RPS 2413 - Upcoming technologies and deep learning reconstruction across different imaging modalities

Author Block: A. Entezam¹, A. Pakzad¹, A. Maksimenko¹, C. Hall¹, M. John Cameron¹, P. Brennan², D. Hausermann¹, T. Gureyev¹, H. M. Quiney¹; ¹Melbourne/AU, ²Sydney/AU
Purpose: To develop and validate a patient-specific Monte Carlo (MC) dosimetry framework for propagation-based X-ray phase-contrast breast computed tomography (BCT) at the Imaging and Medical Beamline (IMBL) of the Australian Synchrotron, enabling accurate estimation of mean glandular dose (MGD) for clinical translation.
Methods or Background: BCT provides 3D imaging without breast compression, improving comfort and enhancing visualization of internal structures—crucial for accurate cancer detection. Propagation-based phase-contrast BCT offers superior soft-tissue contrast at equal resolution and comparable or lower doses than conventional absorption-based BCT. Accurate radiation dosimetry is vital for patient safety and optimal imaging performance. Most existing MC-based MGD studies use homogeneous phantoms, neglecting patient-specific anatomy that affects dose distribution. To overcome this, a voxel-based MC dosimetry framework was developed using EGSnrc to calculate MGD in realistic anthropomorphic breast phantoms derived from synchrotron BCT images, tailored to IMBL beam properties. Corresponding homogeneous phantoms were also generated, and air kerma–to–MGD conversion factors were obtained and compared with heterogeneous phantom results. Simulations were performed for varying breast densities, dimensions, and skin thicknesses across 28–38 keV.
Results or Findings: Breast anatomy and beam energy strongly influenced MGD. Higher glandular density and larger breasts increased MGD, while a 2 mm thicker skin layer reduced it by about 10% at equal air kerma. MGDs decreased with rising beam energy, though less markedly in larger breasts. Comparisons between heterogeneous and homogeneous phantoms showed clear differences in conversion coefficients, underscoring the need to include anatomical heterogeneity in dosimetric modelling.
Conclusion: The voxel-based MC framework offers an anatomically accurate, patient-specific approach to dosimetry in synchrotron phase-contrast BCT, enabling precise dose estimation and supporting safe, effective clinical implementation at the Australian Synchrotron..
Limitations: This study focused on simulations performed using only EGSnrc MC code.
Funding for this study: Funding was provided by Australian National Health and Medical Research Council (NHMRC) Synergy Grant (APP2011204).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethical approval was received from the Human Research Ethics Committee of Monash University (project number: CF15/31382015001340).

Author Block: A. Belarra¹, M. Castillo¹, I. Hernandez-Giron², N. Amallal El Ouahabi¹, A. Tejerina¹, P. Homolka³, M. Chevalier¹; ¹Madrid/ES, ²Dublin/IE, ³Vienna/AT
Purpose: Contrast-enhanced mammography (CEM) is a widespread imaging modality to improve lesion detection. In this study, a 3D-printed anthropomorphic phantom including iodinated resin disks is employed to evaluate CEM performance.
Methods or Background: A breast-CT volume segmented into gland/skin and fat and further compressed (thickness: 4.8 cm) was used to generate four 1.2cm-thick slices, which were 3D-printed (BambuLab X1C, 0.2 mm nozzle, 0.14 mm layer height, 100% infill-lines) selecting PLA for glandular tissue/skin and ABS for fat. One of the intermediate slices included five cylindrical recesses (8mm diameter, 0.8mm height) located over different backgrounds. Three sets of four resin iodinated disks (Iomeron400) were SLA-printed with varying concentrations (0.50, 0.75, 1.00 mg/cm²). A fifth disk was PLA-printed as a control. Three CEM phantom images were acquired (Hologic-3Dimensions)for each set of resins+PLA disks. On the images, the difference (S) between MPV of a 7mm-circular-ROI inside each disk and MPV of a 1mm-band-ROI on the background were computed. For each set, the average S was obtained across images (and all disks for resins ones).
Results or Findings: The phantom background was successfully removed in CEM images. S values were [averageSD(range)]: [0±2(-2,1) for the PLA disk] and for iodinated disks: [23±5(17,28); 0.50mg/cm2], [30±7(24,39); 0.75mg/cm2] and [44±4(40,49); 1.00mg/cm2]. S value and iodine concentration were linearly dependent (43.1 slope; R2=0.998).
Conclusion: In CEM images, both PLA (glandular tissue) and background (parenchyma) were successfully removed, whereas resin-disk signals remained visible. As expected, S values increased with iodine concentration. These results support the use of 3D-printed anthropomorphic phantoms for evaluating the CEM performance under conditions that better mimic the clinical scenario.
Limitations: A wider range of concentrations must be evaluated and a better recess-disk fitting achieved to reduce the air artefact on the disk borders.
Funding for this study: This work was supported by the Spanish Ministry of Science and Innovation under grants PID2021-123390OB-C22 and PID2021-123390OB-C22.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: L. D'Hondt¹, D. Buytaert¹, P-J. Kellens¹, A. Snoeckx², K. Bacher¹; ¹Gent/BE, ²Antwerp/BE
Purpose: While phantoms support objective CT image quality and dose evaluations, off-the-shelf models insufficiently reflect complex morphologies and radiodensity heterogeneities, thereby limiting clinical translation of results. Part-solid pulmonary nodules exemplify intricate structures inadequately represented by generic models. Adaptation of 3D-printing in research could improve phantoms. This study aimed to 3D-print part-solid nodule phantoms that mimic patient nodules, as validated by a human reader study.
Methods or Background: Seven part-solid pulmonary nodules on patient-derived CT series were segmented and solid core and ground glass densities were measured through Hounsfield Unit (HU) thresholding. Four target radiodensity ranges were defined and stereolithography-based 3D-printing parameters were optimised towards 3D-designs of nodules with a solid core (>0HU) and three gradients mimicking ground glass opacity ([0; -300HU[, [-300; -500HU[, [-500; -750HU]). Low-dose CT acquisitions of 3D-prints incorporated in a Lungman phantom were randomised with patient scans. A multicentre, single-blinded reader study, involving seventeen radiologists, assessed whether 3D-printed nodules were distinguishable from real part-solid nodules using χ2 goodness-of-fit testing.
Results or Findings: Final 3D-printed nodule models consisted of the four target radiodensities measured on CT-images of clinical nodules. Overall accuracy of all scored nodules was 53.50%. Importantly, the false negative rate indicates that, across all radiologists, 47.50% of 3D-printed nodules was incorrectly classified as a real patient nodule, while 44.54% of patient nodules were thought to be 3D-prints. There was significant evidence (χ2= 136.13; p= 1.864e-31; 5% confidence level) against readers reliably distinguishing patient nodules from our 3D-printed phantoms.
Conclusion: We were able to manufacture nodule phantom models with morphological and radiological properties that are highly customisable, relatively inexpensive and indistinguishable from real clinical examples. As such, this study demonstrates that adaptation of 3D-printing can enhance clinical relevance of phantom research.
Limitations: No limitations were identified.
Funding for this study: Funding was provided by the FWO “Kom op tegen Kanker”-project for lung cancer screening research in Belgium. (Project number: G0B1922N).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The retrospective use of patient CT series was approved by the Antwerp University Hospital Ethics Committee (Approval Number: EC/PM/eva/2024.048).

Author Block: L. Kaster¹, M. Lochschmidt¹, A. M. Bauer¹, T. Dorosti¹, S. Demianova¹, T. Koehler², D. Pfeiffer¹, F. Pfeiffer¹; ¹Munich/DE, ²Hamburg/DE
Purpose: Dark-field chest radiography has demonstrated superiority over conventional radiography for diagnosing and staging pulmonary diseases by visualising microstructural lung changes. However, rib- and clavicle-induced beam hardening artefacts generate false dark-field signals that impair diagnostic accuracy. This study introduces an AI-based segmentation framework for anatomically adaptive beam hardening correction (BHC) in clinical dark-field radiographs.
Methods or Background: Rib and clavicle segmentations were obtained using U-Net models trained on the VinDr-RibCXR dataset and clinical attenuation images. Masks were refined into anterior and posterior rib components and combined with attenuation-contribution weights derived from dual-energy CT material decomposition (aluminium and water as surrogates for bone and soft tissue). Patient-specific correction maps were generated and subtracted from raw dark-field images. The method was evaluated on 174 radiographs (healthy, COPD, COVID-19) using qualitative assessment and quantitative measures of signal homogeneity and diagnostic separability.
Results or Findings: Uncorrected dark-field radiographs showed elevated signals at rib and clavicle locations, obscuring the lung parenchyma and reducing interpretability. The proposed segmentation-based BHC suppressed these artefacts, yielding a more homogeneous dark-field signal and improved visualisation of pulmonary structures. Improvements were consistently observed across all cohorts. Quantitatively, intra-pulmonary signal variability decreased (coefficient of variation −16%, p < 0.001), and diagnostic separability improved, with overlap between healthy and diseased lungs reduced by up to 18.8%.
Conclusion: AI-segmentation-based BHC enables regionally adaptive correction of bone-induced artefacts in dark-field chest radiography, improving image homogeneity and visual differentiation of pathological changes. This addresses a key limitation of the modality and strengthens its clinical potential for pulmonary disease assessment.
Limitations: The limitations of the study are that correction performance depends on segmentation accuracy and on contribution weights derived from limited spectral CT datasets. Patients with atypical anatomy or metallic implants may still exhibit residual artefacts.
Funding for this study: Funding was provided by the European Research Council (ERC Synergy Grant SmartX, SyG 101167328), the Free State of Bavaria under the Excellence Strategy of the Federal Government and the Länder, and the Institute of Advanced Study, Technical University of Munich.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the Ethics Commission of the Medical Faculty, Technical University of Munich (reference number 166/20S and 587/16S) .

Author Block: R. Villa, N. Paruccini, E. De Ponti; Monza/IT
Purpose: To characterize the image quality performance of a deep learning-based (DL) reconstruction algorithm, Philips Precise Image, and compare it with iterative reconstruction (IR) algorithms (iDose 4 and IMR) under clinically relevant conditions. The goal is to evaluate its potential in optimizing CT protocols.
Methods or Background: Two phantoms were used for this study: an anthropomorphic abdomen phantom (PhantomX, Germany) and a Catphan phantom (The Phantom Laboratory, USA). Both phantoms were acquired at five dose levels (1.7-6.8 mGy), with each acquisition repeated five times. Two CT scanners were used: one with Philips iDose 4 and IMR, and another with Philips iDose 4 and Precise Image. Image quality was evaluated in terms of noise, spatial resolution, blur, visual artifacts, and detectability index with a channelized Hotelling observer.
Results or Findings: Precise Image significantly improved all image quality parameters compared to iDose 4. Furthermore, noise reduction was up to 70%, spatial resolution improved by up to 20%, and CHO scoring increased by up to 90%. The improvement was less pronounced with IMR; however, Precise Image demonstrated behaviour more similar to filtered back projection (FBP) than IMR, with less dependency of image quality on contrast. This similarity makes it easier to optimize CT protocols. Additionally, Precise Image produced fewer visual artifacts than IMR, resulting in images that were much more similar in appearance to FBP images and reducing the synthetic texture typically associated with iterative reconstruction techniques.
Conclusion: The DL reconstruction algorithm demonstrates superior image quality performance compared to conventional iterative reconstruction methods. Its image characteristics, which closely resemble those of FBP, suggest it could be a valuable tool for simplifying and improving CT protocol optimization.
Limitations: Phantom size, small size compared to an adult patient.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: D. W. Shin¹, J. Im², L. Liu², B. E. Hoppel¹, T. Hagio³, A. Dhanantwari¹, K. Boedeker³, P. Noël²; ¹Tustin, CA/US, ²Philadelphia, PA/US, ³Ōtawara/JP
Purpose: To use 3D-printed anthropomorphic abdominal phantoms to compare trends in low contrast detectability (LCD) across various reconstruction algorithms and doses.
Methods or Background: An anthropomorphic abdominal phantom, consisting of two modules, was 3D-printed with PixelPrint, a technique that converts patient CT images and/or synthetic datasets into printer instructions which modulate printed density at sub-resolution scale to mimic realistic tissue densities and textures. The first module consisted of native anatomy, while the second incorporated ten digitally added rods (5 mm diameter, 30 HU contrast) within the same native anatomical background. Scans were acquired using an abdominal CT protocol on a clinical CT scanner (Aquilion ONE INSIGHT, Canon Medical Systems) across four water-equivalent diameters (Dw = 15, 27, 31, and 35 cm) at dose levels from 4-30 mGy. Images were reconstructed with Filtered Back Projection (FBP), Hybrid Iterative Reconstruction (HIR), two Deep Learning Reconstructions (DLR1 and DLR2), and Super Resolution-Deep Learning Reconstruction (SR-DLR). The LCD was determined via a Non-Prewhitening Model Observer with an eye filter. The trend in detectability index (d′) across reconstruction algorithms was compared at various dose levels and phantom sizes.
Results or Findings: At the highest dose (30 mGy) and largest size (Dw 35 cm), the d’ were 1.84, 1.56, 1.44, 1.23, and 0.75 for SR-DLR, DLR1, DLR2, HIR, and FBP, respectively. At the lowest dose (4 mGy), the corresponding d’ were 0.95, 0.99, 0.85, 0.89, and 0.30. The SR-DLR generally showed better LCD performance than conventional algorithms across dose levels and phantom sizes.
Conclusion: Lifelike 3D-printed abdominal phantoms replicate clinical imaging conditions, demonstrating in this study that SR-DLR improves visualization of subtle lesions—even in larger patients or at reduced doses—thereby supporting greater diagnostic confidence.
Limitations: Future reader studies are required for full clinical translation.
Funding for this study: Canon Medical Systems USA
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: I. Fitton¹, S. Alkhazzam², M. H. Kharita², M. Dimailing¹, M. Khalife¹, C. Van Ngoc Ty¹, V. Tsapaki³; ¹Paris/FR, ²Doha/QA, ³Vienna/AT
Purpose: Slit-beam radiography is recognized as a low-dose imaging technique particularly suitable for spine, hips, and knees. Additional clinical applications could be expanded with photon-counting technology. The aim of this study was to evaluate the image quality of photon-counting (PCD) compared to pressured Xenon gaseous linear (GLD) detectors, using the ATIA International Atomic Energy Agency (IAEA) methodology.
Methods or Background: Seven EOS imaging systems including two models based on PCD and GLD technologies were studied. The IAEA test-object was propped along each detector. A block of 30 × 30 × 10 cm3 PMMA plates was placed at the output of the X-ray tube as an attenuator. Acquisition parameters were fixed : 70kV, 20mA, scanning speed of 6, no modulation. Image acquisitions were repeated five times on each of the detectors. « For-processing » (FPRO) and « For-presentation » (FPRE) image formats were compared when available. Signal-difference-to-noise ratio (SDNR), Normalized Noise Power Spectrum (NNPS), Modulation Transfer Function (MTF), detectability indexes (d') for small lesions with diameters of 0.3mm and 4mm were automatically quantified using PyATIA software. A Wilcoxon-test was used to compare both technologies.
Results or Findings: Results showed that task-based detectability metrics (SDNR and d′) were substantially higher on photon-counting. The spatial resolution was higher on PCD compared to GLD at high frequencies in FPRE: 11.21 lp/mm[8.55–12.24] vs. 2.64 lp/mm[2.16–3.44] (p<0.05).
The noise variance was greatly reduced on PCD and close to zero. The noise level was four times lower on PCD compared to GLD.
Conclusion: The IAEA ATIA methodology proved valid for inter-technology comparison, enabling benchmarking between PCD and GLD detectors. PCD demonstrated statistically superior image quality compared to GLD. This paves the way to a range of clinical applications of slit-beam radiography beyond spinal conditions.
Limitations: None
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: C. Graf¹, A. Jaffray¹, A. Rund¹, S. Steinerberger², A. Rauscher¹; ¹Vancouver, BC/CA, ²Seattle, WA/US
Purpose: The T1 relaxation time of the MRI signal can characterize brain tissue microstructure, e.g. to monitor disease progression and treatment response in conditions such as multiple sclerosis, tumours, or brain trauma. Variable flip angle (VFA) T1-mapping is fast but prone to bias from off-resonance, B1-inhomogeneity, imperfect spoiling, and magnetization transfer (MT). Our aim was to design excitation radiofrequency (RF) pulses with optimal control (OC) that are robust to B0/B1-variations and satisfy controlled-saturation MT (CSMT) to reduce MT-related bias, enabling accurate VFA-T1 without B1-mapping.
Methods or Background: Non-selective OC excitation pulses were designed for α=6° and 15° at 3T and 7T. Robustness levels covered B1-scales of 80-120% of nominal and B0-offsets of ±2.4ppm. Pulses were optimized by a semi-smooth quasi-Newton method with exact discrete derivatives; Bloch dynamics were solved via symmetric operator splitting. To enforce CSMT, pulses for both flip angles used identical duration and constant maximum amplitude, yielding matched RF energy​. Numerical VFA-T1 simulations compared OC versus block pulses at 3T and 7T, sampling wide B0/B1-ranges. Summary metrics within the target robustness box included min/max/mean T1 and the central 90%-range.
Results or Findings: OC pulses produced tighter T1-distributions than block pulses. At 3T, the 90% T1-range for OC was 748-882ms (≤10% from nominal 800ms), versus 540-1129ms (>40%) for block; at 7T, OC yielded 1255-1384ms (<7% from nominal 1300ms) versus 879-1842ms (~40%) for block. OC-T1 remained largely insensitive to B1 within 80-120% of nominal B1, whereas block-pulse T1 increased quadratically with B1.
Conclusion: OC-designed, CSMT-compliant excitation pulses substantially reduce B0/B1-related VFA-T1 bias at 3T and 7T while removing the need for B1-mapping. The approach is vendor-agnostic, patient-independent, and requires no specialized hardware, supporting practical deployment for robust quantitative T1-mapping.
Limitations: Future work will verify results experimentally.
Funding for this study: nA
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: V. Atamaniuk, L. Hanczyk, M. Obrzut; Rzeszów/PL
Purpose: Magnetic resonance elastography (MRE) is an advanced non-invasive imaging technique that quantifies the biomechanical properties of soft tissue. While clinical MRE is mainly applied to the liver, stiffness-related biomarkers may be valuable for other organs, including the thyroid, where tissue stiffening occurs in multiple pathological conditions. However, due to the gland’s small size and anatomical location, MRE of the thyroid is challenging without dedicated hardware. This study aimed to assess the feasibility of thyroid MRE using a custom-designed passive driver.
Methods or Background: 6 adult volunteers (aged 29–59 years) underwent thyroid MRE. A custom two-branch passive driver was built and positioned on the neck to cover both lobes of the thyroid. Vibrations at 120 Hz were generated by an active driver and transmitted through the passive driver to induce shear waves. Imaging was performed on a 1.5T whole-body MRI scanner using a SE-EPI-based 3D vector MRE sequence. Shear modulus magnitude and its storage and loss components were reconstructed using direct inversion. Wave field quality was assessed with octahedral shear strain SNR (OSS-SNR) and a confidence metric. Regions of interest were manually defined on T2-weighted images by a radiologist.
Results or Findings: The custom driver consistently provided excellent shear wave illumination in the thyroid. Mean ± SD OSS-SNR was 4.09±2.11, with confidence 81.39±3.45%. Displacement amplitudes ranged from 6.45 to 149.16 µm (mean ± SD: 51.24±21.45 µm). Mean ± SD stiffness, storage modulus, and loss modulus were 4.11±0.67 kPa, 3.87±0.70 kPa, and 1.22±0.61 kPa, respectively.
Conclusion: A dedicated passive driver makes thyroid MRE technically feasible, yielding reliable shear wave propagation and stiffness quantification. These findings support further research into the role of MRE for non-invasive thyroid characterization.
Limitations: Small sample size and lack of clinical data.
Funding for this study: National Science Centre, Poland (Grant ID: 2024/53/N/ST7/00358)
Polish National Agency for Academic Exchange (Grant ID: 25/NAWA-PROM-UR/2019)
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approved by the Regional Medical Chamber ethics committee (Resolution No 60/2022/B); written informed consent was obtained from all participants.