The Effect of CT Scan Parameters on the Measurement of CT Radiomic Features: A Lung Nodule Phantom Study

Young Jae Kim, Hyun-Ju Lee, Kwang Gi Kim, Seung Hyun LeeComput. Math. Methods Medicine 2019 · DOI

TL;DR

This study investigates how common CT scan parameters (slice thickness, mAs, and reconstruction algorithm) affect radiomic features extracted from lung nodules. The findings reveal that these parameters significantly influence radiomic measurements, highlighting the critical need for standardization in clinical radiomics and deep learning applications to ensure reproducible and reliable results.

What Problem Does This Paper Address?

Radiomics, the extraction of quantitative features from medical images, holds great promise for improving disease diagnosis and prognosis. However, the reliability of radiomic features is heavily influenced by the parameters used during image acquisition and reconstruction. Different CT scanners, protocols, and reconstruction algorithms can lead to variations in image quality and, consequently, in the radiomic features extracted. This variability poses a significant challenge for comparing results across studies, developing robust diagnostic models, and translating radiomic findings into clinical practice. This paper specifically addresses the need to quantify the impact of common CT scan parameters on radiomic features derived from lung nodules, a critical area for lung cancer diagnosis and management.

Methods

The researchers employed a systematic approach using an anthropomorphic thorax phantom containing spherical lung nodule phantoms with known attenuation values (100 HU for solid nodules and -630 HU for ground-glass nodules). CT scans were performed using a 64-channel multi-detector row CT scanner under 12 different combinations of scan parameters:

  • Slice Thickness: 1.25 mm, 2.5 mm, and 5.0 mm
  • mAs (Milliampere-seconds): 30 mAs and 120 mAs
  • Reconstruction Algorithms: "Lung" and "Standard"

Each scan parameter combination was repeated 10 times with nodule relocation to assess reproducibility. Radiomic features were extracted from the segmented nodules. A total of 20 features were analyzed, encompassing histogram-based features (e.g., mean, standard deviation, skewness, kurtosis, energy, entropy) and texture features derived from Gray-Level Co-occurrence Matrix (GLCM) and Gray-Level Run Length Matrix (GLRLM). Multiple regression analysis was used to evaluate the effect of each scan parameter on these features, and Cohen's d was calculated to assess the absolute effect size of these differences.

Datasets

The study utilized an anthropomorphic thorax phantom and two types of nodule phantoms (100 HU and -630 HU) with known attenuation values. No new benchmark datasets were introduced; instead, the study focused on evaluating the impact of scan parameters on features extracted from these controlled phantom datasets.

Key Insights and Ideas

  • Scan Parameters Matter Significantly: The core takeaway is that CT scan parameters are not trivial and have a substantial impact on radiomic feature values. This underscores the inherent variability that can arise when comparing radiomic data acquired under different conditions.
  • Slice Thickness Dominates for Solid Nodules: For the 100 HU (solid) nodules, slice thickness emerged as a particularly influential parameter. This is attributed to its strong relationship with partial volume effects, which are more pronounced in denser, solid structures.
  • Reconstruction Algorithm is Key for Ground-Glass Nodules: In contrast, for the -630 HU (ground-glass) nodules, the reconstruction algorithm had a greater effect. This is likely due to the increased noise in these lower-attenuation nodules, which is more susceptible to variations introduced by different reconstruction kernels.
  • Need for Standardization: The findings strongly advocate for standardization of CT acquisition and reconstruction protocols when performing radiomics studies or developing deep learning models for lung nodule analysis. Without standardization, comparing results across different institutions or even within the same institution over time becomes problematic.
  • Texture Features are Sensitive: Many texture features (GLCM and GLRLM) were found to be significantly affected by scan parameters, indicating their sensitivity to image noise and resolution.

Key Results

  • Widespread Significance: Across both 100 HU and -630 HU nodule phantoms, a large majority of the 20 radiomic features showed statistically significant differences (p ≤ 0.05) when comparing different scan parameters.
  • Slice Thickness Impact (100 HU): For 100 HU nodules, 18-20 out of 20 features showed significant differences between various slice thickness comparisons (e.g., 5.0 mm vs. 1.25 mm).
  • Reconstruction Algorithm Impact (-630 HU): For -630 HU nodules, 18-19 features showed significant differences between reconstruction algorithms.
  • mAs Impact: Both nodule types showed significant differences in 18-19 features when comparing 30 mAs and 120 mAs.
  • Effect Size: The study quantified the magnitude of these effects using Cohen's d, revealing that slice thickness had a large effect size for 100 HU nodules, while the reconstruction algorithm had a large effect size for -630 HU nodules.

Limitations and Future Work

  • Phantom vs. Real Data: The study was conducted using phantoms, which, while controlled, do not perfectly replicate the complexity and variability of real patient data. Future work should validate these findings on clinical datasets.
  • Limited Parameter Range: The study explored a specific set of parameters. Further research could investigate a wider range of parameters, including different CT vendors, tube voltages, and advanced reconstruction techniques.
  • Standardization Methods: The authors suggest the need to develop methods for standardizing computer features obtained from different scan parameters. This could involve developing robust normalization techniques or feature selection strategies that are less sensitive to parameter variations.
  • Clinical Correlation: While the study quantifies the technical impact, future work should aim to correlate these parameter-induced feature changes with actual clinical outcomes or diagnostic performance.

Summary updated

BibTeX
@article{kim2019effect,
  author = {Young Jae Kim and Hyun-Ju Lee and Kwang Gi Kim and Seung Hyun Lee},
  title = {{The Effect of CT Scan Parameters on the Measurement of CT Radiomic Features: A Lung Nodule Phantom Study}},
  journal = {Comput. Math. Methods Medicine},
  year = {2019},
  doi = {10.1155/2019/8790694},
}
Figure