RPS 1605 - Hot Topic: generative AI in radiology

Author Block: G. Della Pepa, G. Irmici, V. Molinari, C. De Berardinis, E. D'Ascoli, L. Corradini, G. Rossini, C. Depretto, G. P. Scaperrotta; Milan/IT
Purpose: To design and evaluate a generative artificial intelligence (AI) tool that automatically produces personalized and structured breast imaging reports (mammography, ultrasound, and interventional procedures) from minimal user input.
Methods or Background: A prototype based on a large language model (LLM) was developed to learn each radiologist’s unique reporting style. For each user, 50 anonymized previous breast imaging reports were combined and analyzed to generate an automatically created JSON (JavaScript Object Notation) structure encoding linguistic patterns, recurrent terminology, and section hierarchy. A second, handcrafted JSON file defined the functional core of the LLM, including task logic, prompts, and behavioral instructions. The system combines both components to generate full reports from brief notes summarizing the current exam, optionally integrating previous reports for automated comparison. Internal testing was conducted comparing reporting time and usability across manual typing, voice dictation, and the AI-assisted workflow.
Results or Findings: The AI tool, tested by five breast radiologists, generated complete, correct, and style-consistent reports within a few seconds from brief inputs. Reporting time was significantly reduced compared with both manual typing and voice dictation. Radiologists reported smoother workflow and lower cognitive load after the initial setup. Mean satisfaction scores were 9.2/10 for usability and 9.0/10 for accuracy. Generated reports required only minimal editing before validation.
Conclusion: The developed tool demonstrates that personalized, fast, and consistent breast imaging reporting can be achieved through generative AI, improving workflow efficiency and ensuring high reporting accuracy.
Limitations: Preliminary single-center testing. A refined model with expanded parameters and larger datasets is under development to improve precision and scalability
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: M. Meyer, S. Steinmetz, A. Othman, D. Graafen; Mainz/DE
Purpose: Pulmonary embolism (PE) is a potentially life-threatening event requiring accurate imaging and diagnostics. CT pulmonary angiography (CTPA) is widely established as imaging standard for PE detection but can face difficulties in poorly contrasted images or in the detection of small, peripheral PEs. This study aims to evaluate the potential benefits of Deep Learning-based iodine contrast enhancement and denoising in CTPA (DLe-CTPA) compared to conventional CTPA (c-CTPA).
Methods or Background: This retrospective, multi-center study analyzed data from 224 patients who underwent CTPA for suspected pulmonary embolism (PE). Imaging data were collected from eight different CT scanners, and a vendor-agnostic Deep Learning algorithm was utilized to enhance iodine contrast. For the quantitative analysis, Signal intensity (SI), Signal-to-Noise Ratio (SNR) and Contrast-to-Noise Ratio (CNR) where obtained at eleven distinct pulmonary artery segments using a Matlab tool. For the qualitative analysis, two radiologists independently evaluated both data sets and performed qualitative assessments using a 4-point Likert scale to evaluate various parameters, including image quality, contrast, image noise and diagnostic confidence.
Results or Findings: DLe-CTPA significantly enhanced SI, SNR, and CNR across all pulmonary artery sites and scanners compared to c-CTPA (SI, CNR: p<0.001; SNR: p<0.03). Qualitative assessment showed improvements for overall image quality, contrast, image noise and diagnostic confidence for DLe-CTPA (all p<0.001).
Conclusion: The application of DLe-CTPA significantly improves SI, SNR and CNR for all pulmonary artery segments and CT scanners, as well as the qualitative measures overall image quality, contrast, image noise and diagnostic confidence when compared to c-CTPA.
Limitations: We are limited by retrospective study design.
Funding for this study: No.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the local ethic commission.

Author Block: M. Solak, M. TÖREN, B. ASAN, E. Kaba, M. Beyazal, F. BEYAZAL ÇELİEKR; Rize/TR
Purpose: Contrast-enhanced MRI (CE-MRI) is essential for diagnosing and monitoring brain metastases (BM). However, potential risks of gadolinium-based contrast agents (GBCA) call for alternative methods. This study aimed to generate synthetic CE-T1 and CE-FLAIR from non-contrast MRI in BM patients using Generative adversarial networks (GAN)-based models.
Methods or Background: This retrospective study utilized brain MRI scans (3.0T GE Discovery MR750w) acquired between January 2023 and April 2025 from 183 patients with brain metastases (T1: 83, T2: 100). The dataset included 16,250 T1-related and 2,810 T2-related images (T1, CE-T1, T2, and CE-FLAIR sequences). For CE-MRI synthesis, four GAN models (Pix2PixHD, CycleGAN, C-CycleGAN, and CGAN) were applied using sample boundary maps instead of semantic labelling. Data were divided into training (70%), validation (20%), and testing (10%) sets. Generator realism was assessed via G_GAN loss, and visual quality via VGG-based perceptual loss. Model outputs were quantitatively evaluated using MSE, SSIM, PSNR, and RMSE. Qualitative evaluation was performed through a visual Turing test by experts and repeated after one month.
Results or Findings: G_GAN losses generally decreased, and T1 and T2 VGG losses showed strong correlation (r = 0.90). Pix2PixHD achieved thr best performance (T1: SSIM 0.80, PSNR 29.2 dB; T2: SSIM 0.90, PSNR 27.3 dB), while CGAN performed worst (T1: SSIM 0.30, PSNR 19.1 dB; T2: SSIM 0.20, PSNR 18.5 dB). Qualitative evaluation showed 61.4% accuracy with moderate inter-observer agreement; repeat testing showed 60% accuracy.
Conclusion: GAN-based models can generate CE-T1 and CE-FLAIR images from non-contrast MRI, achieving strong quantitative metrics and moderate diagnostic accuracy, supporting their potential to reduce exposure GBCA in neuro-oncologic imaging.
Limitations: Not applicable.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The ethics committee notification can be found under the number Decision No: 2025/280.

Author Block: Y. Ni, C. Liu, W. Li, H. F. V. Lee, A. E. Helali, Y-L. Wong, G. Ren, J. Cai, T. Li; Hong Kong/HK
Purpose: Multiparametric MRI (mpMRI) is essential for glioblastoma (GBM) management but is protracted and costly. Deep learning synthesis from conventional MRI suffers from poor generalizability across scanners. We propose a novel solution: a quantitative synthesis network (QS-Net) that generates diagnostic-quality mpMRI contrasts directly from quantitative Magnetic Resonance Fingerprinting (MRF) maps, inherently bypassing scanner-specific contrast variability.
Methods or Background: A generative adversarial network with deep-supervised residual blocks was developed to synthesize standard mpMRI (T1w, T2w, FLAIR, SWI) from core MRF-derived T1/T2 maps. The model was trained on 32 healthy volunteer scans and fine-tuned on 9 GBM patient scans, with hold-out testing on 9 independent GBM patients. Performance was quantitatively (MAE, SSIM, PSNR) and qualitatively compared against Res-Unet, cGAN, and Swin-Transformer models. A critical generalizability test evaluated all architectures when trained on MRF maps versus conventional MRI inputs.
Results or Findings: QS-Net significantly outperformed all state-of-the-art models, achieving superior quantitative metrics across all synthesized contrasts: T1w, T2w, FLAIR, and SWI (e.g., MAE: 1.01-1.45e-02; SSIM: 0.926-0.939; PSNR: 27.56-29.69) and qualitatively reproducing critical tumor boundaries and internal texture. Crucially, models trained on quantitative MRF inputs demonstrated consistently superior generalizability (p < 0.005) across all architectures compared to those trained on conventional qualitative MRI.
Conclusion: QS-Net establishes a new paradigm for rapid, high-fidelity mpMRI synthesis by leveraging quantitative MRF as a scanner-agnostic source. This directly addresses the critical limitation of generalizability in prior methods, with immediate clinical potential to streamline GBM imaging protocols, reduce acquisition time, and standardize diagnostics across healthcare institutions.
Limitations: The study's cohort size, while sufficient for a proof-of-concept, is limited. Further validation on a larger, multi-scanner cohort is necessary to confirm robust clinical deployment.
Funding for this study: This study has been supported by: (1) The National Natural Science Foundation of China Young Scientist Fund (NSFC-YSF 82202941) from China; (2) The Innovation and Technology Support Program (ITS/049/22FP) from the Hong Kong Special Administrative Region (HKSAR), China; (3) The General Research Fund (GRF 15104822, GRF 15102219) from HKSAR, China; (4) The Health and Medical Research Fund (HMRF 10211606, HMRF 06173276) from HKSAR, China.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: S. Pasumarthi Venkata¹, S. Colombo Serra², J. Vymazal³, A. Shankaranarayanan¹; ¹Menlo Park, CA/US, ²Colleretto Giacosa/IT, ³Prague/CZ
Purpose: Higher-relaxivity gadolinium-based contrast agents (GBCAs) like Gadopiclenol have recently been introduced which are being used at lower dose levels to achieve comparable lesion visualization. In this work, we show that a deep learning (DL) based contrast boosting algorithm further amplifies the benefits of such higher-relaxivity GBCAs, by improving lesion visualization and image quality without increasing the dosage.
Methods or Background: T1w pre-contrast and standard-contrast (SC) images from 20 patients (2D and 3D scans) were obtained for this study. The patients were injected with 0.05 mmol/kg of Gadopiclenol. The pre-contrast and SC images were used to generate contrast boosted (CB) images using an FDA-cleared DL algorithm that boosts the contrast signals present in the SC images. Rectangular regions-of-interest (ROIs) were drawn on all enhancing lesions and on the healthy parenchymal tissues. These ROIs were used to compute the contrast-to-noise ratio (CNR), lesion-to-brain ratio (LBR) and contrast enhancement percentage (CEP).
Results or Findings: The mean CNR, LBR and CEP of CB images (10.38±1.56, 5.12±1.03 and 1.96±0.35) were greater than that of SC images (2.52±0.85, 2.55±1.85 and 0.99±0.28). The percentage increase of CNR, LBR and CEP from SC to CB images are 314.19%, 118.12% and 335.18% respectively. We separately calculated the percent increase of CNR, LBR and CEP for patients injected with Gadobenate dimeglumine (at 0.1 mmol/kg) and found it to be 205.97%, 60.50% and 83.19%.
Conclusion: Quantitative analysis has shown that the CB algorithm has improved the benefits of Gadopiclenol. From the percent increase in CEP it can be seen that the CB images are equivalent to a double dose injection. Comparative analysis with normal GBCAs has revealed that the CB algorithm performance is better in higher relaxivity GBCAs.
Limitations: The study is limited to a small sample size.
Funding for this study: n/a
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Internal IRB

Author Block: D. Wang¹, S. Pasumarthi Venkata¹, T. C. Arnold¹, A. Shankaranarayanan¹, G. Zaharchuk²; ¹Menlo Park/US, ²Stanford/US
Purpose: Gadolinium-based contrast agents (GBCAs) are widely used in brain MRI to enhance lesion visibility for diagnosing and monitoring brain tumors. However, due to safety concerns such as long-term gadolinium retention in tissues, there is growing interest in dose-reduction strategies. In this study, we investigate the feasibility of a fast diffusion model to synthesize contrast-enhanced images from pre-contrast scans.
Methods or Background: The proposed method integrates a conditional denoising diffusion probabilistic model with adversarial learning to achieve high-fidelity medical image synthesis. The model maps source domain images to conditioning features to guide the diffusion model during generation. During inference, the model runs on latent features, thus enabling high throughput at 22 fps. The studied dataset includes 126 clinical cases (113 for training and 13 for testing) acquired using a Philips Insignia 3T scanner with Gadoterate meglumine contrast agent. The cohort consists of 55 female and 71 male patients, with a mean age of 48 ± 16 years. All pre-contrast and SOC-Gad images were mean-normalized, skull-stripped, and affine co-registered using the pre-contrast scan as reference.
Results or Findings: The Syn-Gad images showed strong similarity to SOC-Gad images, with a mean PSNR of 34.13, SSIM of 0.8906, RMSE of 0.0202, and CNR of 0.0523, further supporting the fidelity of the synthesized images. Qualitative analysis shows that Synth-Gad images preserve enhancement patterns comparable to those in SOC-Gad, particularly around enhancing tissue regions.
Conclusion: The proposed diffusion-based synthesis model demonstrates strong potential for contrast dose reduction in brain MRI. This approach offers a promising path toward reducing patient and environmental exposure to gadolinium-based contrast agents in routine brain MRI exams.
Limitations: The work still requires clinical validation in larger, pathology-rich cohorts.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: P. Mathur, B. S. Kelly, R. P. Killeen, A. Lawlor; Dublin/IE
Purpose: To investigate the determinism and consistency of diffusion-based generative models for synthesising longitudinal magnetic resonance imaging (MRI) progression in Multiple Sclerosis (MS), comparing a state-of-the-art diffusion model, Brownian Bridge Diffusion (BBDM), and its Mask-Guided extension (MGBBDM).
Methods or Background: Baseline and follow-up brain MRI slices from 110 patients were used for training, 30 for validation, and 30 for testing, yielding approximately 6,000 paired slices for model training. Three diffusion-based generative models were trained to simulate follow-up images conditioned on baseline scans and corresponding lesion masks (except BBDM). For each model, five synthetic samples were generated per input to assess pixel-wise variability and reproducibility. Consistency was quantified using the coefficient of variation (CV) across Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and L1 distance metrics.
Results or Findings: The MGBBDM demonstrated the lowest inter-sample variability, achieving PSNR CV = 9.41% and SSIM CV = 8.89% (both Excellent), with moderate L1 CV = 20.64%. In contrast, the Improved Diffusion and standard BBDM exhibited higher variability (PSNR CV ≈ 12–13%, L1 CV > 30%). The BBDM without mask guidance showed no consistent lesion progression, producing follow-up images resembling the baseline. These findings indicate that mask guidance is essential for achieving deterministic and clinically meaningful lesion evolution.
Conclusion: Mask-Guided Brownian Bridge Diffusion offers superior reproducibility and anatomical stability compared to standard diffusion models. By enforcing lesion-specific constraints during synthesis, it produces consistent and clinically interpretable MS progression—an essential step toward reliable generative modelling in longitudinal neuroimaging. Deterministic synthesis supports the creation of reproducible imaging datasets, enabling clinically trustworthy simulation and validation for Multiple Sclerosis tracking.
Limitations: The study is limited to 2D modelling with a moderate dataset. Future work will extend to 3D evolution and clinical validation.
Funding for this study: This research was supported by Science Foundation Ireland (SFI) under Grant Number SFI/12/RC/2289_P2 and the Irish Centre for High-End Computing (ICHEC)
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: N. Benchekroun, G. Webber, M. Alshammari, A. Hammers; London/UK
Purpose: Combined MRI and PET imaging enables fusion of anatomical and metabolic information in neuroimaging, but simultaneous acquisition is limited by the availability of hybrid PET-MR scanners, cost, and patient burden. Shortage of paired PET-MRI datasets further constrains research requiring both modalities. This study aims to evaluate generative AI methods for synthesising MRI brain images from FDG-PET scans.
Methods or Background: A Denoising Diffusion Probabilistic Model (DDPM) with U-Net architecture was developed for cross-modality synthesis from FDG-PET to MRI. The model was trained on over 1,400 axial slices extracted from 149 paired PET-MRI volumes and evaluated on 62 independent test subjects. Two approaches were systematically compared: a 2D method processing slices independently, and a 2.5D method incorporating adjacent slice context to enhance spatial consistency and inter-slice coherence. Quantitative performance was assessed using structural similarity index measure (SSIM) and Pearson correlation coefficients, while visual assessment evaluated anatomical structure preservation and artefact patterns.
Results or Findings: Both approaches achieved reasonable synthesis quality with modest improvements for the 2.5D method. Quantitative evaluation demonstrated 3-10% improvements across metrics for the 2.5D approach, with SSIM increasing from 0.626±0.064 to 0.651±0.062 and correlation improving from 0.798±0.043 to 0.822±0.049. Visual assessment revealed that while both approaches adequately preserved anatomical structures, the 2.5D method produced more systematic error patterns concentrated at tissue boundaries, with more randomly scattered artefacts for the 2D approach.
Conclusion: Diffusion-based cross-modality synthesis can generate structurally coherent MRI images from FDG-PET scans, with the 2.5D context-aware approach providing modest but consistent improvements in synthesis quality. These methods demonstrate potential for supplementing missing MRI data in research and clinical applications.
Limitations: While results are promising, SSIM is still limited, and clinical validation outstanding.
Funding for this study: The School of Biomedical Engineering and Imaging Sciences is supported by the Wellcome EPSRC Centre for Medical Engineering at King’s College London (WT 203148/Z/16/Z) and the Department of Health via the National Institute for Health Research (NIHR) comprehensive Biomedical Research Centre award to Guy’s & St Thomas’ NHS Foundation Trust in partnership with King’s College London and King’s College Hospital NHS Foundation Trust. MA is supported by the Saudi Arabia Cultural Bureau in London under the Saudi scholarship program. NB was supported by a CDT summer studentship.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: J. Pfann Hossbach, H. Schreiter, L. Brock, T-T. Nguyen, S. Heidarikahkesh, A. George, R. Janka, M. Uder, S. Bickelhaupt; Erlangen/DE
Purpose: Avoiding dynamic contrast enhanced (DCE) prostate MRI acquisition can accelerate clinical workflows, increasing the use of prostate MRI. We aim to generate virtual DCE (vDCE) images from multiparametric non-contrast-enhanced sequences using artificial intelligence as potential substitute.
Methods or Background: This IRB-approved retrospective study included n=2092 patients who underwent clinical prostate examinations with T1w-DCE at 3T scanners (Siemens Healthineers MAGNETOM Skyra/Vida). T1w, T2w, and DWI (b-values: 50, 800, 1500 s/mm²) acquisitions were used to train a GAN network; a 2.5D U-Net with 2 discriminators (full/half resolution). Data were resampled to a mutual FOV/resolution, sequence-wise normalized, and split into train=1450, validation=419, and test=213 subjects (⌀25 slices). To standardize and reduce temporal resolution, DCE images at 15, 30, 45 and 60s after acquisition start were selected and registered with a separately trained VoxelMorph-Network forming the targets. The training for 100 epochs minimized the combined adversarial (binary cross entropy), perceptual, L1 and SSIM loss between predicted and target slices using the non-DCE images with their ±1 neighboring slices as input.
Results or Findings: The generated test data achieved a MS-SSIM of 0.9649, 0.9298, 0.8762 and 0.853, SSIM of 0.9251, 0.8516, 0.7801 and 0.7407 and NRMSE of 0.0318, 0.0484, 0.0657 and 0.0713, respectively, outperforming state-of-the-art single timepoint predictions. Radiologist classified the overall image quality of n=200 targets/predictions into real/generated, yielding near-equal counts: real (n=140 real, n=138 generated) and generated (n=60 real, n=62 generated).
Conclusion: Multi-timepoint vDCE image generation was technically feasible and indistinguishable from real images for the reader. Further work is necessary to improve the method and to assess its potential for prostate MRI in clinical practice.
Limitations: A lesion enhancement comparison and diagnostic value evaluation was not conducted in this retrospective single-centre study. Furthermore, truly dynamic image generation was not addressed.
Funding for this study: This research was funded by the Bavarian State Ministry of Economic Affairs, Regional Development and Energy.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Study was approved by ethics Committee of FAU with waived infromed consent

Author Block: K. Nechypurenko, T. Nechypurenko; Kyiv/UA
Purpose: To synthesize evidence on hidden feature propagation and demographic bias in generative AI models for synthetic medical image generation, evaluating implications for diagnostic algorithm performance and equity across patient populations.
Methods or Background: Narrative synthesis of peer-reviewed literature (2024-2025) examining generative AI in radiology. We analyzed 8 key publications including: one systematic survey of 103 diffusion model studies, three foundation model validation studies with datasets exceeding 20 million images (SA-Med2D-20M, TotalSegmentator, UK Biobank), two multi-institutional bias assessments, and two radiologist evaluation studies. We extracted data on diagnostic performance metrics, demographic bias patterns, and validation methodologies across CT, MRI, chest radiography, and histopathology.
Results or Findings: Synthetic image quality varied by modality. Pathologists could not distinguish synthetic histopathology images while radiologists identified 96.6% of synthetic radiological images as non-authentic. Algorithms trained with combined synthetic-authentic data showed improvements: isocitrate dehydrogenase mutation prediction area under curve 0.75, glioma volumetry concordance 0.782, lesion detection (sensitivity improvements up to 25% for specific tasks). However, foundation models exhibited persistent demographic disparities: decreased accuracy for female patients (normal chest radiographs) and patients of African ancestry (pleural effusion), despite massive training datasets. Algorithms extracted patient race through hidden features imperceptible to radiologists. Demographic-balanced fine-tuning provided only partial mitigation.
Conclusion: Synthetic medical images enhance diagnostic algorithm performance for rare pathologies and data-scarce scenarios but systematically propagate demographic biases through hidden features undetectable by visual assessment. Clinical deployment requires demographic-stratified validation protocols and continuous performance monitoring across patient populations to ensure equitable diagnostic accuracy.
Limitations: Secondary literature synthesis with heterogeneous validation methodologies across studies. Limited long-term clinical outcome data. Geographic and demographic representation varied across included datasets.
Funding for this study: No funding was received for this study
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: