Impact of CT slice thickness reduction algorithm on AI-based lung nodule detection in chest CTs of colorectal cancer patients
Quantitative Imaging in Medicine and Surgery, 2026
Abstract
Introduction: Radiology departments running AI-assisted lung nodule detection face a persistent challenge: imaging protocols vary widely across sites, and thick CT slices routinely used in oncology surveillance degrade AI performance. This study examined whether deep learning-based super-resolution (SR) reconstruction could recover enough structural detail from thick-section chest CTs to meaningfully improve AI detection rates — a question with direct implications for departments seeking to deploy or standardise AI tools across heterogeneous scanner environments.
Methods: In this retrospective multicentre study, researchers analysed chest CT scans from 96 colorectal cancer patients undergoing routine metastatic surveillance. The focus was pulmonary nodules smaller than 10 mm, with up to three nodules per patient assessed. Three image sets were fed into a commercial AI computer-aided detection system (AVIEW LCS+, Coreline Soft): the original thin-section images (used as the reference standard), original 5 mm thick-section images, and SR-converted thin-section images derived from the thick-section data. Detection performance was compared at both nodule and patient level, and image quality metrics including noise levels and nodule morphology were assessed.
Results: Among 105 reference nodules, the AI system's sensitivity rose sharply from 31.4% on thick-section images to 61.0% on SR-converted images, while positive predictive value improved from 42.9% to 80.0%. Patient-level sensitivity climbed from 41.5% to 67.1%. SR reconstruction significantly reduced image noise and preserved nodule anatomy in 95.3% of cases, with 80% solid-feature consistency. Nodule dimensions remained comparable across image types, though SR-converted images showed slightly lower density values. A notable caveat: SR reconstruction produced hallucinated nodules in approximately 20% of cases — 16 out of 80 SR-derived findings — most of which represented benign or artefactual structures rather than true malignancies.
Conclusion: SR reconstruction offers a practical route to improving AI-based nodule detection in thick-slice CT datasets without requiring re-scanning — potentially valuable for multisite AI research and retrospective data harmonisation. Radiology directors should note that while detection gains are substantial, the hallucination rate warrants human oversight and further prospective validation before clinical deployment at scale.
