# Impact of CT acquisition settings on the stability of radiomic features and the performance of pulmonary nodule classification models

## TL;DR

This study shows that CT scan settings affect the stability of radiomic features, and using stable features improves the reliability of models for classifying lung nodules.

## Contribution

The study evaluates how CT acquisition settings influence radiomic feature stability and model performance, emphasizing the importance of stable features for generalizability.

## Key findings

- Slice thickness had the greatest impact on radiomic feature stability, while image transmission methods had the least.
- Models using stable features showed consistent performance across different CT settings, unlike those using unstable features.
- Full-feature and intermediate stable models had higher AUCs in training but performed poorly in test sets with varying CT parameters.

## Abstract

To evaluate the stability of radiomic features under different CT acquisition settings and investigate its impact on diagnostic model performance and generalizability.

198 patients with 1227 pulmonary nodules underwent chest CT scans using varied settings (three slice thicknesses, two reconstruction matrices, six convolution kernels, two transmission methods). 1394 radiomic features were extracted per nodule. Feature stability was evaluated using the Intraclass Correlation Coefficient (ICC, stable: ICC ≥ 0.8, intermediate stable: 0.4 < ICC < 0.8, unstable: ICC ≤ 0.4). Four diagnostic models (Full-feature, Stable, Unstable, Intermediate stable) were developed using two datasets (lung cancer screening, n = 184; clinical scenarios, n = 1192). In addition, three combination models were constructed for ablation analysis. Model performance and generalizability were assessed via fivefold cross-validation and independent test sets with different CT parameters.

Slice thickness and image transmission methods had the greatest and least impacts on feature stability (7.0% and 83.0% stable features, respectively). In training and validation sets, the Full-feature and Intermediate stable models showed higher AUCs than the Stable and Unstable models (p < 0.05). However, in test sets with varying CT parameters, the Stable model maintained consistent performance (AUC: 0.693–0.728), while the Unstable model exhibited the greatest variability (AUC: 0.523–0.800). Notably, the Full-feature and Intermediate stable models largely predicted nodules as benign, exhibiting limited ability to discriminate malignant cases.

Radiomic feature stability is significantly affected by CT reconstruction parameters, especially slice thickness. Models based on stable features demonstrate better generalizability across varying CT settings, underscoring the importance of assessing feature stability in radiomic-based diagnostics.

Radiomic feature stability is significantly affected by CT acquisition parameters. Stable radiomic features enhance diagnostic model consistency and reliability across diverse CT settings. Therefore, feature stability analysis and selection of stable features are crucial to enhance model generalizability and stability.

How do CT settings affect radiomic feature stability and model performance?Feature stability varies with CT parameters, but stable features enhance model generalizability.Stable feature models boost diagnostic reliability and clinical applicability.

How do CT settings affect radiomic feature stability and model performance?

Feature stability varies with CT parameters, but stable features enhance model generalizability.

Stable feature models boost diagnostic reliability and clinical applicability.