Robustness of pulmonary nodule radiomic features on computed tomography as a function of varying radiation dose levels—a multi-dose in vivo patient study

Gijs A. Bartholomeus, Wouter A. C. van Amsterdam, Annemarie M. den Harder, Martin J. Willemink, Robbert W. van Hamersvelt, Pim A. de Jong, Tim LeinerEuropean 2023 · DOI

TL;DR

This study investigates how radiation dose levels affect radiomic features extracted from lung nodules in CT scans. The key finding is that most radiomic features are not robust to dose variations, but shape-based features show remarkable stability. While a linear model can correct for some dose-related variability, its accuracy decreases at lower doses.

What Problem Does This Paper Address?

Radiomics, the extraction of quantitative features from medical images, holds great promise for improving lung cancer diagnosis, prognostication, and treatment monitoring. However, for radiomic features to be clinically useful, they must be reproducible and reliable. A significant challenge is the variability introduced by different imaging parameters, particularly radiation dose. Lowering radiation dose in CT scans is desirable to reduce patient exposure, but it can increase image noise, potentially altering radiomic feature values. While phantom studies have explored this, this paper addresses the critical need for in vivo validation of radiomic feature robustness across varying radiation doses in actual patients. The authors aim to determine which radiomic features are stable and whether dose-induced variations can be corrected.

Methods

The researchers conducted an in vivo, multi-dose study on 19 patients with pulmonary nodules. Each patient underwent four CT scans in a single session at different radiation dose levels (60, 33, 24, and 15 mAs). Pulmonary nodules were manually segmented. To assess feature robustness, they calculated the Intra-Class Correlation Coefficient (ICC), a measure of agreement between measurements. Additionally, they fitted linear regression models for each feature to quantify bias (average deviation from the full-dose value) and R² (proportion of variability explained by dose). This allowed them to evaluate how well dose-induced variations could be corrected.

Datasets

The study utilized a dataset of 32 pulmonary nodules from 15 patients (initially 19 patients were included, but some were excluded due to lung masses). The scans were acquired using a 256-slice CT system. No new benchmark datasets were introduced; the focus was on analyzing existing CT data under controlled, multi-dose conditions.

Key Insights and Ideas

  • Shape Features Reign Supreme: The most striking finding is the exceptional robustness of shape-based radiomic features. These features consistently showed high ICC values, indicating they are the least affected by changes in radiation dose. This suggests that geometric properties of nodules are inherently more stable than texture-based features.
  • Most Features Are Fragile: Contrary to the stability of shape features, the vast majority of commonly used radiomic features (from GLCM, GLDM, GLRLM, GLSZM, and first-order categories) were found to be significantly influenced by radiation dose. This highlights a major hurdle for widespread clinical adoption of many radiomic approaches.
  • Linear Correction Offers Partial Hope: For a substantial portion of the unstable features, a simple linear model could effectively correct for dose-induced variability. This is a promising development, suggesting that some features might be salvageable for clinical use even with dose variations.
  • Low Dose is the Achilles' Heel: While linear correction is possible, its accuracy diminishes significantly at lower radiation doses (e.g., 15 mAs). This implies that extremely low-dose CT scans may still pose a challenge for reliable radiomic analysis, even with correction methods.
  • Filters Can Matter (Sometimes): The study explored the impact of image filters. While most filters did not improve robustness, the Laplacian of Gaussian (LoG) filter showed a notable ability to make features more robust compared to the original, unfiltered images.

Key Results

  • Only 15% of the 100 features analyzed (without filters) were considered "stable" (ICC > 0.9).
  • Shape features were disproportionately represented among the most robust features, with 10 out of 14 shape features achieving an ICC > 0.9.
  • Bias generally increased, and R² decreased as radiation dose decreased.
  • Shape features also demonstrated better "correctability" (higher R²) by linear models compared to other feature classes.
  • The study identified specific features with high ICC values, such as original_glrlm_GrayLevelNonUniformity (0.966) and various shape features like shape_VoxelVolume (0.96).

Limitations and Future Work

The authors acknowledge several limitations:

  • Manual Segmentation: Nodules were manually delineated by a single investigator. Previous studies suggest automated or semi-automated segmentation might yield more robust results. Future work should explore the interaction between segmentation methods and dose variation.
  • Small Sample Size: The study included a relatively small number of nodules (32). Larger datasets would strengthen the findings.
  • Focus on Stability, Not Clinical Value: The study concentrated solely on feature stability and did not assess the prognostic or diagnostic value of these features. Future research should investigate if these stable or correctable features can indeed improve clinical decision-making.
  • Limited Pre-processing: The study did not investigate the impact of other pre-processing steps (e.g., bin width, reconstruction algorithms) on feature reproducibility.
  • FBP Reconstruction: Radiomic features were extracted from Filtered Back Projection (FBP) reconstructed scans. With the increasing use of iterative reconstruction methods for low-dose CT, future studies should examine their impact on radiomic feature robustness.

The authors suggest that future studies should confirm the prognostic/diagnostic value of the identified robust features and investigate the interplay of other variables like segmentation and reconstruction algorithms with dose variation. They also point to the potential of unsupervised deep learning techniques as an alternative approach that might be less sensitive to dose variations.

Summary updated

Abstract

• The vast majority of commonly used radiomic features are strongly influenced by variations in radiation dose level. • A small minority of radiomic features, notably the shape feature class, are robust against dose-level variations according to ICC calculations. • A large subset of radiomic features can be corrected by a linear model taking into account only the radiation dose level.
BibTeX
@inproceedings{bartholomeus2023robustness,
  author = {Gijs A. Bartholomeus and Wouter A. C. van Amsterdam and Annemarie M. den Harder and Martin J. Willemink and Robbert W. van Hamersvelt and Pim A. de Jong and Tim Leiner},
  title = {{Robustness of pulmonary nodule radiomic features on computed tomography as a function of varying radiation dose levels—a multi-dose in vivo patient study}},
  booktitle = {European},
  year = {2023},
  doi = {10.1007/s00330-023-09643-8},
}
Figure