CBCT-IQ: A Publicly Available Annotated Cone-Beam CT Dataset for Image Quality Assessment and Benchmarking

Background Image quality assessment (IQA) is essential for evaluating cone-beam computed tomography (CBCT) systems, optimizing acquisition protocols, and validating image reconstruction and enhancement algorithms. Despite the increasing use of CBCT in image-guided interventions, radiotherapy, orthopaedics, and other clinical applications, there is currently no publicly available benchmark dataset that combines systematically acquired CBCT images with expert image quality annotations. Existing studies typically rely on proprietary datasets or objective IQA measures alone, limiting reproducibility and fair comparison between methods.The CBCT-IQ dataset addresses this limitation by providing a publicly available collection of CBCT images acquired under controlled imaging conditions together with expert reader assessments of image quality. In addition to manual annotations, the repository contains benchmark results for 28 established IQA measures, enabling direct comparison between human perception and objective image quality measures. Methods Two anthropomorphic phantoms representing the thorax and pelvis were scanned using a Siemens ARTIS pheno C-arm CBCT system. Imaging parameters were systematically varied to generate a broad spectrum of image quality levels while maintaining reproducible acquisition conditions. Three independent experts evaluated image quality using paired comparisons against reference images through the Speedy IQA software.Each selected image slice received two independent quality scores: Overall Image Quality (OIQ) Region of Interest (ROI) Image Quality using a four-level Likert scale ranging from very poor to very good. Expert annotations were subsequently analysed together with objective IQA measures using Spearman rank correlation to establish benchmark performance. Dataset Description Image Acquisition Protocols CBCT acquisitions were performed using the Siemens ARTIS pheno system installed in a research hybrid operating theatre. Starting from the standard clinical acquisition protocol, three imaging parameters were systematically varied: Tube voltage (60, 90 and 125 kV) Pulse width (3.2, 5.0, 6.4 and 8.0 ms) Reconstruction type (Normal, Smooth and Very Smooth) The default acquisition protocol (90 kV, 8.0 ms, Normal reconstruction) served as the reference acquisition for all full-reference image quality analyses.Overall, 36 CBCT volumes were reconstructed, from which 1,764 representative image slices were selected across the two anatomical phantoms. Images were converted to NIfTI format and normalized using the reference acquisition before annotation and quantitative analysis. Dataset Structure: The repository contains CBCT image data together with expert annotations and benchmark image quality measurements.The dataset includes: degraded.zip: Degraded CBCT images acquired under systematically varied acquisition settings saved as NIfTI files in the format ”./DATASETNAME/SLICE/KERNEL/KV/PULSEWIDTH/RECONSTRUCTIONMETHOD.nii.gz”. reference.zip: Reference CBCT images acquired using the default imaging protocol saved as NIfTI files in the format ”./DATASETNAME/SLICE/EE/090/8.0/NORMAL.nii.gz”. For each degraded image, the image with kernel EE, kV 90, pulse width 8.0 and reconstruction type NORMAL is selected as reference image. annotations.csv: Expert annotations for OIQ and ROI quality from three independent readers. Each row is formatted as:filename,OIQ_1,OIQ_2,OIQ_3,OIQ_MOS_ZSCORE,ROI_1,ROI_2,ROI_3,ROI_MOS_ZSCORE,OIQ_IQA-rating,ROI_IQA-rating where filename corresponds to the image file, and OIQ_1 to OIQ_3 to the overall image quality ratings, and ROI_1 to ROI_3 to the region of interest ratings, and OIQ_MOS_ZSCORE to the mean of the z-scored OIQ annotations, and ROI_MOS_ZSCORE to the mean of the z-scored ROI annotations, and OIQ_IQA-rating to the consensus IQA measure-based rating for OIQ, and ROI_IQA-rating to the consensus IQA measure-based rating for ROI. software_versions.txt: Contains the versions of the employed IQA measures. All images are stored in NIfTI format, while annotations and benchmark results are provided in CSV format to facilitate reproducible analysis. We normalized the images by computing the minimal (r) and maximal (R) pixel values over all reference slices of the volume, then clipping the pixel values of each image to R. Finally, for each image I, we applied the scaling (I - r)/(R - r). Findings Benchmark evaluation of 28 IQA measures showed that several full-reference measures achieved the highest correlation with expert perception.The study further demonstrated that reconstruction type had a substantially larger influence on perceived image quality than variations in tube voltage or pulse width. While expert observers were generally unable to distinguish subtle image quality differences arising from acquisition parameter changes alone, several objective IQA measures consistently detected these differences, highlighting their complementary role alongside subjective evaluation. Interpretation The CBCT-IQ dataset establishes the first openly available benchmark specifically designed for quantitative IQA in cone-beam CT using clinically relevant expert annotations. By combining systematically acquired CBCT images, standardized reader assessments, and benchmark objective IQA measures within a single resource, the dataset enables reproducible evaluation of existing and newly developed IQA methods.Beyond benchmarking, the dataset supports research on AI-based image quality prediction, optimization of acquisition protocols, reconstruction algorithm development, quality-aware image enhancement, and automated quality assurance for CBCT imaging. The inclusion of both subjective expert ratings and objective IQA measure values provides a valuable reference for investigating the relationship between human visual assessment and computational image quality measures. Funding This study was supported by NÖ FTI Grundlagenforschung project (Project number: GLF21-1-001). In addition, this study was funded by ACMIT COMET Module FFG project (FFG number: 879733, application number: 39955962). We also acknowledge funding from the Austrian Science Fund (FWF) through project T1307. Furthermore, this study was also supported by a University of Sydney Robinson Fellowship and an NHMRC Investigator Grant (APP ID: 1041194). We have decided to make the data set available to the research community under the Creative Commons Attribution 4.0 International license. If you use the data in your research, please make sure to cite the repository and the relevant publication.