Computer recognition of regional lung disease patterns

Abstract

We have developed an objective, reproducible, and automated means for the regional evaluation of the pulmonary parenchyma from computed tomography (CT) scans. This method, known as the Adaptive Multiple Feature Method (AMFM) assesses as many as 22 independent texture features in order to classify a tissue pattern. In this study, the six tissue patterns characterized were: honeycombing, ground glass, bronchovascular, nodular, emphysemalike, and normal. The lung slices were evaluated regionally using 31 x 31 pixel regions of interest. In each region of interest, an optimal subset of texture features was evaluated to determine which of the six patterns the region could be characterized as. The computer output was validated against experienced observers in three settings. In the first two readings, when the observers were blinded to the primary diagnosis of the subject, the average computer versus observer agreement was 44.4 +/- 8.7% and 47.3 +/- 9.0%, respectively. The average interobserver agreement for the same two readings were 48.8 +/- 9.1% and 52.2 +/- 10.0%, respectively. In the third reading, when the observers were provided the primary diagnosis, the average computer versus observer agreement was 51.7 +/- 2.9% where as the average interobserver agreement was 53.9 +/- 6.2%. The kappa statistic of agreement between the regions, for which the majority of the observers agreed on a pattern type, versus the computer was found to be 0.62. For regional tissue characterization, the AMFM is 100% reproducible and performs as well as experienced human observers who have been told the patient diagnosis.