Impact of deep learning image reconstruction on CT radiomics reproducibility in patients with pulmonary nodules
Abstract
Objective: Ensuring CT feature reproducibility across imaging protocols and reconstruction algorithms is essential for the generalizability of radiomics applications. This study aimed to evaluate the reproducibility of CT radiomic features in pulmonary nodules using deep learning image reconstruction (DLIR) compared with conventional reconstruction techniques under different radiation doses.
Materials and methods: We included 94 patients between March and June 2023. Standard- and low-dose CT images were scanned to assess the impact of dose reduction on feature reproducibility. DLIR strengths at low (DLIR-L), medium (DLIR-M), and high (DLIR-H), representing progressive levels of noise suppression, were compared against filtered back projection (FBP) and adaptive statistical iterative reconstruction-V at 50% strength (ASIR-V-50%) reconstructions. 2264 features were extracted from pulmonary nodules. The concordance correlation coefficient (CCC) and the coefficient of variation (CV) were used to evaluate inter-dose reproducibility, inter-reconstruction reproducibility, and feature variability.
Results: The mean CCC for inter-dose reproducibility using DLIR-H, ASIR-V-50%, and FBP was 0.934 ± 0.104, 0.922 ± 0.128, and 0.922 ± 0.113, respectively ( p = 0.055), with 29.8%, 27.9%, and 20.2% of features showing near-perfect agreement. In low-dose CT, DLIR-L achieved the highest inter-reconstruction reproducibility against ASIR-V-50% (0.950 ± 0.094), with 39.4% of features showing near-perfect agreement. The overall CV across reconstructions was 9.87 ± 20.32%, with 74.0% of features highly stable. Part-solid nodules showed the lowest variability (7.17 ± 13.48%, p < 0.001). Reproducibility improved significantly with increasing nodule size ( p < 0.001).
Conclusion: DLIR improves the radiomic feature reproducibility in pulmonary nodules, making it particularly suitable for low-dose CT protocols. However, the robustness tends to decrease with higher DLIR strengths. Nodule characteristics significantly influence variability.
Key Points: Question The impact of DLIR on radiomic feature reproducibility across radiation doses and reconstruction methods in pulmonary nodules remains unknown.
Findings DLIR-H demonstrated high inter-dose reproducibility, while DLIR-L achieved the highest inter-reconstruction reproducibility; feature variability decreased significantly with increasing nodule size, with 21 robust features identified.
Critical relevance statement This study critically assesses reconstruction-related instability in CT radiomics and identifies DLIR settings and nodule characteristics that influence feature robustness, supporting protocol optimization and more reliable quantitative pulmonary nodule assessment in clinical radiology.
Related
- Reconstruction-aware radiomics for risk stratification of subsolid pulmonary nodules: technical evidence, clinical limits, and translation priorities.
- Radiomics-based interpretable machine learning model for identifying high-risk patients with small bowel obstruction: a multicenter study.
- Multiparametric MRI radiomics and machine learning for outcome prediction in acute ischemic stroke: a retrospective study of 353 patients
- Nomogram models integrating multiregion radiomics, quantitative computed tomography, and clinical indicators for distinguishing connective tissue disease-associated interstitial lung disease from idiopathic pulmonary fibrosis and improving the prediction of gender-age-physiology stage
- Development and validation of an interpretable machine learning model for early prediction of mortality in ICU patients with chronic obstructive pulmonary disease and heart failure
BibTeX
@article{zhu2026impact,
title = {{Impact of deep learning image reconstruction on CT radiomics reproducibility in patients with pulmonary nodules}},
author = {Huiyuan Zhu and Tingle Wei and Luna Wang and Lin Zhu and Qunhui Chen and Guangyu Tao and Linlin Sun and Xiwen Sun and Zike Huang and Hong Yu},
journal = {Insights into imaging},
year = {2026},
month = oct,
doi = {10.1186/s13244-026-02417-y},
url = {https://europepmc.org/article/PMC/PMC13661886},
note = {CC BY},
}
RIS
TY - JOUR
TI - Impact of deep learning image reconstruction on CT radiomics reproducibility in patients with pulmonary nodules
AU - Huiyuan Zhu
AU - Tingle Wei
AU - Luna Wang
AU - Lin Zhu
AU - Qunhui Chen
AU - Guangyu Tao
AU - Linlin Sun
AU - Xiwen Sun
AU - Zike Huang
AU - Hong Yu
PY - 2026
DA - 2026/10/13
JO - Insights into imaging
DO - 10.1186/s13244-026-02417-y
UR - https://europepmc.org/article/PMC/PMC13661886
AB - Objective: Ensuring CT feature reproducibility across imaging protocols and reconstruction algorithms is essential for the generalizability of radiomics applications. This study aimed to evaluate the reproducibility of CT radiomic features in pulmonary nodules using deep learning image reconstruction (DLIR) compared with conventional reconstruction techniques under different radiation doses. Materials and methods: We included 94 patients between March and June 2023. Standard- and low-dose CT images were scanned to assess the impact of dose reduction on feature reproducibility. DLIR strengths at low (DLIR-L), medium (DLIR-M), and high (DLIR-H), representing progressive levels of noise suppression, were compared against filtered back projection (FBP) and adaptive statistical iterative reconstruction-V at 50% strength (ASIR-V-50%) reconstructions. 2264 features were extracted from pulmonary nodules. The concordance correlation coefficient (CCC) and the coefficient of variation (CV) were used to evaluate inter-dose reproducibility, inter-reconstruction reproducibility, and feature variability. Results: The mean CCC for inter-dose reproducibility using DLIR-H, ASIR-V-50%, and FBP was 0.934 ± 0.104, 0.922 ± 0.128, and 0.922 ± 0.113, respectively ( p = 0.055), with 29.8%, 27.9%, and 20.2% of features showing near-perfect agreement. In low-dose CT, DLIR-L achieved the highest inter-reconstruction reproducibility against ASIR-V-50% (0.950 ± 0.094), with 39.4% of features showing near-perfect agreement. The overall CV across reconstructions was 9.87 ± 20.32%, with 74.0% of features highly stable. Part-solid nodules showed the lowest variability (7.17 ± 13.48%, p < 0.001). Reproducibility improved significantly with increasing nodule size ( p < 0.001). Conclusion: DLIR improves the radiomic feature reproducibility in pulmonary nodules, making it particularly suitable for low-dose CT protocols. However, the robustness tends to decrease with higher DLIR strengths. Nodule characteristics significantly influence variability. Key Points: Question The impact of DLIR on radiomic feature reproducibility across radiation doses and reconstruction methods in pulmonary nodules remains unknown. Findings DLIR-H demonstrated high inter-dose reproducibility, while DLIR-L achieved the highest inter-reconstruction reproducibility; feature variability decreased significantly with increasing nodule size, with 21 robust features identified. Critical relevance statement This study critically assesses reconstruction-related instability in CT radiomics and identifies DLIR settings and nodule characteristics that influence feature robustness, supporting protocol optimization and more reliable quantitative pulmonary nodule assessment in clinical radiology.
SN - 1869-4101
ER -