Research & Projects

I build clinically grounded AI systems for medical imaging research.

2026
A Digital Phantom for MR Spectroscopy Data Simulation
Simulated data is increasingly valued for validating MR spectroscopy (MRS) processing and analysis algorithms, but there is no consensus on the optimal simulation models or parameters. This work introduces a novel MRS digital brain phantom framework, a comprehensive and modular foundation for MRS data simulation, generated by combining anatomical and tissue label information with literature metabolite data. Paired with a signal-based model, the phantom generates various spectral datasets, saved in the NIfTI-MRS format for downstream use. Implemented across two anatomical templates at different resolutions and tested over a range of user-defined parameters, the phantom produced spectra with realistic signal characteristics and structural variation, validated against in vivo MRS data acquired under similar conditions.
Paper
2026
A Foundation Model Framework for MRI Classification of EVI and MFI in Rectal Cancer
Objectives: Accurate MRI-based identification of extramural vascular invasion (EVI) and mesorectal fascia invasion (MFI) is crucial for risk-stratified rectal cancer treatment. However, subjective visual assessment and inter-institutional variability limit diagnostic consistency. This study developed and evaluated a multi-center, foundation model-driven framework that automatically classifies EVI and MFI on axial and sagittal MRI. Materials and methods: A total of 331 pre-treatment rectal cancer T2-weighted MRI scans from three European hospitals were retrospectively recruited. A self-supervised frequency domain harmonization strategy was applied to reduce scanner variability. Three classifiers, SeResNet, the universal biomedical pretrained model (UMedPT) with a multilayer perceptron head, and a logistic-regression variant using frozen UMedPT features (UMedPT_LR), were trained (n = 265) and tested (n = 66). Gradient-weighted class activation mapping (Grad-CAM) visualized model predictions. Results: UMedPT_LR achieved the best EVI performance with multiplanar fusion (AUC = 0.82, test set). For MFI, UMedPT trained on axial harmonized images yielded the highest performance.
Paper
2026
An Automated AI Tool for Diagnosing Cardiac Disease from Conventional CMR
Cardiovascular magnetic resonance (CMR) imaging enables non-invasive assessment of myocardial structure, function, and pathology, but requires substantial experience in interpretation of CMR images that could be supported by artificial intelligence (AI)-based models. We developed an automated data curation pipeline using locally-run large language models to extract diagnostic labels from CMR reports, then fine-tuned three vision foundation models (DINO, VST, UMedPT) on multimodal imaging data across 988 curated cases. The ensemble approach achieved diagnostic performance with AUC-ROC values up to 0.966 for cardiac amyloidosis and 0.959 for hypertrophic cardiomyopathy. Training code and model weights are publicly available.
Paper Code
2026
Data-Driven Synthesis of Magnetic Resonance Spectroscopy Data Using a Variational Autoencoder
The development of deep learning methods for magnetic resonance spectroscopy (MRS) is often hindered by limited availability of large, high-quality training datasets. We propose using a variational autoencoder trained on measured spectroscopy data to generate synthetic samples. Our approach learns a compressed representation of complex spectra and can create new samples through latent-space operations. Testing revealed the model successfully reconstructs dominant spectral patterns and generates synthetic data matching in-vivo characteristics. When augmenting limited data subsets, synthetic spectra improved noise and signal quality metrics. However, the results also reveal limitations of the generative approach, including under-representation of stochastic noise and reduced accuracy in absolute metabolite quantification. The work establishes an evaluation framework emphasizing that validation should consider downstream applications when employing synthetic data.
Paper
2026
Extreme Cardiac MRI Analysis Under Respiratory Motion: Results of the CMRxMotion Challenge
Deep learning models achieve state-of-the-art performance in automated cardiac MRI (CMR) analysis, but their efficacy depends heavily on high-quality, artifact-free images, and their robustness to the respiratory motion artifacts common in clinical practice remains underexplored. We organized the MICCAI CMRxMotion challenge, curating and releasing a dataset of 320 CMR cine series from 40 healthy volunteers who followed controlled breathing protocols to induce a spectrum of motion artifacts. The challenge comprised two tasks, automated image-quality classification by motion severity and myocardial segmentation robust to motion, with 22 algorithms submitted and evaluated. This paper gives a comprehensive overview of the challenge design and dataset, reports evaluation results for the top-performing methods, and investigates how motion artifacts affect five clinically relevant biomarkers. All resources and code are public.
Paper Code
2026
Foundation Models Adaptation for Multi-View Multi-Modal Cardiac MRI Segmentation and Direct Ejection Fraction Estimation
Foundation models show strong transferability in cardiac MRI (CMR), but their effectiveness for heterogeneous multi-view and multi-sequence CMR analysis remains unclear. We explore fine-tuning and combining different CMR foundation models for the Universal Multi-Sequence, Multi-Center and Multi-View CMR Segmentation (CMR-Multi) Challenge. CineMA was fine-tuned for cine and late gadolinium enhancement (LGE) segmentation across short-axis and long-axis views; for direct left-ventricular ejection fraction (LVEF) estimation, two frozen CMR foundation models extracted embedding vectors combined via attention-based multiple-instance learning. On the challenge validation set, cine segmentation reached Dice scores of 0.862 to 0.902 across views, LGE segmentation reached 0.621 to 0.846, and direct LVEF regression achieved an MAE of 4.96 percentage points and a Pearson correlation of 0.91. These results show foundation models can be effectively adapted and combined for multi-view CMR analysis, while accurate LGE scar segmentation remains a challenging task.
Paper Code
2026
Large-Scale Modality-Invariant Foundation Models for Brain MRI Analysis
The field of computer vision is undergoing a paradigm shift toward large-scale foundation model pre-training via self-supervised learning (SSL). Leveraging large volumes of unlabeled brain MRI data, such models can learn anatomical priors that improve few-shot performance in diverse neuroimaging tasks. However, most SSL frameworks are tailored to natural images, and their adaptation to capture multi-modal MRI information remains underexplored. This work proposes a modality-invariant representation learning setup and evaluates its effectiveness in stroke and epilepsy lesion segmentation, following large-scale pre-training. Experimental results suggest that despite successful cross-modality alignment, lesion segmentation primarily benefits from preserving fine-grained modality-specific features. Model checkpoints and code are made publicly available.
Paper
2026
Pre- to Post-Contrast Synthesis of Breast DCE-MRI Using Latent Bridge Matching
Dynamic contrast-enhanced MRI (DCE-MRI) is central to breast cancer imaging, but gadolinium administration increases scan burden and motivates contrast-reduced alternatives such as synthetic contrast generation. This work proposes a latent bridge matching (LBM) framework that synthesizes peak-enhanced breast DCE-MRI from pre-contrast images for the MAMA-SYNTH challenge. Rather than starting from Gaussian noise like conventional latent diffusion models, it learns a conditional bridge between paired pre-contrast and peak-enhanced VAE latents, with a latent UNet predicting the correction toward the peak-enhanced latent so the trajectory stays anchored to patient-specific anatomy. Evaluated on 91 DUKE validation cases, tumor-conditioning improved performance over pre-contrast conditioning, reducing MSE from 1.023 to 0.940 and FRD from 7.523 to 4.716 while raising tumor SSIM from 0.355 to 0.429, outperforming the evaluated diffusion baseline. Further work is needed to validate generalization and remove dependence on ground-truth tumor masks at inference.
Paper
2026
Towards Brain MRI Foundation Models for the Clinic: Findings from the FOMO25 Challenge
Clinical deployment of automated brain MRI analysis faces a fundamental challenge: clinical data is heterogeneous and noisy, and high-quality labels are prohibitively costly to obtain. The FOMO25 challenge, organized at MICCAI 2025, provided researchers with the FOMO60K pretraining dataset and evaluated foundation models on clinical workflow data across three tasks: infarct classification, meningioma segmentation, and brain age regression, comparing nineteen models from sixteen teams. Self-supervised pretraining improved performance on clinical data experiencing domain shift. Different pretraining approaches suited different tasks, masked autoencoders for segmentation and hybrid objectives for classification, and smaller pretrained models achieved strong results while scaling provided unreliable benefits.
Paper
2025
An AI-Powered Data Curation and Publishing Virtual Assistant
Ensuring high quality and reusability of personal health data is costly and time-consuming; an AI-powered virtual assistant for health data curation and publishing could help patients improve harmonization, quality, and interoperability. This formative evaluation study assessed the usability of the first-generation (G1) prototype from the AIDAVA Horizon Europe project, recruiting patients with breast cancer and cardiovascular disease from three European countries. An intuitive front-end, supported by AI and non-AI curation tools and developed across two generations, let patients ingest and curate their personal health data into a personal health knowledge graph representing their integrated, high-quality medical records. Usability of G1 was assessed using the system usability scale, informing the design of the next-generation prototype.
Paper
2025
An Extensive Generative Framework for Synthesising CT Images of Diverse Anatomies
The conventional approach of testing the safety and efficacy of medical devices/drugs is done through clinical trials. These trials are time-consuming, expensive and pose ethical issues. In Silico Clinical Trials (ISCTs) leverage computer simulations to assess medical devices or drugs by replicating physiological processes. Acquiring large datasets for ISCTs is challenging due to legal and ethical constraints, but synthetic data generation offers a solution. In this study, we present a widely applicable generative framework which has the capability to synthesise CT images of visceral organs and hard tissues, thereby demonstrating its extensiveness across various anatomical structures. In addition to data synthesis, we show that synthesised images are comparable to real images in terms of quality metrics and expert evaluations (data-level verification). Hence, this research shows the feasibility to use generative frameworks for clinically realistic data synthesis. Furthermore, we introduce a multiphasic verification strategy with an additional verification step (model-level verification) for prospective utilisation for ISCTs.
Paper
2025
Enhancing Reconstruction-Based Out-of-Distribution Detection in Brain MRI
Out-of-distribution (OOD) detection is crucial for safely deploying automated medical image analysis, since abnormal patterns can hamper performance, yet OOD detection in medical imaging remains an open challenge. This study optimizes a reconstruction-based autoencoder specifically for OOD detection, addressing the underexplored potential of a simple model, the lack of deep learning strategies tuned for the task, and the choice of reconstruction metric. We evaluated the model's reconstruction capability, the impact of training epoch and metric choice, and model and metric ensembles, for unsupervised detection of synthetic local and global artifacts in brain MRI. Learned perceptual image patch similarity (LPIPS) and the contrast component of SSIM consistently outperformed other metrics, and combining two well-converged models with LPIPS and contrast reached a pixel-level area under the Precision-Recall curve of 0.66. On a more realistic OOD dataset, local artifacts proved harder to detect than global ones, underscoring the need for tailored, carefully validated approaches.
Paper
2025
Explainable Anatomy-Guided AI for Prostate MRI: Virtual Biopsy Risk Assessment
This work presents a fully automated, anatomically guided deep learning pipeline for prostate cancer risk stratification from routine MRI, integrating an nnU-Net module for segmenting the prostate gland and its zones, a classification module based on the UMedPT Swin Transformer foundation model fine-tuned with anatomical priors and clinical data, and a VAE-GAN framework generating counterfactual heatmaps that localize decision-driving regions. Developed on 1,500 PI-CAI cases for segmentation and 617 biparametric MRIs from the CHAIMELEON challenge for classification, segmentation reached Dice scores of 0.92 to 0.95 across zones, and a three-scale classification ensemble with anatomical priors reached an AUC of 0.79, outperforming the 2024 CHAIMELEON winners. In a prospective multi-center in-silico trial with 20 clinicians, AI assistance raised diagnostic accuracy from 0.72 to 0.77, raised inter-rater agreement from 0.43 to 0.53, and cut review time by 40%, supporting the pipeline's potential as a virtual biopsy in clinical practice.
Paper Code
2025
Generative AI for Synthetic Data Across Multiple Medical Modalities: A Systematic Review
This systematic review covers generative models (GANs, VAEs, diffusion models, and LLMs) used to synthesize medical data, spanning imaging (dermoscopic, mammographic, ultrasound, CT, MRI, X-ray), text, time-series, and tabular EHR data. Unlike prior narrowly focused reviews, it encompasses a broad array of modalities, offering insight into current and future applications across synthesis, generation techniques, and evaluation methods, alongside a GitHub repository as a living resource. The multi-modality scope surfaces cross-modality insights unavailable in single-modality surveys; core generative techniques are transferable, but synthesis methods often lack sufficient integration of patient-specific context and clinical knowledge. Findings are structured around three themes, synthesis applications, generation techniques, and evaluation methods, revealing an absence of standardized benchmarks and a need for large-scale, privacy-aware validation and cross-team benchmarking.
Paper
2025
Leveraging Hand-Crafted Radiomics on Multicenter FLAIR MRI for Predicting Disability Worsening in MS
Background: Multiple sclerosis (MS) is an autoimmune disease of the central nervous system, leading to varying degrees of functional impairment. Conventional tools, such as the Expanded Disability Status Scale (EDSS), lack sensitivity to subtle disease worsening. Radiomics provides a quantitative imaging approach to address this limitation. This study applied machine learning (ML) and radiomics features from T2-weighted Fluid-Attenuated Inversion Recovery (FLAIR) magnetic resonance imaging (MRI) to predict disability worsening in MS. Methods: A retrospective analysis was performed on real-world data from 247 PwMS across two centers. Disability worsening was defined as a change in EDSS over two years. FLAIR MRIs underwent preprocessing and super-resolution reconstruction to enhance low-resolution images. White matter lesions (WML) were segmented using the Lesion Segmentation Toolbox (LST), and tissue segmentation was performed using sequence Adaptive Multimodal Segmentation. Radiomics features from WML and normal-appearing white matter (NAWM) were extracted using Pyradiomics, harmonized with Longitudinal ComBat, followed by recursive feature elimination for feature selection.
Paper
2025
Privacy-preserving open-source LLMs for clinical CMR report analysis
We investigated the utilization of privacy-preserving, locally-deployed, open-source large language models to extract diagnostic information from free-text cardiovascular magnetic resonance (CMR) reports. Nine open-source LLMs were evaluated on their ability to identify diagnoses and classify patients into cardiac diagnostic categories based on descriptive findings in 109 clinical CMR reports, scored with accuracy, precision, recall, and F1, with confusion matrices used to examine misclassification patterns. Most models performed exceptionally well: Google's Gemma2 achieved the highest average F1 score of 0.98, followed by Qwen2.5:32B and DeepseekR1-32B at 0.96 and 0.95. All other evaluated models scored above 0.93, apart from Mistral and DeepseekR1-7B. The top four LLMs outperformed a board-certified cardiologist (F1 score of 0.94) across every evaluation metric. These findings demonstrate the feasibility of using open-source, privacy-preserving LLMs in clinical settings for automated, accurate, fast, and resource-efficient diagnostic categorization of imaging reports.
Paper Code
2025
Radiomics Quality Score 2.0: Towards Radiomics Readiness Levels and Clinical Translation
Radiomics could play a relevant role in precision oncology by offering precise quantitative support for clinical decision-making. The Radiomics Quality Score (RQS), developed to assess the rigour of radiomics studies, has been widely adopted by researchers since its 1.0 benchmark, but an updated framework is needed to account for evolving knowledge around fairness, explainability, quality control, and harmonization. This review introduces RQS 2.0, which keeps the scientific rigour of its predecessor while addressing these contemporary needs to potentially accelerate clinical translation. It also introduces radiomics readiness levels, inspired by the technology readiness level framework and integrated into RQS 2.0, reflecting nine distinct levels of incremental improvement toward clinical implementation, alongside a strategic vision for precision oncology's future direction.
Paper Website
2025
Report2CT: Radiology Report-Conditional 3D CT Generation
Text-to-image latent diffusion models have advanced medical image synthesis, but 3D CT generation remains limited, and existing approaches rely on simplified prompts that neglect the rich semantic detail in full radiology reports. Report2CT is a radiology report-conditional latent diffusion framework that synthesizes 3D chest CT volumes directly from free-text reports, integrating three pretrained medical text encoders (BiomedVLP CXR BERT, MedEmbed, and ClinicalBERT) to capture nuanced clinical context. The model was trained on 20,000 CT volumes from the CT-RATE dataset, conditioned on complete reports and voxel spacing. Evaluated with Frechet Inception Distance and CLIP-based semantic alignment against the GenerateCT baseline, Report2CT generated anatomically consistent volumes with strong text-image alignment; multi-encoder conditioning improved CLIP scores, and classifier-free guidance further enhanced alignment with only a minor FID trade-off. Report2CT ranked first in the VLM3D Challenge at MICCAI 2025 on Text-Conditional CT Generation, achieving state-of-the-art performance across all evaluation metrics.
Paper Code
2025
WAND: Wavelet Analysis-Based Neural Decomposition of MRS Signals for Artifact Removal
Accurate metabolite quantification in MR spectroscopy (MRS) is challenged by low signal-to-noise ratio, overlapping metabolites, and unknown, unparameterized baseline artifacts that obscure low-concentration metabolites. This paper introduces wavelet analysis-based neural decomposition (WAND), a data-driven method that decomposes MRS signals into metabolite-specific signals, baseline, and artifacts, exploiting how cleanly these components separate in the wavelet domain. A U-Net is trained to predict masks over wavelet coefficients from the continuous wavelet transform, isolating each component before inverse-transforming back to separated signals; the artifact mask is obtained by inverting the sum of all known signal masks, letting WAND capture and remove even artifacts it was never explicitly shown. WAND's accurate decomposition is demonstrated across both simulated and in vivo data.
Paper
2024
Automated Cardiovascular MR Myocardial Scar Quantification with Unsupervised Domain Adaptation
Quantification of myocardial scar from late gadolinium enhancement (LGE) CMR images can be facilitated by automated AI-based analysis, but AI models are susceptible to domain shifts that degrade performance on data with different characteristics than the training data. This study trained CycleGAN models to translate local hospital data to the appearance of a public LGE CMR dataset, then evaluated an existing AI scar-quantification pipeline (myocardium segmentation, scar segmentation, scar burden) on an external test set of 44 patients clinically assessed for ischemic scar, without retraining on local labels. Dice similarity between manual and AI-predicted segmentations was similar to previously reported results: 0.76 for myocardium and 0.75 for scar overall, 0.41 for scar in pathological scans, with a Bland-Altman mean bias in scar burden of -0.62%. This shows domain adaptation alone can carry an existing pipeline to a new site.
Paper
2024
Brain MR Image Simulation for Deep Learning Based Medical Image Analysis Networks
Training and validating deep-learning medical image analysis needs large annotated MRI datasets, and existing brain simulation data is limited in anatomical variety, tissue classes, and overall realism. This work proposes a realistic simulation framework combining patient-specific phantoms with Bloch-equation analytical solutions for fast, accurate MRI simulation, with ground-truth labels derived automatically from open-source high-resolution T1-weighted MRI via an automated brain classification tool. The framework generates 3D brain MRI across variable anatomy, sequence, contrast, SNR, and resolution, and a 3D brain segmentation network trained purely on this simulated data reached Dice scores of 0.818 to 0.832 on the real MRBrainS18 challenge dataset and 0.901 to 0.939 on OASIS data, closely matching the established FSL toolkit despite never training on a real image. This is an initial step toward physics-based MRI generation for training deep learning models without real annotated data.
Paper
2024
Deep Learning Synthesis of White-Blood from Dark-Blood Late Gadolinium Enhancement CMR
Dark-blood late gadolinium enhancement (DB-LGE) cardiac MR has been proposed as an alternative to standard white-blood LGE (WB-LGE) protocols to enhance scar-to-blood contrast, but acquiring both has the drawback of extra scan time. This study developed a deep learning method to generate synthetic WB-LGE images from DB-LGE, allowing assessment of both contrasts without additional scanning. DB-LGE and WB-LGE data from 215 patients trained two unpaired image-to-image translation models, CycleGAN and contrastive unpaired translation, across five loss-function settings each. CycleGAN produced the most realistic synthetic WB-LGE images: visual readers could barely distinguish real from synthetic (55% correctly classified), and scar-burden quantification from the synthetic images was highly correlated with the real WB-LGE, with a mean bias of 0.44%. Further clinical evaluation is warranted before adoption.
Paper
2024
Results of the 2023 ISBI Challenge to Reduce GABA-Edited MRS Acquisition Time
This conference challenge compared machine-learning reconstruction models for gamma-aminobutyric acid (GABA)-edited MR spectroscopy using only one-quarter of the transients typically required for a complete scan, across three tracks: simulated data, in vivo data with matched acquisition parameters, and in vivo data with different parameters. Performance was scored on mean squared error, signal-to-noise ratio, linewidth, and a proposed shape-score metric, using an openly released baseline model and simulated data. Of three compared submissions, a covariance-matrix CNN performed best on simulated data, while a vision transformer on a spectrogram representation performed best on both in vivo tracks. Reconstructions from just 80 transients matched or beat conventional 320-transient results on SNR, linewidth, and fit error, though some models optimized those metrics without truly improving spectral quality, pointing to a need for more robust evaluation metrics.
Paper
2023
Cardiac MRI Synthesis for Reducing Segmentation Failures
Deep-learning cardiac MRI segmentation methods show great promise but are rarely applicable to realistic clinical scenarios, since most are trained on largely homogeneous datasets without the acquisition variation typical of multi-vendor, multi-site, and pathological data, and degrade on outlier cases with difficult pathologies, artifacts, or extensive changes in tissue shape and appearance. This work presents a model for segmenting all three cardiac structures in a multi-center, multi-disease, and multi-view scenario, addressing the different challenges of such heterogeneous data through a pipeline of heart-region detection, augmentation via image synthesis, and a late-fusion segmentation approach. Extensive experiments demonstrate the ability of the approach to handle outlier cases during both training and testing, allowing better adaptation to unseen and difficult examples. Reducing segmentation failures on outlier cases improves not only average segmentation performance but also the estimation of downstream clinical parameters, giving more consistent derived metrics.
Paper
2023
Cardiac Pathology Synthesis
We propose a method for synthesizing cardiac MR images with plausible heart pathologies and realistic appearances, to generate labeled data for supervised deep-learning training. Synthesis combines label deformation and label-to-image translation: the former via latent-space interpolation in a VAE model, the latter via a label-conditional GAN model. We devise three approaches to label manipulation in the VAE's latent space: intra- subject synthesis to interpolate intermediate slices and increase through-plane resolution, inter-subject synthesis to interpolate geometry and appearance between dissimilar subjects from different scanner vendors, and pathology synthesis to generate pseudo-pathological subjects with characteristics of a desired heart disease. We also model the relationship between 2D slices in the VAE's latent space prior to reconstruction, generating 3D-consistent subjects from stacked 2D slice-by-slice generations. This diversifies and enriches an available cardiac MR database, and we quantitatively show it improves generalization and robustness to multi-vendor, multi-disease data for image segmentation.
Paper Code
2018 - 2023
Simulation and Synthesis for Cardiac Magnetic Resonance Image Analysis
My PhD thesis, defended at Eindhoven University of Technology within the OpenGTN Marie Curie ITN-EID project under Prof. Dr. Ir. Marcel Breeuwer, develops methods for generating realistic cardiac MR images with corresponding ground-truth labels for supervised segmentation, addressing the scarcity of labelled cardiac MRI data. It connects two directions: physics-driven image simulation grounded in MR image-formation physics and the Bloch equations, published as a flexible framework for simulating cardiac MR images with variable anatomy and contrast, and data-driven image synthesis using conditional deep generative models, published as a label-conditional GAN framework and a VAE-based pathology synthesis method. Across both directions, the thesis shows that simulated and synthesized training data can match or improve the performance of deep-learning segmentation models trained on real data alone, particularly under limited or homogeneous training sets.
Paper
2022
Cardiac MR Scar Quantification
The clinical utility of late gadolinium enhancement (LGE) cardiac MRI is limited by a lack of standardization and time-consuming postprocessing. This work tests whether a cascaded deep learning pipeline, trained with augmentation by synthetically generated data, improves model accuracy and robustness for automated scar quantification. A bounding-box regression network first identifies a region of interest around the left-ventricular myocardium, then two further nnU-Net models segment the myocardium and any scar, trained on the EMIDEC challenge data supplemented with an extensive synthetic dataset from a conditional GAN. The cascaded pipeline significantly outperformed a single nnU-Net segmenting both structures directly (myocardium Dice 0.84 vs 0.63; scar Dice 0.72 vs 0.46, both p < 0.01). Adding synthetic data as augmentation improved scar Dice by a further 0.06. On the challenge test set, the augmented cascaded pipeline reached a mean per-subject Dice of 0.86 for myocardium and 0.67 for scar, comparable to a manual operator.
Paper
2022
Cardiac MRI Simulation Framework
One limiting factor for deep-learning medical image analysis is the scarcity of labeled data, which simulation can address by generating ample training data with ground-truth labels. This work develops a flexible framework for simulating cardiac MR images with variable anatomical and imaging characteristics, aimed at creating a diversified virtual population. Parameters are defined to alter anatomy, assign MR tissue properties to different tissue types, and manipulate image contrast through acquisition parameters. A database of virtual subjects was simulated and evaluated for aiding a deep-learning segmentation method. Training entirely on simulated images performed comparably to training on real images for heart-cavity segmentation in mid-ventricular slices. Moreover, simulated data could be combined with classical augmentation to boost performance when training data is limited, particularly by increasing contrast and anatomical variation, leading to better regularization and generalization. The database and simulation code are public.
Paper Code
2022
Conditional Cardiac MR Image Synthesis
Synthesizing high-quality medical images with variability in anatomy and appearance can help address the scarcity of properly annotated data in medical image analysis research. We propose a framework combining segmentation and synthesis based on mask-conditional GANs to generate high-fidelity, diverse cardiac MR images: a segmentation module trained on a physics-based simulated database provides multi-tissue labels on real CMR images, and a synthesis module trained on paired real images and labels translates masks into realistic cardiac images, so anatomy follows the labels while appearance is learned from training data. We study the effects of tissue-label count, training-data quantity, and multi-vendor data on synthesis quality, and evaluate the synthetic data for cardiac cavity segmentation in replacement and augmentation scenarios. Replacement with only synthetic data matched real-data performance, and augmentation improved Dice score by up to 4% and reduced Hausdorff Distance by up to 40%.
Paper Code
2022
On the Usability of Synthetic Data
Deep-learning segmentation methods for cardiac MRI perform well on images from the same source as their training data, but degrade significantly on images from different scanners or sites, and acquiring large multi-site, multi-vendor training data is generally infeasible. This work explores using synthesized short-axis CMR images from a segmentation-informed conditional GAN, trained on paired real images and segmentation maps for the heart and surrounding tissue, to improve the robustness of heart- cavity segmentation models. A model trained solely on synthetic data only slightly underperformed one trained on real data, and combining real with synthetic data during training substantially improved segmentation performance, up to 4% in Dice score and 40% in Hausdorff distance, across multiple datasets from different sites and scanners, and across both 2D and 3D segmentation networks. The work also analyzes the quality of the synthetic data and its ability to replace real MR images during training.
Paper
2021
Influence of Image Artifacts on Simulations of Cardiac Electrophysiology
Myocardial infarct patients face an increased risk of scar-based ventricular tachycardia, and computational electrophysiological models built from late gadolinium enhanced MR images can provide a personalized prediction of that risk. This work investigates how respiratory slice-alignment image artifacts affect such image-based simulations, using two model series: one from a clinical MR image with slice translations applied to artificially induce and correct misalignment, and one from computer-simulated MR images with and without slice misalignment, built from a mechanistic anatomical torso phantom. Personalized models built from these images had electrical stimuli applied to attempt to induce tachycardia, and the response of aligned versus misaligned models to different stimuli was used to assess risk. Slice misalignments were found to affect simulation outcomes, with the extent depending on infarct geometry: the number of inducible tachycardia patterns and stimuli locations varied with misalignment severity, and channel-dependent tachycardias could disappear entirely if the conducting channel was lost from the image.
Paper Code