected as region of interest (ROI) with 32 x 32 pixels on a HRCT image. The input image was decomposed by several Gabor filters in the pre-processing. The boosting method was a kind of meta-learning algorithm. It required plural classifier and integrated each classifier's result based on the confidences of the classifier. We applied "Ada-Boost" method for the classifying problem, and adopted multi-layered perceptron (MLP) as each classifier. In the boosting method, the Gabor filtered images were provided to a MLP, and the MLP was trained by back-propagation (BP). The confidence rate of the MLP was calculated, and misclassified patterns were focused. Then, the next MLP was trained by BP with the focused patterns selectively.

In the experiment using 48 ROI images, a classification error rate of 4.2 % (2/48) was obtained. In the COP group, UIP group, and normal group, whole examples were classified with correctness. In the NSIP group, however, 2 cases were mistaken, one was classified as COP group, and the other was classified as UIP group. So that the error rate of the NSIP group was 11% (2/18).

Our computerized classification system for IIPs achieved low error classification rate for three IIP groups (COP, NSIP, and UIP) and normals. This system would be useful for assisting radiologists in the diagnosis of IIPs.

Kuwahara, M, Shouno, H, Kido, S, Classification of Idiopathic Interstitial Pneumonia Using Boosting Method on High-Resolution Computed Tomography Images.  Radiological Society of North America 2005 Scientific Assembly and Annual Meeting, November 27 - December 2, 2005 ,Chicago IL. http://archive.rsna.org/2005/4414154.html