Unraveling Radiomics Complexity: Strategies for Optimal Simplicity in Predictive Modeling

TL;DR

This study tackles the challenge of radiomics' high dimensionality by introducing a principled framework to identify the smallest predictive feature subset for a given clinical problem. It defines radiomics complexity levels, builds an open-source MEDimage tool, and uses a rigorously controlled workflow (FDA feature reduction and XGBoost modeling) across five cancer cohorts to determine context-specific optimal feature types. Key findings show that morphological, intensity, or texture features dominate predictive performance depending on modality and endpoint, with HU-range tuning further boosting results. The approach offers a practical path toward simplified, interpretable radiomics with potential for broader generalizability and clinical translation.

Abstract

Background: The high dimensionality of radiomic feature sets, the variability in radiomic feature types and potentially high computational requirements all underscore the need for an effective method to identify the smallest set of predictive features for a given clinical problem. Purpose: Develop a methodology and tools to identify and explain the smallest set of predictive radiomic features. Materials and Methods: 89,714 radiomic features were extracted from five cancer datasets: low-grade glioma, meningioma, non-small cell lung cancer (NSCLC), and two renal cell carcinoma cohorts (n=2104). Features were categorized by computational complexity into morphological, intensity, texture, linear filters, and nonlinear filters. Models were trained and evaluated on each complexity level using the area under the curve (AUC). The most informative features were identified, and their importance was explained. The optimal complexity level and associated most informative features were identified using systematic statistical significance analyses and a false discovery avoidance procedure, respectively. Their predictive importance was explained using a novel tree-based method. Results: MEDimage, a new open-source tool, was developed to facilitate radiomic studies. Morphological features were optimal for MRI-based meningioma (AUC: 0.65) and low-grade glioma (AUC: 0.68). Intensity features were optimal for CECT-based renal cell carcinoma (AUC: 0.82) and CT-based NSCLC (AUC: 0.76). Texture features were optimal for MRI-based renal cell carcinoma (AUC: 0.72). Tuning the Hounsfield unit range improved results for CECT-based renal cell carcinoma (AUC: 0.86). Conclusion: Our proposed methodology and software can estimate the optimal radiomics complexity level for specific medical outcomes, potentially simplifying the use of radiomics in predictive modeling across various contexts.

Figures (12)

• Figure 1: Overview of the study workflow. The workflow in a typical radiomics analysis starts with acquisition and reconstruction of medical images. Subsequently, images are segmented to define regions of interest (ROIs). Following this step, the proposed radiomics software processes the images and computes features characterizing the ROIs, which are then organized by complexity levels for model training. Machine learning begins with feature cleaning to remove or replace invariant features, followed by feature set reduction to retain features exhibiting a high and stable correlation with the clinical endpoint, while removing inter-correlated features. All models are constructed using XGBoost. The final step involves results analysis through two stages: identification of the optimal complexity level, characterized by: the minimum number of features; minimum complexity; the highest and statistically significant performance, and explanation based on feature importance. Experiments can be conducted via programming or through the interface, with the code generation option facilitating the shift between the two approaches.
• Figure 2: Results overview per optimal complexity level and per cohort. Levels are identified as blue (morphological), green (intensity), and red (texture). A. Comparison of heatmaps showing aggregated AUC across splits for various radiomics signatures. Each column indicates a complexity level. The numerical suffixes in column names indicate the number of features retained after feature set reduction. In the bottom rows, green lines indicate significant improvements ($\text{p-value}<.05$) and red lines indicate non-significant improvements. Green boxes point to the identified optimal levels and the red one is selected for a further discussion in supplementary material. B: Explanation section displaying histogram of feature importance or importance tree, highlighting features with high importance at the identified optimal level.
• Figure 3: CT images of low-grade (right) and high-grade (left) NSCLC showing the application of cluster shade texture analysis. Images were selected based on the highest difference in high dependence low grey level emphasis (hdlge) feature (376.98 for high-grade and 0.02 for low-grade cancer), which had the highest importance after training. A contrast difference is visible between low and high-grade cancers (top). Overlay of the cluster shade revealed intratumoral heterogeneity (bottom).
• Figure 4: (A) Axial contrast enhanced computed tomography (CECT) comparison of a clear cell versus non-clear cell renal carcinoma with the highest disparity in median intensity. (B) Different views of 3D heatmap illustrating the influence of Hounsfield Unit (HU) range $[X,Y]$ on machine learning results in clear cell and non-clear cell renal carcinoma classification. Range limits highlighted in red denote the pre-optimization limits, while those in green signify the updated limits post-optimization.
• Figure 6: Illustration of multiple machine learning pipelines in the MEDimage interface.