Retrospective Validation of AccuPulmo CT Portal for Detecting Pulmonary Fibrosis on Chest CT

Overview

This retrospective observational study evaluates the diagnostic performance of AccuPulmo CT Portal, an artificial intelligence-assisted medical imaging software, for detecting pulmonary fibrosis on pre-existing chest computed tomography images. A total of 900 chest computed tomography examinations obtained at Taichung Veterans General Hospital between January 1, 2020, and December 31, 2024, will be retrospectively selected. The planned sample includes 300 examinations with pulmonary fibrosis and 600 examinations without pulmonary fibrosis. All study images will be de-identified and coded before evaluation. Three qualified specialists in pulmonology or radiology will independently review each image without access to the original radiology report or the artificial intelligence output. The reference standard will be established by majority agreement of at least two of the three specialists. AccuPulmo CT Portal will retrospectively analyze the coded images. An artificial intelligence-derived pulmonary fibrosis area greater than 10 percent will be classified as positive, and an area of 10 percent or less will be classified as negative. The primary performance measures are sensitivity and specificity. Secondary measures include accuracy, positive predictive value, negative predictive value, and performance across clinically relevant subgroups. The software results will not be returned to treating physicians and will not affect participant diagnosis, treatment, or clinical management.

This is a single-center, retrospective, non-interventional diagnostic performance study using pre-existing chest computed tomography images and associated clinical information from Taichung Veterans General Hospital. Eligible participants are adults aged 20 years or older who underwent chest computed tomography for pulmonary disease between January 1, 2020, and December 31, 2024. The study plans to include 900 chest computed tomography examinations, comprising approximately 300 pulmonary fibrosis-positive examinations and 600 pulmonary fibrosis-negative examinations. Potentially eligible examinations will initially be identified from existing institutional radiology records. Images with missing data or image characteristics that substantially interfere with lung texture assessment, including cardiac implants, extensive pneumonia, or pleural effusion, will be excluded according to the prespecified eligibility criteria. Before specialist review, positive and negative samples will be combined, de-identified, assigned blinded study codes, and randomly allocated for review. Three qualified specialists holding board certification in pulmonology or radiology will independently evaluate all study images. The specialists will not have access to participant identifiers, the original radiology reports, or the AccuPulmo CT Portal results. Each specialist will estimate the proportion of pulmonary fibrosis and classify each examination as pulmonary fibrosis positive or negative. The reference standard will be determined by majority agreement, defined as concordant classification by at least two of the three specialists. All coded images will subsequently be analyzed by AccuPulmo CT Portal. According to the prespecified classification rule, a pulmonary fibrosis area greater than 10 percent will be classified as positive, designated as Critical Risk, whereas a pulmonary fibrosis area of 10 percent or less will be classified as negative, designated as Low Risk. After completion of the blinded assessments, the database will be unblinded for statistical analysis. The AccuPulmo CT Portal classifications will be compared with the specialist-derived reference standard. The primary endpoints are sensitivity and specificity. Secondary endpoints are accuracy, positive predictive value, negative predictive value, and performance consistency across relevant clinical subgroups. The study uses only pre-existing images and records. AccuPulmo CT Portal will be operated in an offline research environment, and its outputs will not be returned to treating physicians or used for clinical decision-making. No additional imaging examination, clinical procedure, treatment assignment, or participant contact will occur as part of this study.