Measurement Phantoms in Medical Imaging

• Measurement phantoms are engineered or simulated reference objects designed to calibrate imaging systems and validate reconstruction algorithms.
• They span physical, anthropomorphic, computational, and stochastic types, each tailored with specific materials and model assumptions to mimic real-world conditions.
• Phantoms enable precise quantification of metrics like TTF, NPS, DSC, and dose errors, supporting virtual trials, algorithm training, and standardization in imaging research.

Searching arXiv for the cited phantom papers to ground the article in the referenced literature. arXiv search query: "all:measurement phantoms CT Mercury Phantom TTF NPS" Measurement phantoms are used in contemporary imaging and sensing research as controlled reference objects for quantifying system performance, validating reconstruction and analysis methods, and establishing reproducible test conditions. The literature spans engineered physical test objects, anthropomorphic phantoms, individualized voxelized “digital-twin” phantoms, analytical phantoms with exact forward models, and stochastic numerical phantoms for virtual imaging trials. Across CT, ultrasound, photoacoustics, optoacoustics, quantitative phase tomography, EEG and transcranial electric stimulation, radiation therapy, and dynamic tomography, their defining function is to provide known geometry, composition, or generative structure against which resolution, noise, dose, oxygenation, morphology, conductivity, or motion can be assessed (Fricks et al., 2020, Fu et al., 2020, Ziemczonok et al., 2024, Li et al., 2021).

1. Classes of measurement phantoms

The literature distinguishes several recurring phantom classes. Engineered physical phantoms are built from materials selected to mimic specific radiological, acoustic, optical, electrical, or mechanical properties. Anthropomorphic phantoms embed these surrogates in anatomically realistic geometry, often derived from clinical CT or MRI and realized through molding or 3D printing. Computational phantoms and digital twins assemble voxelized structures from segmented patient data or matched templates. Stochastic numerical phantoms generate ensembles with controlled anatomical and property variability for task-based assessment and virtual imaging. Analytical phantoms occupy a distinct category in which the forward model, such as the Radon transform, is available in closed form (Gallas et al., 2014, Fu et al., 2020, Park et al., 30 Sep 2025, Dessole et al., 2023).

Category	Representative example	Primary purpose
Physical CT phantom	Mercury Phantom	TTF and NPS estimation
Patient-specific computational phantom	iPhantom	CT organ dosimetry
Tissue-mimicking ultrasound phantom	Gel wax phantom	Uniformity, geometric accuracy, target sizing
Photoacoustic validation phantom	Blood-oxygenation flow phantom	sO$_2$ validation with online reference sensing
Anthropomorphic optical phantom	Forearm oximetry phantom	HSI and PAT sO$_2$ validation
Stochastic numerical phantom	3-D USCT breast phantom / breast qOAT NBP	Virtual imaging trials

This classification is technically important because the choice of phantom determines which uncertainties are controlled and which remain externalized. A simple test pattern can isolate a single forward-model assumption, whereas an anthropomorphic or stochastic phantom can expose interactions among anatomy, contrast mechanisms, reconstruction error, and observer variability. The cited work therefore uses phantoms not merely as calibration objects but as experimental infrastructures for system design, algorithm development, and protocol validation (Fricks et al., 2020, Gröhl et al., 2023, Xu et al., 2024).

2. Physical phantoms and surrogate materials

Physical measurement phantoms are typically designed around modality-specific surrogate materials and explicit exclusion criteria. In CT quality assurance, the Mercury Phantom V3.01 is a cylindrical phantom with outer diameter approximately 200 mm, mounted on a removable support ring. It contains cylindrical inserts of varying diameter from 3 mm to 25 mm for TTF estimation, a uniform polymethyl methacrylate region for NPS estimation, tapered transition sections, and air gaps and thin-wall regions that must be excluded because they introduce partial-volume artifacts. Under helical acquisition with 5 mm slice thickness and a 400 mm field-of-view, about 114 contiguous axial slices are reconstructed, and only canonical TTF and NPS slices are suitable for analysis (Fricks et al., 2020).

Ultrasound and photoacoustic phantoms emphasize acoustic and optical mimicry. The gel-wax ultrasound phantom embeds nylon filaments and a stainless-steel disc in gel wax and was used to evaluate image uniformity, geometric accuracy, and disc diameter. Its reported acoustic properties are $c_{us}=1418\pm31$ m/s, $\rho=0.87\pm0.02$ g/cm$^3$, $Z=1.23\pm0.05$ MRayl, and attenuation coefficient $0.88\pm0.05$ dB/cm/MHz; the maximum error in ultrasound distance measurement was $7.14\%$, and phantom volume decreased by about $1.83\%$ over 62 weeks (Phani et al., 2022). The blood-oxygenation photoacoustic phantom uses a 19 mm diameter agar cylinder containing an embedded PVC flow tube and integrates online absorption spectrometry and pre- and post-phantom fluorescence-quenching pO$_2$ probes, enabling dynamic validation from $_2$0 to $_2$1 sO$_2$2 (Gehrung et al., 2019). Anthropomorphic forearm phantoms for multispectral optical imaging were molded from MRI-derived 3D-printed forms using a stable copolymer-in-oil material, with vessel-like structures at five distinct sO$_2$3 levels between 0 and 100% (Dreher et al., 29 Mar 2025).

Other physical phantoms target anatomical realism, electrical conduction, motion, and mechanics. An anthropomorphic multimodality head phantom for proton therapy was fabricated from patient CT, 3D printed in epoxy resin, and filled with a dipotassium phosphate-based cranial bone surrogate, agarose gel, distilled water, and normoxic dosimetric gel; CT measurements yielded Hounsfield units agreeing with reference values, while the bone surrogate produced undesirably high MRI signal intensity (Gallas et al., 2014). For EEG and transcranial electric stimulation, fired clay was proposed as a skull material because it is formable, ion-permeable, and tunable in conductivity; stabilized conductivities spanned 0.191 S/m down to 0.024 S/m depending on clay type and firing temperature, with reproducible reuse after deionized-water rinsing (Hunold et al., 2018). Mechanically powered motion phantoms based on a printed mainspring or a LEGO elastic-band drive produced rotary motion suitable for MRI and ultrasound, with derived velocities reported as consistent and reproducible within a small error (Gomez et al., 2019). A cystic lung phantom made from NIST-recommended polymer foam demonstrated that individual cyst volumes were increasingly underestimated as cyst size decreased, reaching about $_2$4 to $_2$5 for 5 mm cavities (Shah et al., 25 Mar 2025). Polyurethane-gel surface phantoms were used to compare impact-based analysis with macroindentation, suction, damped oscillation, and durometry for superficial mechanical characterization (Bouffandeau et al., 13 Nov 2025).

3. Computational, individualized, and stochastic phantoms

Computational measurement phantoms extend the same calibration logic into patient-specific and ensemble-based settings. The iPhantom framework constructs full-body CT digital twins directly from patient images in three steps: segmentation of 22 high-contrast anchor organs with a 3D U-Net, parameterized template matching using trunk height and effective diameter $_2$6, and symmetric diffeomorphic registration to transfer low-contrast non-anchor organs from a matched XCAT template. The resulting phantom contains 34 structures and supports GPU-accelerated Monte Carlo organ dosimetry; reported results include anchor-organ DSC greater than 0.60 for all anchor organs, DSC of about 0.30 to 0.90 for non-anchor organs after diffeomorphic deformation, and organ-dose errors below 10% for most organs under strong tube-current modulation (Fu et al., 2020).

Stochastic numerical phantoms are designed for virtual imaging trials rather than one-off calibration. For ultrasound computed tomography, a 3-D stochastic breast-phantom pipeline adapts the FDA-VICTRE generator, prunes to USCT-visible tissues, inserts lesions, assigns tissue-class acoustic properties by truncated normal or normal sampling, and adds Gaussian random field textures in fat and glandular tissues. The framework was released with open-source phantom-generation code and 52 sets of simulated USCT measurement data to enable reconstruction development and task-based image-quality assessment (Li et al., 2021). For quantitative optoacoustic tomography of the breast, stochastic numerical breast phantoms include adipose, glandular parenchyma, ligaments, ducts, TDLUs, arteries, veins, nipple, muscle, a two-layer skin model, and lesion classes including viable tumor cell region, necrotic core, peripheral angiogenesis region, fibroadenomas, and cysts. The associated virtual imaging framework models optical absorption, scattering, acoustic propagation, and transducer impulse responses, and a public library of 1,020 phantoms and associated data was released for system evaluation (Park et al., 30 Sep 2025).

A related line of work addresses transcranial photoacoustic computed tomography. Stochastic numerical head phantoms are created from adjunct CT by segmenting skull plates, diploë, scalp, and cortical region, then adding vasculature synthesized by constrained constructive optimization. Optical properties are assigned at 690, 830, and 1064 nm, and acoustic-elastic properties are assigned plate-wise and layer-wise. In a case study, reconstructions using the exact heterogeneous skull model were nearly artifact-free, whereas homogeneous-skull modeling produced strong artifacts that overshadowed cortical signals, increased RMSE by about fourfold, and reduced CNR below 10 dB (Huang et al., 10 Oct 2025).

This body of work suggests that the term “measurement phantom” now includes not only fabricated objects but also voxelized and stochastic object models whose realism is sufficient to drive end-to-end simulation, observer studies, and design optimization. A plausible implication is that phantom realism is increasingly defined by the fidelity of both anatomy and forward physics rather than by geometry alone.

4. Quantitative models, metrics, and ground truth

Measurement phantoms are tightly coupled to explicit quantitative targets. In CT system characterization, the task transfer function is defined by

$_2$7

and the noise power spectrum is the two-dimensional Fourier transform of the noise autocovariance,

$_2$8

The Mercury Phantom is explicitly organized around the measurement assumptions required by these two metrics, including straight cylindrical edges for TTF and homogeneous regions for NPS (Fricks et al., 2020).

In individualized computational phantoms, overlap and dose metrics become central. iPhantom uses the Dice Similarity Coefficient,

$_2$9

and organ dose error

$c_{us}=1418\pm31$0

Its registration model minimizes an energy that combines image mismatch with a smoothness penalty, thereby tying anatomical plausibility directly to downstream dosimetric accuracy (Fu et al., 2020).

In photoacoustic imaging, phantoms are designed around the forward relation

$c_{us}=1418\pm31$1

which links initial pressure to optical absorption and fluence. The blood-oxygenation phantom couples this model to linear spectral unmixing and to pO$c_{us}=1418\pm31$2-to-sO$c_{us}=1418\pm31$3 conversion via the Severinghaus equation, while also implementing an exponential correction for spectral coloring by nigrosin in the surrounding agar (Gehrung et al., 2019). In quantitative phase tomography, tailored microphantoms provide known refractive-index distributions and enable dry-mass validation with

$c_{us}=1418\pm31$4

together with RMSE and morphological metrics such as Dice coefficient or Hausdorff distance (Ziemczonok et al., 2024).

Analytical phantoms serve a different but related role. For convex objects such as ellipses, squares, and rectangles with constant attenuation, the Radon transform can be computed exactly; the resulting exact sinograms eliminate discretization error at the sinogram stage and isolate the error introduced by reconstruction algorithms. The reported relative $c_{us}=1418\pm31$5 errors comparing analytical and numerical sinogram pipelines are of the same order but differ by phantom class, illustrating how even “simple” phantoms can expose reconstruction-dependent bias (Dessole et al., 2023).

5. Automation, learning, and virtual experimentation

A major recent development is the use of measurement phantoms as supervised training infrastructure. In CT QA, manual slice annotation in the Mercury Phantom requires labeling each slice as Outside Phantom, NPS pattern, TTF pattern, Tapered Section, or Unsuitable air-gap/partial-volume. Fricks et al. adapted VGG19 with ImageNet pretraining to five-class classification using three-channel axial context formed from previous, current, and next slices, together with runtime augmentation implemented via Albumentations. With $c_{us}=1418\pm31$6 and full augmentation, the model achieved approximately 98% validation accuracy, 97% test accuracy, and 86% extreme-test accuracy; random initialization, triplicate input without axial context, and omission of augmentation all degraded performance, particularly on misaligned data (Fricks et al., 2020).

Photoacoustic imaging provides an especially clear example of the phantom-to-learning pipeline. A collection of 102 well-characterized tissue-mimicking phantoms and their digital twins enabled supervised training of a U-Net for pixel-wise recovery of optical absorption coefficients from experimental photoacoustic data. The simulated and experimental slices were paired through manual segmentation, optical-property assignment from double-integrating sphere measurements, 3D Monte Carlo light transport, 2D k-Wave acoustic propagation, and common reconstruction. On held-out phantoms, the experimentally trained model produced smoother and sharper spatial maps than the simulation-trained model and outperformed it by about 8 percentage points in relative error for inclusions, while Monte Carlo fluence correction from reference optical properties yielded quantification error of approximately 20% (Gröhl et al., 2023).

Phantom-driven learning also appears in stochastic object modeling. AmbientCycleGAN translates a known mathematical phantom domain to a realistic stochastic object model using noisy measurement data, combining adversarial losses in both domains with cycle-consistency and identity losses. In the reported studies, the method markedly improved Frechet Inception Distance relative to CycleGAN and preserved interpretable control over features such as cluster number and location in clustered lumpy background models (Xu et al., 2024). Data-centric phantom resources support algorithm benchmarking beyond learning as well: the dynamic X-ray gel phantom dataset provides 17 time frames, dense reference frames, metadata for ASTRA- or Spot-based forward operators, and example MATLAB code for sparse-angle testing, time-resolved reconstruction, and comparison to dense-angle references by RMSE, SSIM, and PSNR (Heikkilä et al., 2020).

6. Limitations, biases, and standardization pressures

The literature is equally explicit about the limitations of measurement phantoms. Material surrogates may match one contrast mechanism while failing in another: the aqueous K$c_{us}=1418\pm31$7HPO$c_{us}=1418\pm31$8 bone surrogate in the CT/MRI proton-therapy phantom approximated radiological properties but produced undesirably high MR signal intensity (Gallas et al., 2014). Small structures are a recurring failure mode: in the cystic lung phantom, individual cyst volumes were increasingly underestimated as diameter decreased, with underestimation exceeding 3–5% below about 10 mm and reaching roughly 16% at 5 mm (Shah et al., 25 Mar 2025). In quantitative photoacoustics, absolute recovery remains difficult: the best reported relative errors in the phantom study were about 22% for ground-truth-fluence correction and about 29% for the experimentally trained network on inclusions, and none of the evaluated methods fully corrected spectral coloring above 850 nm in the blood-flow phantom (Gröhl et al., 2023, Gehrung et al., 2019).

Patient-specific and anthropomorphic phantoms also expose unresolved coverage gaps. iPhantom identifies training data size, registration speed, and out-of-field structures as limitations; organs entirely outside the scan range can exhibit larger dose errors, and larger annotated datasets are expected to improve accuracy (Fu et al., 2020). In anthropomorphic multispectral oximetry phantoms, nonlinear mixing of proxy dyes complicates generation of intermediate sO$c_{us}=1418\pm31$9 states, long-term dye stability in the copolymer matrix was untested, and residual air bubbles can induce reverberation artifacts in deeper vessels (Dreher et al., 29 Mar 2025). For fired-clay skull models, long-term stability under cyclic loading or environmental changes was not assessed, and mechanical properties were not quantified (Hunold et al., 2018).

These limitations have produced a strong standardization pressure. Several studies release datasets, code, CAD workflows, or recipe details explicitly to support reproducibility, including dynamic tomography data with accompanying examples, open-source USCT phantom-generation scripts, and open-source anthropomorphic oximetry phantom resources (Heikkilä et al., 2020, Li et al., 2021, Dreher et al., 29 Mar 2025). This suggests that a contemporary measurement phantom is increasingly expected to be accompanied by documented fabrication, property characterization, forward modeling assumptions, and public benchmarking assets.