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

arXiv:2607.12399 · cs.CV, cs.AI · Submitted 2026-07-14 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Physically Aware Radiomics Without Interpolation".

Tom: This paper introduces a voxel-spacing–aware radiomic framework designed to address the challenges of voxel anisotropy in CT and MRI imaging.

Jane: First, who's behind it and why it matters.

Title and authors: Tom: So, what are the core findings here? Basically, they developed a voxel-spacing–aware extraction framework that incorporates true voxel dimensions into neighborhood definitions without needing to interpolate the original image signal.

Jane: They compared four different setups: native non-resampled extraction, isotropic resampling, their proposed voxel-spacing–aware extraction, and a fake-isotropic control to isolate geometric metadata effects.

Lu: The key result they found was that the voxel-spacing–aware extraction method showed near-native agreement with native non-resampled extraction in both CT and MRI, achieving median ICC(A,one) values of zero point nine nine seven six for CT and zero point nine nine eight four for MRI.

Meng: That level of agreement is impressive if you're not modifying the original image signal, which is what isotropic resampling does, according to their comparison results.

Lalam: It confirms that preserving the native image signal while accounting for physical spacing leads to more reliable quantitative imaging biomarker analysis.

Tom: And they also noted that while geometry sensitivity varies by feature family, gradient-derived and neighborhood-sensitive texture descriptors showed the most dependence on preprocessing, specifically within the GLDM and GLSZM families.

Jane: So, it’s not a magic fix for every single radiomic feature; some textures are more sensitive to these geometric considerations than others. That's an important caveat we have to keep in mind as we apply this knowledge.

Lu: The methodology involves using a modified PyRadiomics backend that handles direction-dependent descriptors like GLCM with anisotropy-aware angular weighting and uses anisotropic-aware weighted neighborhood averaging for NGTDM.

The paper's summary: Tom: The paper suggests several ways this concept can be expanded, moving beyond just the feature extraction part, and I’m talking about really cool architectural ideas for AI systems.

Jane: They propose three main areas for improvement: using physically aware three dee convolutional kernels that use voxel spacing to adjust their receptive field dynamically (improvement one), geometry-invariant feature extraction modules (improvement two), and spacing-conditioned neural architectures (improvement three).

Meng: The idea of physically aware convolutions sounds very practical; it means the AI can process thick-slice CT or MRI without introducing that signal smoothing you get from interpolation, allowing it to learn from true biological textures (improvement one).

Lu: And improvement two, integrating the VS backend into feature engineering within hybrid CNN-Radiomics models, would generate signatures that stay stable across different scanner protocols and acquisition thicknesses (improvement two).

Lalam: That stability is vital for deployment; if a model trained on one hospital's data can work reliably at another institution with different slice thicknesses, the impact on accessibility in healthcare is huge.

Tom: Plus, improvement three suggests using voxel geometry as a conditioning input through something like Feature-wise Linear Modulation to let the network learn to adapt its feature weighting based on how anisotropic the input is (improvement three).

Jane: That sounds like the AI itself becomes adaptive, learning when it needs to prioritize certain texture descriptors over others depending on the image quality characteristics (improvement three).

The paper's improvements: Tom: So, to wrap up this discussion on "Physically Aware Radiomics Without Interpolation: Disentangling Voxel Geometry and Signal Modification in CT and MRI," the main implication is that we can have a physically grounded alternative to isotropic resampling for quantitative biomarker analysis.

Jane: It’s about separating the geometric modeling from the signal modification introduced by interpolation, which supports using radiomics more reliably in anisotropic CT and MRI data.

Lu: The framework offers a way to incorporate true physical voxel geometry directly into feature computation without modifying the native image signal, which is a coherent alternative to isotropic resampling.

Meng: For practical implementation, the engineering challenge will be making sure these modified PyRadiomics backends are compatible with existing standard workflows while ensuring they handle the complexity of anisotropic data efficiently.

Lalam: This work gives us a solid foundation for building more robust AI tools in medical imaging, moving toward solutions that are less dependent on arbitrary preprocessing choices.

Tom: So, listeners, we’ve seen how this paper moves past the traditional resampling pitfalls by anchoring the analysis to the actual physical structure of the data with this voxel-spacing–aware extraction framework.

Jane: We should keep an eye on how these ideas evolve when they get integrated into those more complex neural network architectures discussed in that third set of improvements we just talked about.

Lu: The potential for creating geometry-invariant radiomic signatures across different acquisition settings is where the most exciting research lies moving forward.

Meng: From an engineering standpoint, seeing how this works in practice will dictate whether this becomes a standard feature in diagnostic pipelines or just an interesting academic concept.

Lalam: Ultimately, the impact is about making the AI tools we use for medical interpretation more consistent and trustworthy across different clinical settings.

Conclusion: Tom: So, we’ve really walked through how this paper on "Physically Aware Radiomics Without Interpolation: Disentangling Voxel Geometry and Signal Modification in CT and MRI" tackles the issue of voxel anisotropy by incorporating true physical spacing into feature extraction rather than just resampling the image.

Jane: Exactly, Tom, it shows us a way to get those high-quality radiomic features while keeping the original image signal intact, which is such a significant step for clinical analysis.

Lu: I think what’s really fascinating is how they handled the different descriptor families—the anisotropic-aware angular weighting for GLCM and the finite-volume zero-order-hold representation for others—it’s really creative how they adapted each mathematical structure to the physical space.

Meng: From an engineering standpoint, I'm really interested in how they implemented that modified PyRadiomics backend; making sure those neighborhood definitions actually respect the true physical dimensions without introducing interpolation artifacts is a tough implementation detail.

Lalam: For me, the most impactful vision here is that this approach could fundamentally improve medical culture by allowing researchers to trust texture features more deeply because they are grounded in physics rather than an arbitrary grid.

Tom: That’s a huge shift, Lu; moving from an index-based neighborhood to a physical-space relationship really lends credibility to what we measure.

Jane: It means less guesswork when we compare results between different scanners or even different acquisition protocols, which is something every radiologist needs.

Lu: And the implications for AI are huge; if these features are geometry-invariant, we can deploy models trained on one institution's data reliably at another without needing complex calibration steps.

Meng: I agree, and that stability would drastically reduce the pipeline complexity for deploying AI diagnostics in diverse clinical environments.

Lalam: It opens up possibilities for creating more universally applicable diagnostic tools that aren't tied to specific scanner hardware limitations or preprocessing scripts.

Tom: So, we’ve seen how this paper on "Physically Aware Radiomics Without Interpolation: Disentangling Voxel Geometry and Signal Modification in CT and MRI" provides a solid foundation for making radiomic analysis physically sound.

Jane: It really gives us a clear path forward for more trustworthy quantitative imaging biomarker development, building on the excellent comparison results they found between native extraction and isotropic resampling.

Lu: I’m genuinely excited about the potential to use these geometry-invariant signatures to build much more robust and generalizable AI models across different medical datasets.

Meng: And from an implementation standpoint, I see this as a major win for building more resilient and practical diagnostic AI systems that don't fail when the input data isn't perfectly uniform.

Lalam: This work could really improve the culture of medical research by providing a method that prioritizes physical reality in our quantitative analyses, leading to more consistent and reliable clinical interpretations.

Department of Radiology, Complejo Asistencial Universitario de León · Department of Radiology, Hospital Universitario Rey Juan Carlos · Health Research Institute of the Jiménez Díaz Foundation · Department of Physical Therapy, Occupational Therapy, Rehabilitation and Physical Medicine, Rey Juan Carlos University · Department of Morphology and Cell Biology, Universidad de Oviedo · Department of Electrical, Systems and Automation Engineering, Universidad de León · Advanced Computing and e-Science Group, IFCA-CSIC · Facultad de Ciencias de la Salud, Universidad Autónoma de Chile

cs.CV, cs.AI

Submitted: 2026-07-14

Updated: 2026-07-14

Comments: Manuscript under peer review

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: This paper introduces a voxel-spacing–aware radiomic framework designed to address the challenges of voxel anisotropy in CT and MRI imaging.

Key concepts

Voxel-spacing–aware extraction: This framework incorporates true voxel dimensions directly into neighborhood definitions when extracting radiomic features. It achieves this without needing to interpolate the original image signal, offering a way to account for physical spacing in anisotropic CT and MRI data.
Isotropic resampling: This is a method where the original image signal is modified by resampling it to be isotropic. The paper compares its results against their proposed method, showing that preserving the native signal while accounting for physical spacing leads to more reliable quantitative analysis.
Geometry-invariant feature extraction modules: This improvement involves integrating the voxel-spacing backend into feature engineering within hybrid CNN-Radiomics models. The goal is to generate feature signatures that remain stable across different scanner protocols and acquisition thicknesses.
Physically aware convolutions: This concept suggests using three-dimensional convolutional kernels that use voxel spacing to dynamically adjust their receptive field. This allows AI systems to process thick slices without introducing signal smoothing from interpolation, learning from true biological textures.

Terminology

Robustness — 18× in this paper · explained in 5 episodes
Anisotropy — 16× in this paper · explained in 3 episodes
External Validation — 14× in this paper · explained in 2 episodes
AUC — 11× in this paper · explained in 2 episodes
Feature Extraction — 9× in this paper · explained in 2 episodes
Feature Selection — 7× in this paper · explained in 4 episodes
Generalization — 6× in this paper · explained in 13 episodes
Calibration — 5× in this paper · explained in 7 episodes
Consistency — 4× in this paper · explained in 3 episodes
Interpretability — 4× in this paper · explained in 15 episodes
Regularization — 4× in this paper · explained in 3 episodes
Cancer Imaging Archive — 3× in this paper
Predictive Modeling — 3× in this paper · explained in 2 episodes
Generalizability — 2× in this paper · explained in 6 episodes
Pairwise Comparisons — 2× in this paper · explained in 2 episodes
Aggregation — 1× in this paper · explained in 1 episode
Benchmarking — 1× in this paper · explained in 3 episodes
Core — 1× in this paper · explained in 2 episodes
Decode — 1× in this paper · explained in 2 episodes
Diagnostic Tools — 1× in this paper · explained in 2 episodes
Feature Engineering — 1× in this paper · explained in 3 episodes
Image Biomarker Standardisation Initiative — 1× in this paper
Information Leakage — 1× in this paper · explained in 2 episodes
No Resampling — 1× in this paper
Segmentation Masks — 1× in this paper · explained in 2 episodes
Selection Bias — 1× in this paper · explained in 4 episodes

Summary

This paper introduces a voxel-spacing–aware radiomic framework designed to address the challenges of voxel anisotropy in CT and MRI imaging. By incorporating physical voxel geometry directly into feature computation without image resampling, the researchers aim to disentangle voxel geometry and signal modification, providing a method that avoids the pitfalls of traditional isotropic resampling while maintaining high predictive performance.

The limitations of current methods

In routine clinical imaging, voxel spacing is frequently anisotropic. The most widely adopted strategy to address this is isotropic resampling, typically performed by interpolating images to a fixed voxel spacing. However, this approach inherently modifies the original image signal, introduces artificial spatial correlations and may suppress biologically relevant heterogeneity. Conversely, direct extraction from native images preserves the original array but implicitly treats voxel neighborhoods as index-based rather than physical-space relationships, which can distort the physical interpretation of texture descriptors.

How it works

The researchers developed a modified PyRadiomics backend to support voxel-spacing–aware extraction directly in native image space. This framework incorporates true voxel dimensions into neighborhood definitions and distance-dependent computations without interpolation. To evaluate the effectiveness of this approach, the study compared four distinct geometric configurations:

Native non-resampled extraction (NR), which preserves the original voxel array but treats local neighborhoods in index space.
Isotropic resampling (RS), which enforces an isotropic voxel grid through interpolation.
Voxel-spacing–aware extraction (VS), which preserves the native voxel array while incorporating true physical spacing into computations.
Fake-isotropic preprocessing (FK), a control designed to isolate the influence of geometric metadata independently from interpolation by overwriting spacing metadata without modifying the image array.

Technical implementation

The implementation incorporates spacing information according to the mathematical structure of each descriptor family. For direction-dependent descriptors like GLCM, features are computed independently for each directional matrix and subsequently combined using anisotropy-aware angular weighting. For NGTDM, the framework utilizes anisotropy-aware weighted neighborhood averaging, allowing neighboring voxels to contribute according to their relative physical separation. To handle GLRLM, GLDM, and GLSZM descriptors, the system employs a finite-volume zero-order-hold representation of the native voxel grid, which allows run length and zone size to be expressed on an isotropic physical lattice while preserving original gray-level values.

Results and validation

The study demonstrated that VS showed near-native agreement with NR in both modalities, with median ICC(A,1) values of 0.9976 in CT and 0.9984 in MRI. In contrast, isotropic resampling (RS) produced lower agreement and larger feature deviations. Geometry sensitivity was found to be feature-family dependent, with gradient-derived and neighborhood-sensitive texture descriptors showing the greatest preprocessing dependence, specifically within the GLDM and GLSZM families. In external validation using independent private cohorts, VS preserved predictive performance comparable to NR in CT, whereas MRI showed greater variability across preprocessing strategies and classifiers. Ultimately, the framework provides a physically grounded alternative to isotropic resampling for quantitative imaging biomarker analysis in anisotropic medical imaging.

Improvements for AI systems

1. Implementation of Physically Aware 3D Convolutional Kernels

Improvement: Replace standard isotropic 3D convolutional layers with Physically Aware Convolutions that utilize voxel spacing metadata (x, y, z) to dynamically adjust the kernel's receptive field in physical space. Instead of assuming a uniform grid, the convolution operation will use a sampling strategy (similar to the paper's zero-order-hold subcell expansion) to ensure that a 3 times3 times3 kernel covers a consistent physical volume regardless of slice thickness.
Capability: The AI system can process anisotropic medical volumes (e.g., thick-slice CT or MRI) without the signal smoothing and artificial spatial correlation introduced by interpolation, allowing the model to learn from true biological textures rather than interpolation artifacts.

2. Geometry-Invariant Radiomics Feature Extraction Modules

Improvement: Integrate a Voxel-Spacing–Aware (VS) backend into the feature engineering pipeline of hybrid CNN-Radiomics models. This replaces index-based neighborhood definitions with physical-space modeling, specifically implementing anisotropy-aware angular weighting for GLCM/NGTDM and isotropic subcell expansion for GLRLM/GLDM/GLSZM.
Capability: The system can generate geometry-invariant radiomic signatures that remain stable across different scanner protocols and acquisition thicknesses. This enables seamless deployment of models trained on one institution's data to another institution with different slice thicknesses without the performance degradation or calibration drift caused by resampling.

3. Spacing-Conditioned Neural Architectures

Improvement: Incorporate voxel geometry as a conditioning input (via Feature-wise Linear Modulation or FiLM layers) within the neural network architecture. This allows the network to modulate its internal feature representations based on the physical dimensions of the input voxels.
Capability: The AI can automatically adapt its feature-weighting strategy; for example, it can learn to de-prioritize gradient-derived features when high anisotropy is detected (as these are shown to be highly sensitive) and prioritize first-order statistics or VS-extracted texture descriptors, improving robustness in heterogeneous clinical datasets.

4. Native-Signal Gray-Level Discretization Layers

Improvement: Replace standard post-interpolation discretization with a Native-Signal Discretization layer that calculates optimal bin widths based on the original, non-resampled intensity distribution of the image.
Capability: The system prevents the bin width shrinkage effect observed in interpolated images (where smoothing reduces local variance), ensuring that the quantized representation of tissue heterogeneity is mathematically consistent with the actual scanner output, leading to more stable training signals and better predictive accuracy for texture-based classifiers.