RATIONALE AND OBJECTIVES: The authors developed a texture-based pattern recognition and segmentation tool for the quantitation of high-resolution computed tomography (HRCT) findings in usual interstitial pneumonia (UIP). METHODS: In HRCT images of five patients with UIP and five patients without UIP, 1022 regions of interest (ROIs) of 5 x 5 pixels were classified by the examiner to be normal, emphysematous, ground-glass lesion, intralobular fibrosis, vessel, or bronchus section. The classes and the texture parameters calculated in the ROIs were the basis for the decision rule, using a multivariate discrimination analysis. The classification was compared with the examiner's diagnosis in 1889 new randomly selected ROIs. RESULTS: Depending on the structure, the sensitivity (the probability that a structure would be recognized correctly) was 68.7% to 80.7%. If the system classified a structure as normal, ground glass or fibrotic region, this was correct in 77.3% to 88.1%. However, the system's diagnosis of a bronchus section was correct in only 16.2%. The overall accuracy was 70.7%. CONCLUSIONS: Texture-based segmentation may be a valuable tool to aid the quantitative assessment of parenchymal disease in HRCT images.
RATIONALE AND OBJECTIVES: The authors developed a texture-based pattern recognition and segmentation tool for the quantitation of high-resolution computed tomography (HRCT) findings in usual interstitial pneumonia (UIP). METHODS: In HRCT images of five patients with UIP and five patients without UIP, 1022 regions of interest (ROIs) of 5 x 5 pixels were classified by the examiner to be normal, emphysematous, ground-glass lesion, intralobular fibrosis, vessel, or bronchus section. The classes and the texture parameters calculated in the ROIs were the basis for the decision rule, using a multivariate discrimination analysis. The classification was compared with the examiner's diagnosis in 1889 new randomly selected ROIs. RESULTS: Depending on the structure, the sensitivity (the probability that a structure would be recognized correctly) was 68.7% to 80.7%. If the system classified a structure as normal, ground glass or fibrotic region, this was correct in 77.3% to 88.1%. However, the system's diagnosis of a bronchus section was correct in only 16.2%. The overall accuracy was 70.7%. CONCLUSIONS: Texture-based segmentation may be a valuable tool to aid the quantitative assessment of parenchymal disease in HRCT images.