Impact of CT slice thickness reduction algorithm on AI-based lung nodule detection in chest CTs of colorectal cancer patients

Introduction

Pulmonary nodule detection is essential for accurate staging and postoperative surveillance in patients with colorectal cancer (CRC), who have a high risk of developing pulmonary metastases throughout the disease course (1-3). Chest computed tomography (CT) is the cornerstone imaging modality for evaluating pulmonary metastases in this population, and early identification of metastatic nodules is associated with improved clinical outcomes. Artificial intelligence (AI)-based computer-aided detection (CAD) systems have recently demonstrated promising results in enhancing detection performance across thoracic oncology applications, including CRC (4-8).

Despite these advances, considerable heterogeneity in CT acquisition parameters—particularly differences in scanner models and slice thickness—continues to limit the generalizability and reproducibility of AI algorithms across institutions. Thick-slice CT data (>4 mm), still commonly encountered in clinical practice, degrade spatial resolution and hinder both the detection and characterization of small nodules (9,10). These limitations pose significant challenges when retrospectively applying AI-based detection algorithms to multicenter chest CT data collected under non-uniform protocols. Therefore, standardizing slice thickness by converting thick-slice CT data into thin-slice-equivalent images is critical to ensure reliable and reproducible AI-based analyses.

Super-resolution (SR) reconstruction has been proposed as a promising deep learning-based approach to address these limitations. By retrospectively transforming thick-slice CT images into thin-slice-equivalent datasets, SR reconstruction enhances spatial resolution and harmonizes image characteristics across heterogeneous scanners and protocols (11,12). This technique has demonstrated improvements in image quality, radiomics reproducibility, and the detectability of subsolid nodules in lung imaging. However, its specific impact on AI-driven lung nodule detection—particularly in patients with CRC—has not been systematically evaluated.

Therefore, this multicenter retrospective study aimed to evaluate whether a deep learning-based SR reconstruction could effectively standardize heterogeneous thick-slice chest CT data across different institutions and enhance AI-driven pulmonary nodule detection in patients with CRC. Specifically, we assessed whether SR-reconstructed images could achieve detection performance comparable to original thin-slice CT and evaluated the feasibility of SR reconstruction as a harmonization strategy for retrospective multicenter AI studies in oncologic chest imaging.

Methods

Study design

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Boards of Wonkwang University Hospital (No. WKUH 2024-01-023-001), Chonnam National University Hwasun Hospital (No. CNUHH-2024-044), and Chonnam National University Hospital (No. CNUH-2024-07-029-009), and individual consent for this retrospective analysis was waived. Consecutive patients with histologically confirmed CRC who underwent chest CT between February 1 and March 31, 2024, at three tertiary academic medical centers were retrospectively identified. Inclusion criteria were as follows: (I) diagnostic-quality chest CT without prior thoracic surgery or major parenchymal abnormalities (e.g., tuberculosis-related lung destruction, pulmonary fibrosis, atelectasis, consolidation, or diffuse ground-glass opacities); and (II) ≤3 noncalcified solid pulmonary nodules, each <10 mm in maximal diameter, as reported by the interpreting radiologist.

The restriction to ≤3 nodules was applied to avoid confounding in cases with diffuse metastatic burden, which could obscure lesion-level matching and bias the evaluation of CAD detection performance across different image reconstructions. In addition, the analysis was restricted to nodules <10 mm because small pulmonary nodules are most susceptible to spatial resolution loss in thick-slice CT, whereas larger nodules are generally detectable regardless of reconstruction thickness and may introduce a ceiling effect when comparing detection performance across reconstruction methods.

After AI-assisted CAD application, 29 patients were excluded due to exceeding the nodule number limit or failing to meet nodule eligibility criteria. The final cohort consisted of 96 patients (Figure 1).

Figure 1 Flowchart depicting case selection and image analysis process. CAD, computer-aided detection; GGO, ground-glass opacity.

CT protocol, SR reconstruction, and CAD analysis

CT acquisition parameters are summarized in Table S1. All contrast-enhanced chest CT examinations were performed using multi-detector scanners from three vendors. Thin-section images (1.0–1.25 mm) were reconstructed according to scanner specifications. For each examination, an additional 5.0 mm thick-slice reconstruction was generated to simulate thick-slice CT data. A commercially available convolutional neural network-based SR reconstruction algorithm (Smart Slicer, version 46.11; Coreline Soft, Seoul, South Korea) was applied to the 5.0 mm thick-slice images to generate SR-converted thin-slice images (1.0–1.25 mm), matching the slice thickness of original thin-section images. Technical details of the SR reconstruction algorithm are provided in Appendix 1. An FDA-cleared AI-assisted CAD system (AVIEW LCS, version 46.11; Coreline Soft) was used to analyze three image sets for each patient: (I) original thin-slice images; (II) original thick-slice images; and (III) SR-converted thin-slice images. The CAD system automatically detected pulmonary nodules and reported quantitative features, including nodule size (mean of the maximum and perpendicular diameters, mm), volume (mm³), and mean attenuation [average Hounsfield units (HU) of all voxels within the segmented nodule], as well as the categorical variable of lobar location.

Visual review of CAD results

All CAD-detected nodules were independently reviewed by a thoracic radiologist with 4 years of experience, blinded to clinical information. Matching and nonmatching classifications were determined by a radiologist based on CAD detection results across image sets, using original thin-slice images as the reference standard: nodules were categorized as matching when the CAD system detected a spatially corresponding nodule in both image sets. Nonmatching nodules were further classified as missing (nodules detected by CAD on original thin-slice images but not detected by CAD on SR-converted images) or hallucinated (nodules detected by CAD only on SR-converted images and not detected by CAD on original thin-slice images).

Nodules detected on original or converted thin-slice images were additionally assessed by the radiologist for pleural and vascular attachment.

Statistical analysis

The detection performance of the CAD system on original thick-slice and SR-converted thin-slice images was evaluated at both nodule and patient levels, using original thin-slice images as the reference standard. Quantitative nodule characteristics and image noise were compared between original and SR-converted thin-slice images (13).

Image noise was defined as the standard deviation of CT attenuation within a manually placed circular region of interest in the tracheal air column, positioned 1 cm above the carina and adjusted to the tracheal lumen (dimensions =5.9 mm × 5.9 mm to 10 mm × 10 mm).

Because quantitative variables exhibited non-normal distributions and were clustered by institution and CT scanner, scanner-specific effects were explicitly accounted for in the analysis. Continuous variables were analyzed using aligned rank transform analysis of variance (ART-ANOVA), with institution and CT scanner included as fixed factors to adjust for scanner-related heterogeneity (Table S2). Categorical variables, including lobar location, were analyzed using the Cochran-Mantel-Haenszel test to control for institutional stratification. A two-sided P<0.05 was considered statistically significant.

Results

Detection performance of AI-based CAD

A total of 96 patients with CRC [mean age, 66.4±10.1 years; 57 men (59.4%)] were included (Table 1). Among 105 pulmonary nodules confirmed on original thin-slice images, 33 (31.4%) were detected on thick-slice images and 64 (61.0%) on the SR-converted thin-slice images. At the nodule level, sensitivity increased from 31.4% to 61.0%, and positive predictive value (PPV) from 42.9% to 80.0%. At the patient level, sensitivity improved from 41.5% to 67.1%, and PPV from 61.4% to 87.2% (Figure 2). These findings indicate that SR reconstruction substantially enhances AI-based detection performance compared with conventional thick-slice CT, although detection sensitivity remained lower than that of original thin-section imaging (Figure 3).

Figure 2 Comparison of AI-assisted CAD performance between original thick-slice and converted thin-slice images. AI, artificial intelligence; CAD, computer-aided detection; NPV, negative predictive value; PPV, positive predictive value.

Figure 3 Representative case of pulmonary nodule detection across different image sets. (A) Original thin-section chest CT image (1.0 mm) shows a small pulmonary nodule (arrow), measuring 5.2 mm in diameter, which was detected by the AI-based CAD system. The image noise level was 26.17 HU. (B) Corresponding original thick-section image (5.0 mm) demonstrates reduced conspicuity, and the nodule was not detected by the AI-based CAD system (arrow). (C) SR-converted thin-section image (1.0 mm) reconstructed from the 5-mm thick-section data improves structural conspicuity, enabling detection of the same nodule by the AI-based CAD system (arrow). The measured nodule size was 4.9 mm, with reduced image noise (17.58 HU). AI, artificial intelligence; CAD, computer-aided detection; CT, computed tomography; HU, Hounsfield units; SR, reconstruction.

Effects of SR reconstruction on nodule characteristics and image noise

SR reconstruction altered image quality metrics while largely preserving nodule size and morphology. Image noise was significantly reduced after SR reconstruction [mean, 12.9 (interquartile range, 9.4–16.6) vs. 20.4 (interquartile range, 13.2–27.4) HU, P<0.001], indicating improved image uniformity and background smoothing (Table 2).

In nodules detected on both original and SR-converted thin-slice images (matched nodules, n=64), anatomic location concordance was high (95.3%), and nodule size showed no significant difference between image sets (5.1±1.6 vs. 5.2±2.0 mm, P=0.985). Although nodule attenuation was significantly lower on SR-converted images (−312.4±142.9 vs. −205.8±112.0 HU, P<0.001), morphologic classification remained stable in most cases, with 80% of nodules maintaining a solid appearance across both reconstructions. These findings indicate that SR reconstruction enhances image quality without materially altering nodule size or essential morphologic characteristics (Table 3).

Hallucinated nodules

Among 80 nodules detected on SR-converted thin-slice images, 16 (20.0%) were classified as hallucinated nodules. These hallucinated nodules were significantly larger than matched nodules (mean size, 7.7±3.5 vs. 5.1±1.6 mm, P<0.001). Most corresponded to benign or artifactual structures, such as subsegmental atelectasis or vascular branches (69%), suggesting a potential risk of overestimating metastatic burden following SR reconstruction (Table 4).

Discussion

This multicenter study demonstrates that deep learning-based SR reconstruction can effectively improve structural conspicuity in thick-slice chest CT, resulting in a substantial improvement in AI-based pulmonary nodule detection in patients with CRC. These findings are particularly relevant in the context of CRC surveillance, where chest CT remains the standard modality for detecting pulmonary metastases (1,2).

Although AI-based CAD systems have improved pulmonary nodule detection and quantification, including in patients with CRC (4-8), their performance remains strongly influenced by CT acquisition parameters, especially slice thickness. Most AI algorithms are optimized for thin-section (≤1.25 mm) CT, whereas real-world oncologic follow-up imaging often relies on 5.0 mm or thicker reconstructions. Prior studies have consistently shown that thicker slices degrade nodule detectability and quantitative reliability (9-12,15-18). Our results extend this body of evidence by demonstrating that SR reconstruction can partially compensate for structural detail degradation associated with thick-slice CT. SR reconstruction nearly doubled both nodule- and patient-level sensitivities by computationally enhancing fine structural conspicuity, resulting in markedly improved detection performance compared with thick-slice images, although performance remained inferior to that of original thin-section CT. This aligns with recent evidence showing that deep learning-based slice-thickness reduction enhances AI performance for lung nodule detection in heterogeneous datasets (19). Importantly, both the SR reconstruction algorithm and CAD system used in this study were developed by the same manufacturer. Therefore, the observed improvements may partly reflect implicit co-optimization between these systems, and the present findings may not be directly generalizable to other combinations of SR reconstruction and CAD software from different vendors.

Beyond improved detection, SR reconstruction enhanced image uniformity while preserving essential nodule characteristics. Significant noise reduction and high morphologic concordance among matched nodules indicate that SR reconstruction acts as both a resolution enhancer and a denoising filter without materially altering nodule size or spatial localization. These findings parallel the known advantages of ultra-high-resolution CT for small pulmonary nodules (20), while further demonstrating that comparable image fidelity can be achieved computationally from standard thick-slice data. Nevertheless, SR reconstruction also resulted in attenuation shifts and apparent morphologic reclassification in a subset of nodules, including conversion from solid to subsolid appearance. These changes may influence nodule risk stratification and downstream clinical interpretation. Because the distinction between solid and subsolid nodules is widely used in guideline-based risk stratification and management decisions, these morphologic shifts may have clinically meaningful consequences if SR-converted images are interpreted as equivalent to native thin-slice CT images. Therefore, SR-converted images should not currently be considered directly interchangeable with native thin-slice CT images for formal nodule characterization or management decision-making.

A key implication of this study is the potential of SR reconstruction as a harmonization tool for multicenter imaging research. Large volumes of thick-slice CT data have accumulated through routine oncologic surveillance but remain underutilized due to heterogeneity in acquisition protocols (11,21). By converting legacy thick-slice CT data into thin-slice-equivalent images, SR reconstruction enables retrospective integration of heterogeneous datasets, facilitating multicenter AI development and validation. This harmonization improves AI generalizability and supports longitudinal and population-level analyses of metastatic disease without additional scanning or radiation exposure. Although reconstruction kernel heterogeneity across institutions may have influenced SR reconstruction behavior and CAD performance, the present results nevertheless suggest the practical feasibility of SR reconstruction after adjustment for CT scanner effects in the statistical analysis. Moreover, this approach aligns with emerging federated and cross-site learning frameworks in medical imaging (22,23) and may serve as a standardized input for multicenter AI validation, as demonstrated in the HANSE LCS harmonization trial (5).

Nevertheless, this study also highlights a key challenge: the emergence of “hallucinated” nodules—false-positive lesions generated by SR algorithms, most of which corresponded to benign or artifactual structures such as subsegmental atelectasis or vascular branches. These false-positive findings likely reflect a trade-off between resolution enhancement and structural fidelity inherent to deep learning-based reconstruction. Although the frequency of hallucinated nodules was not negligible in an oncologic surveillance context, their actual impact on clinical interpretation remains uncertain because reader-performance assessment was beyond the scope of the present study. These findings highlight the need for continued algorithmic refinement and further studies evaluating the clinical consequences of hallucinated nodules in real-world practice.

Although this study was conducted as a multicenter external validation, the relatively small dataset size and focus on commonly encountered 5-mm thick-slice CT images may limit statistical power and generalizability, particularly to other SR-CAD combinations from different vendors. Therefore, the findings should be interpreted as a retrospective preliminary technical feasibility investigation rather than a definitive clinical validation study. Future studies involving larger populations and varying slice thicknesses will be necessary to support broader clinical deployment. Second, the present study evaluated improvements in AI-based CAD performance only and did not assess the impact of SR reconstruction on radiologist interpretation, staging accuracy, or downstream clinical decision-making. In addition, complex real-world imaging conditions, including inflammatory or postoperative pulmonary changes, were not fully represented in the present study cohort. Therefore, the observed improvements remain limited to algorithmic performance metrics, and the potential clinical benefit of SR-assisted reconstruction should be interpreted with caution. Future studies incorporating reader performance analysis and clinical outcome-oriented endpoints will be necessary to determine whether SR reconstruction provides meaningful advantages in real-world oncologic imaging workflows. In addition, individual nodules were not pathologically or longitudinally validated for metastatic status and were reviewed by a single radiologist, as the primary objective of this study was to evaluate lesion detection performance rather than metastatic characterization.

Despite these limitations, the findings consistently demonstrated the robustness of SR reconstruction across heterogeneous CT datasets. Although SR reconstruction inherently involves interpolation of high-frequency image features and may occasionally produce pseudo-nodules, the present results suggest the potential utility of SR-based harmonization for improving AI-assisted analysis of heterogeneous thick-slice CT datasets. Further algorithmic refinement and AI training using datasets with varying slice thicknesses are expected to reduce these minor artifacts.

In conclusion, deep learning–based SR reconstruction represents a promising harmonization strategy for heterogeneous chest CT datasets. By enabling improved AI-based pulmonary nodule detection from thick-slice CT data while preserving essential morphologic information, SR reconstruction bridges historical imaging practices with modern AI analytics and expands the utility of existing oncologic imaging data for multicenter research.

Acknowledgments

None.

Footnote

Funding: This study was supported by Coreline Soft Co., Ltd. (to S.R.K., W.G.J., and G.Y.J.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0880/coif). All authors report a research grant from Coreline Soft Co., Ltd. The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Boards of Wonkwang University Hospital (No. WKUH 2024-01-023-001), Chonnam National University Hwasun Hospital (No. CNUHH-2024-044), and Chonnam National University Hospital (No. CNUH-2024-07-029-009), and individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Benson AB, Venook AP, Adam M, Chang G, Chen YJ, Ciombor KK, et al. NCCN Guidelines® Insights: Rectal Cancer, Version 3.2024. J Natl Compr Canc Netw 2024;22:366-75. [Crossref] [PubMed]
2. Benson AB, Venook AP, Adam M, Chang G, Chen YJ, Ciombor KK, et al. Colon Cancer, Version 3.2024, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 2024;22:e240029. [Crossref] [PubMed]
3. Deboever N, Bayley EM, Eisenberg MA, Hofstetter WL, Mehran RJ, Rice DC, Rajaram R, Roth JA, Sepesi B, Swisher SG, Vaporciyan AA, Walsh GL, Bednarski BK, Morris VK, Antonoff MB. Do resected colorectal cancer patients need early chest imaging? Impact of clinicopathologic characteristics on time to development of pulmonary metastases. J Surg Oncol 2024;129:331-7.
4. Liu JA, Yang IY, Tsai EB. Artificial Intelligence (AI) for Lung Nodules, From the AJR Special Series on AI Applications. AJR Am J Roentgenol 2022;219:703-12. [Crossref] [PubMed]
5. Kondrashova R, Klimeš F, Kaireit TF, May K, Barkhausen J, Stiebeler S, Sperl J, Dettmer S, Wacker F, Vogel-Claussen J. Comparison of AI software tools for automated detection, quantification and categorization of pulmonary nodules in the HANSE LCS trial. Sci Rep 2024;14:27809. [Crossref] [PubMed]
6. Peters AA, Wiescholek N, Müller M, Klaus J, Strodka F, Macek A, Primetis E, Drakopulos D, Huber AT, Obmann VC, Ruder TD, Roos JE, Heverhagen JT, Christe A, Ebner L. Impact of artificial intelligence assistance on pulmonary nodule detection and localization in chest CT: a comparative study among radiologists of varying experience levels. Sci Rep 2024;14:22447. [Crossref] [PubMed]
7. Lee JJB, Suh YJ, Oh C, Lee BM, Kim JS, Chang Y, Jeon YJ, Kim JY, Park SY, Chang JS. Automated Computer-Aided Detection of Lung Nodules in Metastatic Colorectal Cancer Patients for the Identification of Pulmonary Oligometastatic Disease. Int J Radiat Oncol Biol Phys 2022;114:1045-52. [Crossref] [PubMed]
8. Jang S, Kim J, Lee JS, Jeong Y, Nam JG, Kim J, Lee KW. Real-World Diagnostic Performance and Clinical Utility of Artificial Intelligence-Assisted Interpretation for Detection of Lung Metastasis on CT in Patients With Colorectal Cancer. AJR Am J Roentgenol 2025;225:e2533063. [Crossref] [PubMed]
9. Li Q, Li F, Doi K. Computerized detection of lung nodules in thin-section CT images by use of selective enhancement filters and an automated rule-based classifier. Acad Radiol 2008;15:165-75. [Crossref] [PubMed]
10. Goo JM. Deep Learning-based Super-Resolution Algorithm: Potential in the Management of Subsolid Nodules. Radiology 2021;299:220-1. [Crossref] [PubMed]
11. Yu P, Zhang H, Wang D, Zhang R, Deng M, Yang H, et al. Spatial resolution enhancement using deep learning improves chest disease diagnosis based on thick slice CT. NPJ Digit Med 2024;7:335. [Crossref] [PubMed]
12. Fischbach F, Knollmann F, Griesshaber V, Freund T, Akkol E, Felix R. Detection of pulmonary nodules by multislice computed tomography: improved detection rate with reduced slice thickness. Eur Radiol 2003;13:2378-83. [Crossref] [PubMed]
13. Park S, Lee SM, Kim W, Park H, Jung KH, Do KH, Seo JB. Computer-aided Detection of Subsolid Nodules at Chest CT: Improved Performance with Deep Learning-based CT Section Thickness Reduction. Radiology 2021;299:211-9. [Crossref] [PubMed]
14. Brierley JD, Gospodarowicz MK, Wittekind C, editors. TNM Classification of Malignant Tumours. 8th ed. Hoboken: John Wiley & Sons, Inc.; 2017.
15. Sinsuat M, Saita S, Kawata Y, Niki N, Ohmatsu H, Tsuchida T, Kakinuma R, Kusumoto M, Eguchi K, Kaneko M, Morikubo H, Moriyama N. Influence of slice thickness on diagnoses of pulmonary nodules using low-dose CT: potential dependence of detection and diagnostic agreement on features and location of nodule. Acad Radiol 2011;18:594-604. [Crossref] [PubMed]
16. Li WJ, Chu ZG, Zhang Y, Li Q, Zheng YN, Lv FJ. Effect of Slab Thickness on the Detection of Pulmonary Nodules by Use of CT Maximum and Minimum Intensity Projection. AJR Am J Roentgenol 2019;213:562-7. [Crossref] [PubMed]
17. Gupta S, Nayak K, Pendem S. Impact of slice thickness on reproducibility of CT radiomic features of lung tumors. F1000Res 2023;12:1319. [Crossref] [PubMed]
18. Kawel N, Seifert B, Luetolf M, Boehm T. Effect of slab thickness on the CT detection of pulmonary nodules: use of sliding thin-slab maximum intensity projection and volume rendering. AJR Am J Roentgenol 2009;192:1324-9. [Crossref] [PubMed]
19. Jeong J, Park D, Kang JH, Kim M, Kim HY, Choi W, Ham SY. Deep Learning-Based Slice Thickness Reduction for Computer-Aided Detection of Lung Nodules in Thick-Slice CT. Diagnostics (Basel) 2024;14:2558. [Crossref] [PubMed]
20. Higashibori H, Fukumoto W, Kusuda S, Yokomachi K, Mitani H, Nakamura Y, Awai K. Performance of a deep-learning-based lung nodule detection system using 0.25-mm thick ultra-high-resolution CT images. Jpn J Radiol 2025;43:1842-8. [Crossref] [PubMed]
21. Park S, Lee SM, Do KH, Lee JG, Bae W, Park H, Jung KH, Seo JB. Deep Learning Algorithm for Reducing CT Slice Thickness: Effect on Reproducibility of Radiomic Features in Lung Cancer. Korean J Radiol 2019;20:1431-40. [Crossref] [PubMed]
22. Guan H, Yap PT, Bozoki A, Liu M. Federated learning for medical image analysis: A survey. Pattern Recognit 2024;151:110424. [Crossref] [PubMed]
23. Rehman MHU, Hugo Lopez Pinaya W, Nachev P, Teo JT, Ourselin S, Cardoso MJ. Federated learning for medical imaging radiology. Br J Radiol 2023;96:20220890. [Crossref] [PubMed]