Interpretability of radiomics models is improved when using feature group selection strategies for predicting molecular and clinical targets in clear-cell renal cell carcinoma

Cancer Imaging 2023 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Why Radiomics Interpretability Matters in ccRCC

Radiomics, the extraction of quantitative features from medical images, has shown promise for predicting molecular and histopathological characteristics of clear-cell renal cell carcinoma (ccRCC) non-invasively. However, conventional radiomics pipelines often select features that are statistically powerful but clinically difficult to interpret.

Improving interpretability is essential for clinical translation: if clinicians cannot understand which image characteristics drive predictions, they are unlikely to trust or act on model outputs. Feature group selection strategies can help by ensuring selected features represent meaningful, coherent biological concepts rather than arbitrary statistical combinations.

This study used the TRACERx Renal prospective cohort to compare a conventional radiomics pipeline against a novel hierarchical feature group selection approach across 20 clinical and molecular classification targets in ccRCC.

TL;DR: A new hierarchical feature group selection strategy was developed to improve both the accuracy and interpretability of radiomics models for predicting ccRCC molecular and clinical characteristics.
Pages 2-5
TRACERx Renal Cohort and Dual Pipeline Comparison

The study used the first 101 patients from the TRACERx Renal study, a prospective multi-omics cohort designed to study tumor evolution in ccRCC. For each patient, CT-derived radiomic features were extracted from multiple tumor sub-segmentations including whole tumor, core, rim, high enhancing, and low enhancing regions.

A total of 105 radiomic features per region of interest were extracted using the IBSI-standardized pyradiomics framework. Twenty classification targets were assessed: six histopathological (including necrosis, sarcomatoid change, renal vein invasion) and fourteen molecular targets (chromosomal events like Loss 9p21.3 and evolutionary metrics like linear EvoST).

Two pipelines were compared: the Conventional pipeline used standard Correlation-based Feature Reduction (CFR) followed by LR-LASSO feature selection. The Proposed pipeline added a hierarchical CFR step and a feature group selection layer before the final LR-LASSO, forcing selected features to represent distinct, interpretable feature families.

Model performance was assessed by AUC across all 20 targets, with statistical significance set at p-value below 0.05 after permutation testing. Both pipelines were evaluated for predictive performance and interpretability of selected features.

TL;DR: Two radiomic pipelines were compared across 20 clinical and molecular targets in 101 ccRCC patients from TRACERx Renal, with the new pipeline adding hierarchical feature group selection.
Pages 5-8
Eleven Targets Significant, Five Exceed AUC of 0.8

Out of 20 classification targets tested, 11 achieved statistical significance (p-value below 0.05) with either the conventional or proposed pipeline. This demonstrates that CT radiomics captures meaningful biological information across a wide range of ccRCC characteristics.

Five targets achieved AUC above 0.8: necrosis, renal vein invasion, overall stage, linear EvoST (an evolutionary tumor score), and Loss 9p21.3 (a key chromosomal deletion). These represent both clinically actionable histopathological features and biologically informative molecular events.

The proposed pipeline with feature group selection achieved equivalent or better AUC on most targets compared to the conventional pipeline, while simultaneously improving the interpretability of selected features. For sarcomatoid change and overall stage, using tumor sub-segmentations (core versus rim, high versus low enhancing) provided additional predictive value.

TL;DR: Eleven of twenty targets were significantly predicted by radiomics, with five exceeding AUC 0.8, and the new pipeline matched performance while improving interpretability.
Pages 8-10
Hierarchical Feature Group Selection Explained

The hierarchical Correlation-based Feature Reduction (CFR) first groups radiomic features by their semantic category (shape, first-order statistics, texture families such as GLCM, GLRLM, GLSZM). Within each group, redundant correlated features are removed, preserving at least one representative from each group.

The subsequent feature group selection layer then evaluates which feature groups contribute independently to prediction, removing entire groups that do not add information above what other groups already capture. This two-level pruning ensures that the final feature set spans multiple distinct image phenotypes.

The practical effect is that model outputs can be explained in terms of recognizable radiomic concepts, for example, 'the prediction is driven by tumor texture heterogeneity and rim enhancement pattern,' rather than arbitrary combinations of highly correlated features with no unified biological interpretation.

TL;DR: Hierarchical feature group selection works by pruning redundant features within groups and then selecting across groups, producing models explainable in recognizable radiomics terms.
Pages 10-12
Non-Invasive Prediction of Molecular Tumor Features

The ability to predict chromosomal events like Loss 9p21.3 and evolutionary metrics like linear EvoST from CT scans has significant implications for treatment planning. These molecular features are normally accessible only through genomic profiling of tumor biopsies, which captures only a fraction of tumor heterogeneity.

CT-based prediction of renal vein invasion and overall stage can directly inform surgical approach, as confirmed renal vein involvement changes the procedure complexity and intraoperative risk. Accurate preoperative radiomics-based assessment could improve surgical preparedness.

Necrosis and sarcomatoid differentiation, both predictable with AUC above 0.8, are known to correlate with aggressive tumor biology and poor prognosis. Non-invasive identification of these features before surgery could support more nuanced patient risk stratification and treatment selection.

TL;DR: Radiomics models can predict surgical and molecular features of ccRCC non-invasively from CT, enabling better preoperative planning and molecular characterization without biopsy.
Pages 12-14
Interpretability and Performance Can Be Achieved Together

This study demonstrates that improving radiomics model interpretability through hierarchical feature group selection does not require sacrificing predictive performance. The proposed pipeline matched or exceeded conventional approaches across 20 diverse targets.

The use of tumor sub-segmentations, particularly core versus rim and high versus low enhancing regions, added value for specific targets, suggesting that spatial heterogeneity within the tumor captures biological information not available from whole-tumor analysis alone.

Future work should expand to larger cohorts, incorporate multi-time-point CT imaging to capture tumor evolution directly, and prospectively validate whether radiomics-based molecular predictions translate into improved treatment selection and patient outcomes in ccRCC.

TL;DR: Feature group selection enables interpretable, high-performing radiomics models for ccRCC, with tumor sub-segmentation providing additional value for specific molecular and histological targets.
Citation: Open Access, 2023. Available at: PMC10424427.