Deep learning-based CT slice synthesis improves radiomic feature reproducibility and discriminative performance in lung nodule assessment.
Authors
Affiliations (7)
Affiliations (7)
- Department of Computer Center, General Hospital of Ningxia Medical University, Yinchuan, China.
- Department of Radiology, General Hospital of Ningxia Medical University, Yinchuan, China.
- Ningxia Medical University, Yinchuan, China.
- Department of Key Laboratory of Ningxia Stem Cell and Regenerative Medicine, Institute of Medical Sciences, General Hospital of Ningxia Medical University, Yinchuan, China. [email protected].
- Department of Pulmonary and Critical Care Medicine, General Hospital of Ningxia Medical University, Yinchuan, China. [email protected].
- Department of Key Laboratory of Ningxia Stem Cell and Regenerative Medicine, Institute of Medical Sciences, General Hospital of Ningxia Medical University, Yinchuan, China. [email protected].
- Department of Pulmonary and Critical Care Medicine, General Hospital of Ningxia Medical University, Yinchuan, China. [email protected].
Abstract
To investigate the effect of CT slice thickness on radiomic features (RFs) in terms of reproducibility and discriminative power, and to assess whether a deep learning-based CT slice synthesis (DLS) algorithm can mitigate the adverse effects associated with thick-slice CT. This retrospective multicenter study included 506 patients with lung nodules (245 benign, 261 malignant) from two independent cohorts, which were divided into a training set, internal validation set (IVS), and external validation set (EVS). Chest CT reconstructed at 1-mm and 5-mm slice thicknesses was analyzed. A DLS algorithm was applied to convert 5-mm CT into synthetic 1-mm CT. RFs were extracted from all CT types to construct radiomics models. Reproducibility was assessed using the concordance correlation coefficient (CCC) and compared with the Wilcoxon signed-rank test. Discriminative power was evaluated by the area under the receiver operating characteristic curve (AUC) and compared with DeLong's test. The CCCs of DLS 1-mm CT were 0.48 ± 0.37 and 0.49 ± 0.37 in Cohort 1 and Cohort 2, respectively, significantly higher than real 5-mm CT (all p < 0.001). Most RFs from 5-mm CT lacked reproducibility (CCC ≥ 0.85; 0.9% in both Cohort 1 and Cohort 2), whereas DLS 1-mm CT showed marked improvement (Cohort 1, 27.6%; Cohort 2, 26.9%). The discriminative power of RFs from DLS 1-mm CT was superior to that of 5-mm CT and non-inferior to real 1-mm CT, both in model construction and evaluation. CT slice thickness substantially influences the reproducibility and discriminative power of RFs, whereas the DLS algorithm effectively mitigates the limitations associated with thick-slice CT. Deep learning-based CT slice synthesis significantly reduces the slice thickness-related variability in radiomics feature reproducibility and discriminative power, providing a promising methodological approach to improve radiomics standardization and support its clinical translation. CT slice thickness variability substantially impairs radiomic feature reproducibility and discriminative performance, posing a major barrier to standardized radiomics analysis. Across two independent cohorts, deep learning-based slice synthesis mitigated the adverse effects of thick-slice CT on radiomic feature reproducibility and cross-thickness discriminative performance.