Standardised convolutional filtering for radiomics

TL;DR

This work delivers a complete reference manual from the IBSI for standardised convolutional filtering in radiomics, addressing how to define, apply, and report convolutional filters such as LoG, Laws, Gabor, wavelets, and Riesz transforms. It formalises the imaging workflow around padding, convolution, and aggregation, and it discusses geometric invariances, directionality, and spectral properties to enable robust feature extraction. Through benchmarking with digital phantoms and a lung CT dataset, the document establishes reference response maps and feature values to verify software compliance and reproducibility. The practical impact is a comprehensive, testable framework that improves interoperability, reproducibility, and transparency in radiomics analyses across modalities and software implementations.

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 complete version of a reference manual on the use of convolutional filters in radiomics and quantitative image analysis. 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 were found to be poorly reproducible. This reference manual provides definitions for convolutional filters, parameters that should be reported, reference feature values, and tests to verify software compliance with the reference standard.

Figures (69)

• Figure 1: Overall image processing scheme with image filtering (adapted from Zwanenburg et al.ZLV2017). After loading a medical image, the image data are optionally converted (e.g. SUV normalisation). The image may then undergo further post-processing (e.g. noise reduction, bias field correction). Subsequently, a segmentation mask is created or loaded to identify the ROI in the image. The image is subsequently interpolated to ensure that image voxels are isometric, after which image filters may be applied to the image. The ROI mask is interpolated to the same grid as the image, prior to forming intensity and morphological ROI masks. The intensity (but not morphological) mask is optionally re-segmented based on image intensities of the unfiltered image. Subsequently, features are computed from the filtered image and the applicable ROI masks. The intensities in ROI intensity mask may undergo discretisation prior to computing features from e.g. texture families. IH: intensity histogram; IVH: intensity-volume histogram; GLCM: grey level cooccurrence matrix; GLRLM: grey level run length matrix; GLSZM: grey level size zone matrix; NGTDM: neighbourhood grey tone difference matrix; NGLDM: Neighbouring grey level dependence matrix; GLDZM: grey level distance zone matrix; *A different discretisation scheme is usually used for computing IVH features.
• Figure 2: Overview of the filter-based image biomarker extraction process.
• Figure 3: A $16\times 16$ image $f[\boldsymbol{k}]$ (left) is filtered by $g_{ss}$ defined in Eq. \ref{['eq:separabilityExample']} using a separable convolution. The intermediate image $(g_{s}\ast f)[\boldsymbol{k}]$ is shown (centre), where the convolution is performed with $g_s$ aligned along the lines (i.e.$k_1$) of $f$. After convolving this intermediate image with $g_{s}^T$ (i.e. aligned along the columns $k_2$), the result (right) is equivalent to a 2$D$ convolution of $f$ and $g_{ss}$.
• Figure 4: $f[\boldsymbol{k}]$
• Figure 5: $g[\boldsymbol{k}]$