Radiomic Feature Robustness Across Diverse Imaging Scanners

22 concepts21 papers1 proposal

Radiomics operates on a compelling premise: medical images are more than just pictures; they are multi-dimensional arrays of quantitative data. By extracting hundreds of sub-visual features—such as spatial patterns of voxel intensities, shape descriptors, and complex textures—machine learning models can uncover distinct biological signatures. These models have demonstrated remarkable success in predicting molecular subtypes in pancreatic cancer (Georgios Kaissis et al.), forecasting survival outcomes in gastric cancer (Yuming Jiang et al.), and identifying invisible tumor infiltration in glioblastoma (the authors of A Neural Network Approach to Identify the Peritumoral Invasive Areas in Glioblastoma Patients by Using MR Radiomics).

Yet, a profound paradox lies at the heart of radiomics. A model that achieves near-perfect diagnostic accuracy on a dataset from one hospital frequently fails when deployed at another. This drop in performance is rarely due to flaws in the machine learning architecture itself. Instead, it stems from inter-scanner and protocol variability—a technical domain shift where the underlying mathematical features change not because of biology, but because of the hardware and software used to capture the image.

The Root of Instability: Why Features Shift

To understand why radiomic features are fragile, one must look at how they are calculated. Consider a texture feature designed to measure tumor heterogeneity, such as a Gray-Level Co-occurrence Matrix (GLCM). This feature calculates how often pairs of pixels with specific intensities occur next to each other.

In an ideal world, a specific tissue type would always yield the same voxel intensities. In reality, medical images are heavily processed reconstructions. Different scanner manufacturers (e.g., GE, Siemens, Philips) utilize proprietary reconstruction algorithms, varying slice thicknesses, and distinct hardware specifications. As Simon K. B. Spohn et al. observe, quantitative image features are highly sensitive to these reconstruction variations, causing models optimized for one scanner brand to fail on another.

[Biological Tissue] 
       │
       ├─► [Scanner A: Smooth Kernel] ──► Low Noise Voxel Grid  ──► Low Heterogeneity Score
       │
       └─► [Scanner B: Sharp Kernel]  ──► High Noise Voxel Grid ──► High Heterogeneity Score

When a scanner applies a sharp reconstruction kernel, it introduces high-frequency pixel-level noise. A human radiologist easily ignores this noise, but a radiomics algorithm interprets it as increased tumor heterogeneity. Consequently, as Sheng Zhao et al. point out, models trained on single-center data lack the robustness required for multi-vendor clinical environments because they inadvertently learn to classify the technical artifacts of specific scanners rather than the true pathology.

This sensitivity creates significant "batch effects" across institutions (Xinchao Zhang et al.). Beyond hardware, minor alterations in imaging protocols—such as the timing or presence of contrast media—can fundamentally alter high-dimensional feature extractions, as noted by M. Zysman et al.

The Generalizability Bottleneck

The clinical consequence of this feature instability is a severe generalizability bottleneck. A wide array of highly accurate predictive models face this hurdle:

  • Diagnostic Models: Automated pipelines differentiating complex conditions, such as distinguishing COVID-19 from other pneumonias (Xiaofeng Chen et al.) or detecting small airway dysfunction (Zhenyu Shu et al.), struggle when moving away from standardized single-center protocols.
  • Prognostic Models: Algorithms designed to predict chemotherapy response or tumor shrinkage patterns in breast cancer (Ming Fan et al.; Wei Shi et al.) or therapy response in multiple myeloma (Fabian Bauer et al.) experience performance degradation on external datasets.

To combat this, a major focus of current research is the mandate for multi-center external validation. As argued by Jiang Shao et al., Jun Xu et al., Yongsheng Tang et al., and Lu Yu et al., testing models on diverse datasets produced by independent institutions is the only reliable way to expose and eliminate site-specific technical biases. Furthermore, a scoping review by Ariadne L van der Velden et al. emphasizes that standardizing image extraction workflows or engineering entirely hardware-agnostic features is essential for these tools to provide reliable clinical decision support.

Engineering Robustness: Emerging Solutions

To move past the limitations of static pixel data, researchers are developing methods that are inherently resilient to cross-scanner variability.

1. Shifting from Intensities to Physical Deformation

One elegant solution is to capture macro-level biological properties that scanner noise cannot easily obscure. Marwa Ismail et al. introduced the r-DepTH descriptor, which measures how a tumor physically deforms surrounding healthy brain tissue alongside traditional texture analysis. Because physical tissue displacement is a structural reality rather than a pixel-intensity artifact, this descriptor remains stable across different imaging centers, capturing fundamental biological properties instead of site-specific technical variations.

2. Dynamic and Temporal Feature Extraction

Another approach shifts the paradigm from static images to time-series data. Single-time-point images are highly sensitive to variations in scanner resolution and hardware settings. To circumvent this, Hui Qu et al. proposed dynamic radiomics, a framework that extracts quantitative features tracking how a tumor evolves over time or reacts to contrast agents. By focusing on the relative change of intensities across a time series (e.g., contrast wash-in and wash-out curves), the model gains a layer of mathematical normalization, making it significantly more robust against baseline hardware discrepancies.

Static Radiomics:  [Single Snapshot] ──────────────► Highly Sensitive to Scanner Calibration

Dynamic Radiomics: [Snapshot T1] ┐
                   [Snapshot T2] ┼─► Delta (Δ) ────► Robust, Normalized Temporal Signature
                   [Snapshot T3] ┘

3. Hybrid and Deep Learning Signatures

Researchers are also leveraging advanced modeling to bypass raw pixel fragility. For instance, Xiaofeng Chen et al. combined human-interpreted radiological "semantics" with routine clinical data, ensuring high multi-center reliability by anchoring the model in stable clinical observations.

Concurrently, deep learning approaches, such as the 3D convolutional neural networks utilized by Ahmad Chaddad et al., are being deployed to extract deep radiomic signatures linked directly to stable biological markers (like immune cell expressions). However, as the authors note, even deep learning features must undergo rigorous validation across diverse scanner manufacturers to ensure that neural networks are truly learning invariant biological truths rather than complex, hidden domain shifts.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_inter_scanner_variability_1d37b493,
  title = {inter-scanner variability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inter-scanner-variability/rdr_1d37b493009d4ceaa6889f97e9e0ad8f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_inter_scanner_variability_1d37b493_v1,
  title = {inter-scanner variability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inter-scanner-variability/rdr_1d37b493009d4ceaa6889f97e9e0ad8f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake