Assessing robustness of radiomic features by image perturbation (1806.06719v1)

Abstract: Image features need to be robust against differences in positioning, acquisition and segmentation to ensure reproducibility. Radiomic models that only include robust features can be used to analyse new images, whereas models with non-robust features may fail to predict the outcome of interest accurately. Test-retest imaging is recommended to assess robustness, but may not be available for the phenotype of interest. We therefore investigated 18 methods to determine feature robustness based on image perturbations. Test-retest and perturbation robustness were compared for 4032 features that were computed from the gross tumour volume in two cohorts with computed tomography imaging: I) 31 non-small-cell lung cancer (NSCLC) patients; II): 19 head-and-neck squamous cell carcinoma (HNSCC) patients. Robustness was measured using the intraclass correlation coefficient (1,1) (ICC). Features with ICC$\geq0.90$ were considered robust. The NSCLC cohort contained more robust features for test-retest imaging than the HNSCC cohort ($73.5\%$ vs. $34.0\%$). A perturbation chain consisting of noise addition, affine translation, volume growth/shrinkage and supervoxel-based contour randomisation identified the fewest false positive robust features (NSCLC: $3.3\%$; HNSCC: $10.0\%$). Thus, this perturbation chain may be used to assess feature robustness.

• The paper introduces image perturbation as a systematic alternative to test-retest imaging for assessing the robustness of radiomic features, aiming to improve reproducibility.
• Analyzing two patient cohorts (NSCLC, HNSCC), the study identified a specific perturbation chain that effectively minimized false positives in feature robustness assessment.
• This perturbation approach is computationally easier, less resource-intensive, and offers a scalable method to evaluate feature robustness, enhancing the feasibility and reliability of radiomic models in clinical settings.

Insights on Assessing Robustness of Radiomic Features

This paper focuses on a sophisticated methodology to evaluate the robustness of radiomic features through image perturbation, challenging the traditional reliance on test-retest imaging. The authors aim to ensure reproducibility in radiomic models used for analyzing medical images.

Radiomics allows the extraction of a vast array of quantitative features from medical images to support model-based treatment decisions. It is dependent on the robustness of these image features against variations caused by patient positioning, imaging protocols, or segmentation techniques. Commonly, test-retest imaging is employed to identify non-robust features, but it is not always feasible due to logistical and practical constraints. The authors present an alternative by introducing a systematic method of image perturbation, examining 18 combinatorial perturbation techniques to assess feature stability.

The paper involves two distinct cohorts: 31 patients with non-small-cell lung cancer (NSCLC) and 19 patients with head-and-neck squamous cell carcinoma (HNSCC). They analyzed computed tomography (CT) scans, focusing on a large set of 4032 features extracted from the gross tumor volume. Robustness was quantitatively measured employing the Intraclass Correlation Coefficient (ICC), stipulating that features with ICC≥0.90 are robust.

Significant findings include a marked discrepancy in feature robustness between the NSCLC and HNSCC cohorts when assessed using test-retest conditions, with NSCLC showing a higher percentage (73.5% vs. 34.0%). Among the perturbation methods evaluated, a chain including noise addition, affine translation, volume growth/shrinkage, and supervoxel-based contour randomization emerged as most proficient in minimizing false positives regarding feature robustness. This precise perturbation chain identified the fewest false positives: 3.3% for NSCLC and 10.0% for HNSCC.

The paper highlights the improvements in ICC precision, especially evident in the HNSCC cohort, which had a broader confidence interval for test-retest ICC. The computational ease and the reduced resource demand render perturbation methods pragmatic for routine radiomics studies. The perturbation techniques afford a scalable option to assess feature robustness that could surmount the limitations posed by test-retest imaging due to mode-specific constraints or lack of accessible datasets.

The theoretical implications of this paper are significant, suggesting that radiomic model generalizability can be enhanced without the traditional test-retest imaging. The approach of utilizing perturbations aligns with strategies in deep learning, where models are trained with perturbed data to improve robustness across diverse inputs. Practically, this work paves the way for more feasible integration of radiomic models in clinical workflows by ensuring reliability across different imaging modalities without extraneous imaging.

Future work could extend these methodologies to other imaging modalities such as MRI and PET to confirm cross-modal application. Additionally, exploring the incorporation of multiple delineation datasets and investigating the variability introduced by manual segmentation can refine the robustness assessment.

In conclusion, this paper provides a compelling advancement in the field of radiomics by effectively utilizing image perturbations to ascertain feature robustness, offering a viable alternative to test-retest imaging and amplification of model reliability.