PaperlayerView exact revision on arXiv

Source-linked AI summary

Physically Aware Radiomics Without Interpolation: Disentangling Voxel Geometry and Signal Modification in CT and MRI

David Corral Fontecha, Juan Miranda Bautista, Pablo Menendez Fernández-Miranda, Sergio Rubio-Martín, Lara Lloret Iglesias, Jose A. Vega

arXiv:2607.12399v1cs.CVcs.AI

TL;DR

Radiomic texture features can conflate anisotropic voxel geometry with interpolation-induced signal changes. The paper introduces voxel-spacing–aware extraction that incorporates physical geometry without resampling the native image signal. VS showed near-native agreement with NR across CT and MRI, while RS deviated more and MRI exhibited greater variability across strategies and classifiers.

  • Problem

    Radiomic texture computation often uses voxel-index neighborhoods, leaving the independent effects of voxel geometry and combined preprocessing operations insufficiently isolated, particularly for MRI.

  • Method

    The study modifies PyRadiomics to compare native extraction, isotropic resampling, voxel-spacing–aware extraction and fake-isotropic metadata control across CT and MRI.

  • Results

    VS preserved native radiomic behavior and maintained competitive predictive performance, whereas isotropic resampling altered texture structure and MRI showed greater preprocessing variability.

  • Takeaways & Limitations

    Voxel-spacing–aware extraction offers a physically coherent alternative to interpolation-based preprocessing while preserving the native image signal.

  • Takeaways & Limitations

    The implementation requires validation in other software environments, and MRI showed greater variability across preprocessing strategies and classifiers.

Abstract

from arXiv · show

Objective: Radiomic texture features are usually computed in voxel-index neighborhoods, implicitly assuming isotropic spatial relationships. In anisotropic images, this can confound voxel geometry with interpolation-induced signal changes. We developed a voxel-spacing-aware radiomic framework that incorporates physical geometry into texture computation without resampling. Approach: We modified PyRadiomics to account for voxel spacing while preserving the native image signal. Four configurations were compared: native non-resampled extraction (NR), isotropic resampling (RS), voxel-spacing-aware extraction (VS), and fake-isotropic preprocessing (FK), in which spacing metadata were overwritten without altering the image array. Experiments included 685 LIDC-IDRI pulmonary nodules and 209 I-SPY2 breast MRI cases, with 196 radiomic descriptors. Robustness was assessed using ICC, within-subject variability, Friedman testing, feature selection, machine learning, a multilayer perceptron, and external validation. Main results: VS showed near-native agreement with NR: median ICC(A,1) was 0.9976 in CT and 0.9984 in MRI. RS produced lower agreement and larger deviations, while FK showed intermediate behavior, confirming that spacing metadata alone can affect radiomic features. Gradient-derived and neighborhood-sensitive descriptors were most affected by preprocessing. VS preserved predictive performance comparable to NR in external CT validation, whereas MRI showed greater variability across preprocessing strategies and classifiers. Significance: Voxel-spacing-aware extraction separates geometric modeling from interpolation-induced signal modification while preserving the native image signal, offering a coherent alternative to isotropic resampling for radiomic analysis of anisotropic CT and MRI.

1. Introduction

Radiomic texture analysis can be confounded by anisotropic voxel geometry, while isotropic resampling corrects geometry by modifying image signal. The study evaluates voxel-spacing–aware extraction as a non-interpolating alternative across CT and MRI.

  • Motivation: Anisotropic voxel spacing alters the physical interpretation of neighborhood-dependent descriptors, including GLCM, GLRLM and related texture families.The same index displacement can represent different physical distances and directional connectivity across acquisitions.
  • Motivation: Isotropic resampling enforces geometric consistency but interpolation modifies signal, introduces artificial correlations and may suppress biologically relevant heterogeneity.
  • Motivation: Native extraction preserves the voxel array but implicitly treats neighborhoods as index-based rather than physical-space relationships.
  • Research gap: The study tests whether voxel-spacing–aware extraction generalizes beyond CT to MRI, where intensity is non-standardized and acquisition-dependent.
  • Study objective: The framework incorporates true physical voxel geometry into neighborhood computations without interpolating the native image signal and compares NR, RS, VS and FK.The controlled design evaluates robustness, variability, feature selection and predictive performance using statistical, machine-learning and deep-learning analyses.

2. Materials and Methods

The study analyzes public CT and MRI cohorts with consistent geometric preprocessing and a modified PyRadiomics implementation. The design preserves native data while evaluating physical-spacing-aware extraction alongside resampling and metadata controls.

  • Data preparation: Segmentations were reconstructed or mapped to reference geometry, manually reviewed, and corrected when necessary before analysis.
  • Preprocessing: The same four geometric configurations—NR, RS, VS and FK—were applied across modalities, while MRI analyses additionally used native intensity normalization.External validation included private CT and MRI cohorts of 46 and 63 cases, respectively.
  • Implementation: The modified PyRadiomics framework supports spacing-aware texture extraction while retaining standard behavior when the extension is disabled.The implementation spans the Python frontend, C wrapper layer and computational backend.
  • Radiomic features: A total of 196 descriptors were extracted from original and gradient-magnitude images, covering first-order, shape and GLCM, GLRLM, GLSZM and GLDM features.Shape features already operated in physical coordinates and required only minor consistency adjustments.

2.3. Physical-Space Neighborhood Modeling

The physical-space framework incorporates voxel dimensions into neighborhood-dependent texture calculations while preserving native voxel values and topology. Descriptor-specific implementations distinguish directional, neighborhood-averaged and finite-volume computations.

  • Physical-space modeling: Spacing-aware extraction incorporates physical voxel dimensions into texture computation without interpolation-induced intensity modification.The method preserves the native voxel array and expresses neighborhood relationships in physical space.
  • Descriptor-specific modeling: GLCM features use directional co-occurrence matrices followed by anisotropy-aware angular weighting, while NGTDM uses spacing-aware weighted neighborhood means.
  • Descriptor-specific modeling: GLRLM, GLDM and GLSZM use finite-volume isotropic subcells with zero-order hold, preserving original gray levels while measuring runs, dependencies and zones physically.No interpolation, smoothing or intensity mixing is performed.
  • Compatibility: VS is implemented as an optional physical-space extension to selected neighborhood-dependent computations rather than a change to standard IBSI feature definitions.
  • Geometric configurations: The comparison includes NR, RS, VS and FK, with FK overwriting spacing metadata while preserving the original image array to isolate metadata effects.

2.6. Gray-Level Discretization

The analysis fixes gray-level discretization across geometric configurations and evaluates preprocessing effects through paired robustness, variability, statistical, feature-selection and predictive analyses.

  • Gray-level discretization: Fixed bin-width discretization was applied consistently across configurations so observed differences primarily reflected voxel-geometry handling rather than discretization variability.Parameters were estimated separately for original and gradient-filtered images using approximately 50 gray-level bins.
  • Robustness analysis: ICC(A,1) measured numerical absolute agreement, whereas ICC(C,1) measured preservation of relative subject ranking across configurations.
  • Variability analysis: Within-subject variability was quantified using median coefficient of variation and relative median absolute deviation across preprocessing strategies.
  • Robustness analysis: Pairwise preprocessing comparisons used ICC(A,1), ICC(C,1), absolute differences and signed bias to assess preservation of native feature behavior.The comparisons specifically tested whether VS more closely matched NR than RS and whether FK functioned as a negative geometric control.
  • Feature selection: Feature-selection stability combined statistical and machine-learning selectors under 5-fold stratified cross-validation, with training-only preprocessing and selection to avoid leakage.Features receiving at least four selector votes and appearing in at least 40% of folds were labeled stable ensemble descriptors.

2.9. Machine Learning Modeling

The study evaluated how geometric preprocessing affects radiomic-based prediction using leakage-free, train-only modeling and frozen external validation across CT and MRI.

  • 2.9. Machine Learning Modeling: Train-only preprocessing, feature selection, and hyperparameter optimization were transferred as frozen transformations to held-out and independent external cohorts.Models included Logistic Regression, XGBoost, and an exploratory multilayer perceptron.
  • 2.9. Machine Learning Modeling: A common neutral signature enabled direct comparison of NR, RS, VS, and FK while reducing geometry-dependent feature-selection bias.The signature used training-only rankings shared across preprocessing configurations.
  • 2.9. Machine Learning Modeling: Ensemble feature selection retained descriptors selected by at least four independent methods, with models limited to 30 descriptors.Hyperparameters were optimized using 5-fold stratified cross-validation within the training set before final held-out testing.
  • 2.9. Machine Learning Modeling: External validation used independent datasets acquired under different clinical and technical conditions without model retraining or feature reselection.The original transformations, feature signatures, and trained parameters were transferred directly.
  • 2.9. Machine Learning Modeling: The workflow combined robustness statistics, ensemble selection, conventional machine learning, deep learning, and frozen-pipeline external validation.Figure 2 summarizes the multimodal processing and analysis pipeline.

3. Results

Across CT and MRI, voxel-spacing-aware extraction most closely preserved native radiomic features, while isotropic resampling produced larger deviations and geometry-sensitive descriptors showed the greatest variability. Predictive effects were modality- and classifier-dependent, with VS retaining strong CT external performance but no universal winner across settings.

  • Bin Width and Geometry-Dependent Discretization: NR and VS produced identical optimal bin widths, while RS reduced them and FK partially altered Gradient-derived discretization.Differences were more pronounced for Gradient-filtered than Original images.
  • Statistical Robustness Across Geometric Preprocessing: NR–VS comparisons showed the highest preservation, with 141/196 CT and 140/196 MRI features achieving ICC(A,1) ≥0.90.Features below the threshold were predominantly neighborhood-dependent descriptors, especially GLDM, GLRLM, and GLSZM.
  • Statistical Robustness Across Geometric Preprocessing: Isotropic resampling produced lower agreement and larger deviations, whereas FK showed intermediate behavior despite preserving the original voxel array.These findings indicate that spacing reinterpretation alone altered a subset of geometry-sensitive descriptors.
  • Statistical Robustness Across Geometric Preprocessing: Gradient-derived and neighborhood-dependent texture descriptors were less robust than global intensity features, with GLDM and GLSZM showing particularly strong geometry sensitivity.Gradient-filtered representations had lower ICC and increased wCV and rMAD across preprocessing configurations.
  • Ensemble Feature Selection: Stable ensemble signatures were largely comparable between NR and VS, although RS generated more recurrent descriptors without consistently improving held-out performance.RS changes were enriched in texture-derived and Gradient-based features and often showed larger train–test discrepancies.
  • Predictive Modeling Analysis: Across CT external validation, VS preserved predictive performance identically to NR and FK, while RS had lower discrimination and calibration, including ROC AUC 0.284 versus 0.590 for NR with XGBoost.This result was obtained under the common neutral signature setting across Logistic Regression, XGBoost, and MLP models.

4. Discussion

The discussion argues that voxel-spacing–aware extraction separates physical neighborhood modeling from interpolation while largely preserving native radiomic behavior. Across CT and MRI, VS most closely matched NR, whereas RS introduced larger deviations and MRI remained more variable across classifiers and preprocessing strategies.

  • Mechanisms: FK altered several geometry-sensitive descriptors despite preserving the voxel array, but did not reproduce the full behavior of RS.This indicates that spacing reinterpretation and interpolation-induced smoothing affect radiomic descriptors through partially distinct mechanisms.
  • Feature behavior: Gradient-derived and texture-derived descriptors were especially sensitive to preprocessing, and RS produced more recurrent feature-selection patterns without consistently improving prediction or external generalization.NR and VS often yielded sparser consensus signatures while preserving strong pairwise agreement, consistent with less interpolation-induced smoothing.
  • Predictive modeling: VS preserved CT predictive behavior almost identically to NR across Logistic Regression, XGBoost and MLP models, including internal and external validation.RS frequently showed lower external discrimination and calibration stability, while MRI effects depended more strongly on classifier architecture.
  • Cross-modality preservation: NR–VS comparisons consistently showed the strongest preservation across CT and MRI, while analyses involving isotropic resampling demonstrated larger deviations.This cross-modality hierarchy suggests that spatial-neighborhood modeling and physical voxel relationships, rather than modality-specific intensity characteristics, primarily drive the observed effects.
  • Interpretation: VS preserves native radiomic structure while modeling anisotropic spatial relationships without interpolation-induced signal modification.The framework is intended as a geometrically coherent alternative to isotropic resampling, not necessarily as a universal way to maximize predictive performance.
  • Limitations: The study remains limited by modified-PyRadiomics validation, greater MRI variability, restricted image representations, absent phantom or independent IBSI benchmarking, and exclusion of shape descriptors from primary analyses.Future work should test broader transformations, multicenter reproducibility, harmonization robustness, and deep-learning generalization.

5. Conclusion

Voxel-spacing–aware extraction accounts for physical voxel geometry without interpolating the original image signal. Across CT and MRI, it preserved native radiomic behavior and competitive predictive performance while offering a physically coherent alternative to isotropic resampling.

  • Voxel-spacing–aware extraction preserved native radiomic behavior while incorporating physical voxel geometry without interpolating the original image signal.
  • Across CT and MRI, VS maintained competitive predictive performance compared with interpolation-based preprocessing.
  • The framework supports more methodologically coherent, interpretable and reproducible radiomics for anisotropic medical imaging.

Funding

The study reports no specific funding.

  • No specific funding was received for this study.

Ethics Statement

The primary analyses used publicly available de-identified datasets, while external validation used retrospectively collected de-identified clinical imaging data under institutional ethics approval. Informed consent was waived because of the retrospective study design.

  • Primary analyses used publicly available de-identified datasets obtained from The Cancer Imaging Archive.
  • External validation cohorts were retrospectively collected under institutional ethics committee approval using de-identified clinical imaging data.
  • Informed consent was waived because of the retrospective nature of the study.

Code Availability Statement

The modified PyRadiomics implementation and analysis scripts were being prepared for public release, with availability upon reasonable request.

  • The modified PyRadiomics implementation and analysis scripts were being prepared for public release.
  • Code may be made available upon reasonable request.
Loading 2607.12399v1…