RadRepro CBCT: An Open-Access CBCT Phantom Dataset for Improved Standardization and Reproducibility of Radiomics Research

Background

Radiomics, the process of extracting high-dimensional quantitative features from medical images, has demonstrated significant promise in enhancing precision diagnosis, prognosis, and personalized treatment planning. Despite its potential, one of the most critical challenges facing the field is the limited reproducibility of radiomics features. This variability is largely attributed to inconsistencies in imaging devices, acquisition protocols, and image reconstruction techniques.

Methods

This study presents the first open-access CBCT phantom dataset specifically curated to address reproducibility challenges in on-board imaging systems integrated with C-arm linear accelerators. CBCT images were acquired using a widely adopted Catphan phantom across multiple clinical devices from different vendors. Imaging parameters were systematically varied, including tube current (mAs), slice thickness, and reconstruction filters, to simulate realistic variability encountered in clinical workflows. The dataset is structured to allow researchers to evaluate and compare the impact of these variations on radiomics feature robustness.

Image Acquisition Protocols:  CBCT acquisitions of the Catphan phantom were performed on different CBCT scanners:

1) The Elekta X-ray volume imager (XVI, Version 5.0) was used in this study which is an on-board CBCT imaging system on Elekta radiotherapy linacs. It features a flat-panel image detector and the detector-source distance was 153.6cm. The image receptor size was 41×41cm2 and provided an image size of 1024×1024 pixels, however, clinical protocols usually used 512 pixels. The system used a 16-bit pixel grayscale resolution. The total filtration for the X-ray tube was 7.0 mm Al equivalence at 100 kV.  A bow-tie filter was used for all scans on Elekta systems in this study. Three similar Elekta CBCT devices were used in this study. Two systems featured the software version 5.04 and one system version 5.07.

2- Varian radiotherapy system (TrueBeam Version 4.1) was also used in this study which is an  on-board kV imager for CBCT imaging. The system includes a flat-panel image detector (PaxScan4030CB) of 30 x 40 cm² with a source-to-imager distance of 100 cm at isocenter and a source-to-axis distance of 100 cm. The field size ranges from 2 × 2 cm to 50 × 50 cm, with a default image matrix size of 512 × 512 pixels. The X-ray tube has an inherent filtration of 2.7 mm Al at 75 kV. Additionally, the kV beam delivery system utilizes two types of filters: a beam-hardening foil filter (0.89 mm thick titanium) to optimize the X-ray energy spectrum and bowtie filters (Full-Fan for head scans, Half-Fan for body scans) to enhance CBCT projection quality.

The selection of these three scanners was conducted by the following objectives:

•    Intra-vendor comparison: The inclusion of three Elekta XVI scanners allowed for the acquisition of identical measurements, enabling a direct comparison of radiomics features across different imaging devices from the same vendor.
•    Inter-vendor comparison: When performing measurements on scanners from different vendors, such as Elekta and Varian, achieving identical imaging parameters was not feasible due to each vendor’s unique predefined protocols and imaging settings. However, every effort was made to match acquisition and imaging parameters as closely as possible. The aim was to provide an open-access dataset that allows researchers to assess the reproducibility of radiomics features across different vendors when imaging parameters vary. For instance, one can assess the stability of radiomics features by varying the exposure (mAs) on the Elekta system (within a possible range for predefined protocols of the imaging vendor) and comparing the reproducibility level of these features rather than their absolute values  with those obtained from the Varian system. This analysis provides valuable insights into how different image acquisition settings at different CBCT imaging vendors impact radiomics feature consistency.

Various imaging setup were conducted, including varying slice thickness, exposure (mAs values), and the use of different pre-filters before reconstruction. Additionally, a test-retest protocol was implemented on each scanner to evaluate intrascanner repeatability. Furthermore, the phantom was deliberately shifted in different directions to assess the impact of positioning on the extraction of radiomics features in CBCT imaging.

Findings

The dataset comprises 120 CBCT image volumes collected from diverse imaging platforms, accompanied by corresponding region of interest (ROI) segmentations and radiomics features. This allows for extensive intra- and inter-vendor assessments of radiomics feature stability. The controlled setup and standardized segmentation provide a unique opportunity to benchmark radiomics workflows and test harmonization strategies.

Interpretation

By making this dataset publicly available, the study aims to support the radiomics research community in developing more reliable and reproducible analysis pipelines. It contributes to the ongoing efforts to standardize CBCT-based radiomics, enabling the generation of robust and generalizable models for clinical decision support in radiotherapy and beyond.