Se Ri Kang, Won Gi Jeong, Ji Young Rho, Hyemi Choi, Yun-Hyeon Kim, Ju-Hyung Lee, Gong Yong Jin
This retrospective multicenter study evaluated whether deep learning-based super-resolution (SR) reconstruction can enhance structural conspicuity in thick-section chest computed tomography (CT) and improve the detection performance of an artificial intelligence (AI)-based computer-aided detection (CAD) system for lung nodules in 96 patients with colorectal cancer (CRC) undergoing chest CT for metastatic surveillance. Pulmonary nodules <10 mm and ≤3 per patient were analyzed. Three image sets were evaluated using a commercial AI-CAD system: original thin-section images (reference), original 5-mm thick-section images, and SR-converted thin-section images. Nodule- and patient-level detection performances were compared using the Cochran-Mantel-Haenszel test and aligned rank transform analysis of variance (ART-ANOVA). Quantitative nodule metrics, image noise, and morphologic consistency were assessed between original and SR-converted thin-section images. Among 105 reference nodules, AI sensitivity increased from 31.4% with thick-section images to 61.0% with SR-converted images, and positive predictive value (PPV) increased from 42.9% to 80.0%. Patient-level sensitivity improved from 41.5% to 67.1%. SR reconstruction reduced image noise (P<0.001) and preserved nodule morphology, with 95.3% anatomic concordance and 80% solid-feature consistency. Nodule size remained comparable; however, density was lower in converted images. SR reconstruction generated 20% hallucinated nodules (16/80), predominantly benign or artifactual structures. In conclusion, SR reconstruction enhances AI-based pulmonary nodule detection by compensating for structural detail lost in thick-section CT. Despite hallucinated nodules, SR reconstruction may provide a feasible harmonization strategy for retrospective multicenter AI research using heterogeneous CT datasets, although further validation in larger populations is required.