OBJECTIVE: To characterize interstitial lesions in anterior-posterior chest X-rays based on the analysis of textural statistical features that allow the detection of abnormalities with diffuse pattern. MATERIALS AND METHODS: Image analysis begins with the semiautomatic segmentation of the lungs, marking the external contour of the lung manually followed by an automatic division of each lung in six regions. The data base of images used in this study consisted of 482 regions obtained from examinations in which lesions were detected and 324 regions from normal examinations. Textural features were automatically extracted from each area and the selection of the best set of features was made based on the Jeffries-Matusita distance. The regions were classified as normal or suspected using the k nearest-neighbor method and half-half, and cross-correlation methodologies were used for training the classifier. RESULTS: Results were assessed based on the value of the area under the ROC (receiver operating characteristic) curve that indicates an ideal response for an area equal to 1. The results showed an area under the ROC curve (A Z) of 0.887, sensitivity of 0.804, and specificity of 0.793. CONCLUSION: These results indicate that the implemented system has a good potential for computer-aided diagnosis of interstitial lung lesions.

Keywords: Computer-aided diagnosis, Interstitial lung lesions, Texture characteristics, Pattern recognition