[Paper Review] Standardised convolutional filtering for radiomics

Adrien Depeursinge, Vincent Andrearczyk+6 more|SERVAL (Université de Lausanne)|Jan 1, 2020

Radiomics and Machine Learning in Medical Imaging22 citations

TL;DR

This paper proposes a standardized framework for convolutional filtering in radiomics to ensure reproducibility and consistency across software implementations. It defines precise filter configurations, permanent identifiers, and reference values for 11 filter types across 11 test cases, enabling validation of radiomics software and facilitating translation of image biomarkers from research to clinical use.

ABSTRACT

The Image Biomarker Standardisation Initiative (IBSI) aims to improve reproducibility of radiomics studies by standardising the computational process of extracting image biomarkers (features) from images. We have previously established reference values for 169 commonly used features, created a standard radiomics image processing scheme, and developed reporting guidelines for radiomic studies. However, several aspects are not standardised. Here we present a preliminary version of a reference manual on the use of convolutional image filters in radiomics. Filters, such as wavelets or Laplacian of Gaussian filters, play an important part in emphasising specific image characteristics such as edges and blobs. Features derived from filter response maps have been found to be poorly reproducible. This reference manual forms the basis of ongoing work on standardising convolutional filters in radiomics, and will be updated as this work progresses.

Motivation & Objective

• Address the lack of reproducibility and standardization in radiomics by establishing consensus-based filtering protocols.
• Provide a unified framework for implementing convolutional filters used in radiomics to ensure consistent response map generation.
• Enable software validation through benchmark datasets and reference values for key image features.
• Facilitate clinical translation of radiomics by reducing variability in biomarker computation across platforms and research groups.
• Establish permanent identifiers for filters and image processing steps to improve traceability and reproducibility in radiomics studies.

Proposed method

• Define 11 standardized convolutional filters, including Gabor, wavelet, Riesz, and Simoncelli filters, with precise mathematical formulations and implementation parameters.
• Specify filter parameters such as scale (σ), frequency (λ), orientation, and spatial extent to ensure consistent implementation across software platforms.
• Implement boundary conditions (e.g., mirror) and alignment rules (e.g., for Riesz kernels) to reduce implementation variability.
• Introduce permanent identifiers (four-character codes) for filters and image processing steps to enable unambiguous referencing.
• Provide reference values for 11 key image features (e.g., coefficient of variation, energy, root mean square) computed using each filter on standardized test datasets.
• Validate filter implementations across multiple teams using a benchmarking workflow with defined test cases and tolerance thresholds.

Experimental results

Research questions

• RQ1 How can convolutional filtering in radiomics be standardized to ensure consistent and reproducible biomarker extraction across different software platforms?
• RQ2 What precise filter configurations and implementation rules are necessary to minimize variability in response map generation?
• RQ3 What reference values can be established for key image features across standardized filters to enable software validation?
• RQ4 How can permanent identifiers be used to improve traceability and reproducibility in radiomics research?
• RQ5 To what extent do different software implementations agree on feature values when using the same standardized filter settings?

Key findings

• The IBSI standardization framework provides 11 filter configurations with defined parameters, including Gabor, wavelet, Riesz, and Simoncelli filters, ensuring consistent implementation.
• Reference values for 11 image features—such as coefficient of variation, energy, and root mean square intensity—were computed with high consensus (strong or very strong) across multiple teams.
• For the coefficient of variation feature, 10 out of 11 filter configurations achieved strong or very strong consensus, with tolerance levels ranging from 0.001 to 0.925.
• The root mean square intensity feature showed strong consensus (tolerance ≤ 7) for 10 out of 11 filters, with values ranging from 43.0 to 236.
• The energy feature achieved strong consensus for 9 out of 11 filters, with values ranging from 7.03×10⁷ to 1.97×10⁹, and tolerance levels below 1.4×10⁸.
• Filters 8.A and 8.B showed moderate consensus for certain features, indicating potential implementation sensitivity, while filters 10.A, 10.B, 11.A, and 11.B had no reference values due to missing data or low consensus.