Computational Radiology |

Improving diagnosis and interpretability in chest X-rays through federated learning and adversarial training

Thu, 01 Oct 2026 00:00:00 +0000

Reducing Clinician Annotation Fatigue in Open-Set Federated Learning: A Client-Adaptive Vision-Language Gatekeeper

Thu, 01 Oct 2026 00:00:00 +0000

K-MaT: Knowledge-Anchored Manifold Transport for Cross-Modal Prompt Learning in Medical Imaging

Sun, 27 Sep 2026 00:00:00 +0000

PromptGate: Client-Adaptive Vision–Language Gating for Open-Set Federated Active Learning

Sun, 27 Sep 2026 00:00:00 +0000

Eyelid Tumor Triage using Uncertainty-Aware and Explainable Artificial Intelligence

Fri, 25 Sep 2026 00:00:00 +0000

FIDAL: Diversity-Aware Federated Active Learning Under Real-World Distribution Shifts

Thu, 10 Sep 2026 00:00:00 +0000

When multi-institutional dataset aggregation masks shortcut-like behavior in Foundation Models for medical imaging

Tue, 01 Sep 2026 00:00:00 +0000

MRIQT – KI-gestützte Verbesserung der Bildqualität in der neonatalen Ultra-Niedrigfeld-MRT

Mon, 01 Jun 2026 00:00:00 +0000

Federated CT foundation models for multi-center detection of lymph node metastasis in pancreatic cancer

Thu, 09 Apr 2026 00:00:00 +0000

KI-gestützte datenschutzkonforme multizentrische Vorhersage lymphogener Metastasierung beim Pankreaskarzinom mittels eines föderierter CT-Foundation-Modelle

Thu, 01 Jan 2026 00:00:00 +0000

ProSona: Prompt-Guided Personalization for Multi-Expert Medical Image Segmentation

Thu, 01 Jan 2026 00:00:00 +0000

MRIQT: Physics-Aware Diffusion Model for Image Quality Transfer in Neonatal Ultra-Low-Field MRI

Mon, 17 Nov 2025 00:00:00 +0000

Multi-center Collaboration for Improved Lymph Node Metastasis Detection in Pancreatic Ductal Adenocarcinoma

Tue, 03 Jun 2025 00:00:00 +0000

AI in Radiology – From Pixels to Decisions (joint teaching with GUC)

Wed, 01 Jan 2025 00:00:00 +0000

This joint teaching activity was developed within a BMFTR-funded international program to strengthen academic exchange and capacity building in artificial intelligence for medical imaging. The course combines lectures, hands-on exercises, and project-based learning, enabling students to work on clinically motivated imaging problems using AI methods.

Perspective

The initiative contributes to sustainable academic exchange between Germany and Egypt and provides a basis for continued joint teaching activities, student exchange, and integration of research-led topics into international curricula.

Selected Interdisciplinary Research with Methodological AI Contributions

In addition to my core research in artificial intelligence for medical imaging and clinical applications, I have contributed to a small number of exploratory interdisciplinary projects in materials science and food science. These activities focused on transferring methodological expertise in data-driven modeling and machine learning to neighboring domains and align with the University of Bonn’s Transdisciplinary Research Areas Sustainable Futures, Modelling, and Life and Health. Together, they demonstrate the transferability of AI methodologies and my engagement in transdisciplinary research initiatives without detracting from my primary research focus.

Bias and generalizability of foundation models across datasets in breast mammography

Wed, 01 Jan 2025 00:00:00 +0000

Breast Mammography Dataset (Lebanon & international cohorts)

Wed, 01 Jan 2025 00:00:00 +0000

Partner: Dr. Charbel Mourad. Basis of our MICCAI 2025 study on bias and generalisability of foundation models across mammography datasets.

Bridging Regulatory Science and Medical Imaging Evaluation; and Distributed, Collaborative, and Federated Learning: First International Workshop, BRIDGE 2025, and 6th International Workshop, DeCaF 2025, Held in Conjunction with MICCAI 2025, Daejeon, South Korea, September 23 and September 27, 2025, Proceedings

Wed, 01 Jan 2025 00:00:00 +0000

Deep Learning for Lymph Node Metastasis Detection in Pancreatic Ductal Adenocarcinoma

Wed, 01 Jan 2025 00:00:00 +0000

Enhancing patient-specific deep learning based segmentation for abdominal magnetic resonance imaging-guided radiation therapy: a framework conditioned on prior segmentation

Wed, 01 Jan 2025 00:00:00 +0000

FLamby — Federated Learning Benchmark for Healthcare

Wed, 01 Jan 2025 00:00:00 +0000

International collaboration led by Owkin (Jean Ogier du Terrail) with EPFL, Inria, IP Paris, USC and UC Berkeley. Our lab contributed dataset preparation, benchmark design and federated evaluation.

Implementation and clinical evaluation of an in-house thoracic auto-segmentation model for 0.35 T magnetic resonance imaging guided radiotherapy

Wed, 01 Jan 2025 00:00:00 +0000

Leveraging Collective Intelligence for Inclusive Global Healthcare

Wed, 01 Jan 2025 00:00:00 +0000

MR-Linac Adaptive Radiotherapy MRI Dataset

Wed, 01 Jan 2025 00:00:00 +0000

Partners: Prof. Guillaume Landry, Dr. Christopher Kurz (LMU University Hospital) and Prof. Nassir Navab (TUM), with University Hospital Düsseldorf, Heidelberg, Zurich and Gemelli Rome. Basis of CloverNet and ProSona.

Musculoskeletal Photon-Counting CT Dataset

Wed, 01 Jan 2025 00:00:00 +0000

Acquired at the Department of Radiology, UKB, on photon-counting CT. Used together with labelled CT data from NAIST to study domain shift and federated learning feasibility.

Pancreatic Ductal Adenocarcinoma Multi-Center CT Dataset

Wed, 01 Jan 2025 00:00:00 +0000

Partners: Patrick Kupczyk (Radiology) and Alexander Semaan (Surgery), University Hospital Bonn, with Charité – Universitätsmedizin Berlin and UMG Göttingen. Harmonised across sites with a federated evaluation setup, so that data never leave the hospitals.

Prostate MRI Dataset with MR Fingerprinting

Wed, 01 Jan 2025 00:00:00 +0000

Curated with Dr. Barbara Wichtmann and Prof. Julian Luetkens (Radiology, UKB); extended with an external cohort from the Peter MacCallum Cancer Centre, Melbourne, and public data (PI-CAI).

MR Fingerprinting and Apparent Diffusion Coefficient Mapping for Quantitative Characterization of Clinically Significant Prostate Cancer

Sat, 04 May 2024 00:00:00 +0000

Bias-aware foundation models for equitable breast cancer screening

Mon, 01 Jan 2024 00:00:00 +0000

This project addresses fairness and generalization challenges of foundation models for breast mammography, where AI systems trained predominantly on data from high-income countries often underperform in underrepresented populations. Together with international clinical partners, we assembled and analyzed a heterogeneous collection of mammography datasets from Europe, North America, the Middle East, and Asia, including a biopsy-confirmed in-house cohort of approximately 1,500 patients from Lebanon. Methodologically, we conducted a systematic evaluation of foundation models under cross-dataset shift and demonstrated that naïve dataset aggregation improves average performance but does not eliminate subgroup-specific biases. Fairness-aware optimization strategies substantially reduced performance disparities across age groups, breast density categories, and geographic regions while maintaining competitive diagnostic accuracy.

Perspective

The collaboration constitutes an ongoing international research effort at the intersection of medical imaging, fairness-aware AI, and global health. Building on current results, the project is being expanded to additional clinical partners in Palestine and Egypt, supported by the Global One Health Research Academy (GLOHRA), and is planned to scale further through the submitted EU Horizon Europe project SHE-Scan, contingent on funding.

CloverNet--Leveraging Planning Annotations for Enhanced Procedural MR Segmentation: An Application to Adaptive Radiation Therapy

Mon, 01 Jan 2024 00:00:00 +0000

Robust Federated Learning Methodology for Automatic Chest X-Ray Diagnosis

Mon, 01 Jan 2024 00:00:00 +0000

Two-Steps Approach for Breast Cancer Detection and Classification Using Convolutional Neural Networks.

Mon, 01 Jan 2024 00:00:00 +0000

Semi-Supervised Learning for Prostate Lesion Detection and Segmentation

Wed, 01 Nov 2023 00:00:00 +0000

Ascertaining the pose of an x-ray unit relative to an object on the basis of a digital model of the object

Mon, 01 May 2023 00:00:00 +0000

Federated foundation models for multi-center CT staging of pancreatic cancer

Mon, 01 May 2023 00:00:00 +0000

Collaborators: University Medical Center Göttingen, Charité – Universitätsmedizin Berlin, Peter MacCallum Cancer Centre

This project addresses preoperative staging of pancreatic ductal adenocarcinoma (PDAC), where lymph node metastasis is a major prognostic factor but CT-based assessment suffers from low sensitivity and high inter-observer variability. Together with partners across three German centers, we established a retrospective multi-center cohort of 546 patients to enable evaluation under real-world heterogeneity while preserving data privacy. Methodologically, we combined CT foundation model transfer learning (pre-training on 148,000 volumetric CT scans) with heterogeneity-aware federated optimization. The resulting system improved federated performance by +12.6% balanced accuracy over FedAvg and recovered most of the centralized reference performance (centralized BA 0.601; DOR 3.45).

Perspective

The project establishes a sustainable multi-center collaboration for privacy-preserving oncologic imaging in Germany. Ongoing work extends the framework toward multimodal integration and more robust harmonization strategies, and a joint DFG proposal with the clinical partners has been submitted to further advance federated foundation models for PDAC staging and translation.

Deep learning for clinically significant prostate cancer in 3D mpMRI and MR Fingerprinting

Sun, 01 Jan 2023 00:00:00 +0000

Collaborators: Peter MacCallum Cancer Centre, Melbourne

This project addresses the detection and characterization of clinically significant prostate cancer (csPCa) in MRI, where diagnostic performance is limited by inter-reader variability and the complexity of multi-parametric, volumetric imaging. Together with the clinical partners, we curated an initial in-house 3D MRI cohort of 106 patients with histopathological reference, including bi-parametric MRI and quantitative MR Fingerprinting (MRF), which was subsequently extended with external and public datasets to form a multi-center cohort of approximately 500 cases. Methodologically, we developed fully 3D deep learning models that explicitly incorporate anatomical prostate zone information, enabling interpretable and biomarker-aware csPCa detection and segmentation beyond slice-wise analysis.

A dedicated sub-project evaluated transformer-based 3D U-Net architectures for lesion segmentation. On public (PI-CAI) and in-house datasets, transformer-based models achieved Dice scores above 0.60 on detected lesions and demonstrated improved robustness under reduced training data compared to CNN-based baselines, while highlighting persistent challenges in false-positive reduction.

Perspective

The project established a scalable research line on interpretable, anatomy-aware AI for prostate cancer imaging and contributed to early-career training through a completed Bachelor’s thesis and an ISMRM abstract. Ongoing work extends this line toward vision–language foundation models integrating 3D imaging, quantitative MRI biomarkers, and anatomical priors, forming a basis for continued multi-center collaboration and future third-party funding.

Deep learning to estimate aging from chest imaging

Sun, 01 Jan 2023 00:00:00 +0000

Chronologic age is an important risk factor for many chronic diseases. However, as we all age at different rates chronological age is an imperfect measure of aging and more accurate estimates of one’s true biological age are desirable. The proposed study will develop and test artificial intelligence-based measures of aging from chest imaging data and will explore whether these measures can improve chronologic age-based clinical guidelines for prevention of cardiovascular disease and cancer. Over recent decades life expectancy has continuously increased, however, the years of life without significant morbidity have remained stable over the last 5 decades resulting in significant socioeconomic burden to society and the healthcare system. Therefore, delaying the onset of morbidity is a promising way to improve healthy aging. In particular, prevention and early detection of diseases of high socioeconomic importance such as cancer and cardiovascular disease are a priority to increase health and longevity. Biological age is a concept to estimate the differences in rates of aging not captured by chronologic age. Several measures of biological age have been developed in recent years including blood, functional and physiological approaches. However, there relevance in clinical routine is limited. Medical imaging may be an effective way to estimate aging by measuring age-related changes based on anatomical alterations visible in the image such as degenerative changes of the spine, dilation of the heart and vasculature and changes in lung parenchyma. Unlike molecular measures, image-based reflections of aging can be calculated opportunistically using data acquired during routine clinical care. Furthermore, these measures can estimate aging of specific organ systems, which may improve disease-specific preventive measures over general estimates of biological age. The image-based estimates of aging using deep learning proposed in this study will address this unmet need and may help to personalize clinical decision-making for chronic disease screening and prevention. My preliminary data show that deep learning can estimate a chest x-ray age from a chest radiograph image and that this chest x-ray age predicts longevity better than chronologic age in large multicenter clinical trial data. The proposed project will build on these promising initial findings with the potential to broaden the application to other imaging modalities and to leverage multi-omics data to infer causes of aging.

Collaborators:

PD. Dr. , UniKlinikum Freiburg

Outcomes

This exploratory project examined the feasibility of imaging-based phenotyping and biological aging estimation from chest CT and MRI. The work focused on conceptual development and clinical hypothesis generation rather than large-scale validation and resulted in a clinical abstract presented at the European Congress of Radiology.

Perspective

The project provided methodological insights into opportunistic imaging-based risk modeling and informed the design of subsequent projects with clearer data availability, validation strategies, and funding perspectives.

Enhancing oncological care: A guide to setting up a new multidisciplinary cancer cachexia clinic within a tertiary centre

Sun, 01 Jan 2023 00:00:00 +0000

Guest Editorial Special Issue on Federated Learning for Medical Imaging: Enabling Collaborative Development of Robust AI Models

Sun, 01 Jan 2023 00:00:00 +0000

Joint self-supervised image-volume representation learning with intra-inter contrastive clustering

Sun, 01 Jan 2023 00:00:00 +0000

Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching

Sun, 01 Jan 2023 00:00:00 +0000

Open Medical Inference (OMI)

Sun, 01 Jan 2023 00:00:00 +0000

Outcomes

The Open Medical Inference (OMI) initiative is a BMFTR-funded national research infrastructure aimed at improving reproducibility, transparency, and comparability of machine learning methods in medical imaging and clinical AI. Within OMI, I contributed to the development and evaluation of standardized benchmarking protocols, open evaluation pipelines, and reference datasets for medical AI, with a particular focus on robustness, generalization, and fair comparison across institutions and data distributions. These activities complement my research on federated learning and trustworthy AI by addressing methodological gaps in evaluation practices and supporting community-wide standards for medical machine learning research.

Our role

Contributing researcher (Scientific Lead) with responsibility for methodological input on benchmarking design, evaluation protocols, and reproducibility aspects, in close collaboration with national partners.

Perspective

OMI provides a sustained national framework for open and reproducible medical AI research. The experience and infrastructure developed within OMI inform ongoing and future projects on federated learning, multi-center evaluation, and trustworthy AI, and strengthen national coordination across academic and clinical institutions. These activities align with and complement my work on federated learning benchmarks and multi-center evaluation (see also Sections and ).

Transformer based models for unsupervised anomaly segmentation in brain MR images

Sun, 01 Jan 2023 00:00:00 +0000

Utilising Artificial Intelligence in Cancer Imaging To Improve Patient Outcomes

Sun, 01 Jan 2023 00:00:00 +0000

Cachexia is a devastating tumour-induced wasting disease affecting 30% of all cancerpatients. This initiative brings together clinical, computational (artifi cial intelligence),and laboratory research partners to work on the delivery of improved outcomes forpatients with cancer cachexia. Aims of the collaboration: Funding will provideopportunities to broaden digital research links and submit joint grant and conjoint PhDproposals.

Long-term objectives:

1. Developing non-invasive screening tools for cancercachexia, using AI-assisted imaging analysis approaches.
1. Developing a greaterunderstanding of molecular processes of cancer cachexia using animal models andpatient samples.
1. Increase knowledge and awareness about cachexia througheducational events for clinicians, patients and families on cachexia diagnosis andtreatment.

This multidisciplinary and diverse initiative builds on the expertise andexisting collaborations between researchers/clinicians across both universities (Bonn/Melbourne). The group consists of experts in radiology/ image analysis (Ko,Attenberger), computational science /AI (Albarqouni) and cancer biology/ metabolism(Cheng). This team of clinicians and researchers with deep knowledge of the fi eld ofcancer cachexia are ideally placed to educate clinicians, patients and families oncachexia and increase its awareness in the general community.

Partners:

, Australia
, Germany

Outcomes

This collaboration was supported by institutional funding to foster international exchange at the intersection of artificial intelligence, cancer imaging, and translational oncology. Core outcomes included two jointly organized international workshops held in Bonn and Melbourne, reciprocal research visits, and the identification of shared research directions. The collaboration further supported concrete synergies in ongoing projects, including pancreatic cancer staging and multiparametric prostate MRI.

Perspective

The project established a durable international partnership between the University of Bonn and the University of Melbourne, providing a platform for continued joint research, doctoral training, and future collaborative funding applications.

Federated learning for musculoskeletal segmentation in CT

Tue, 01 Nov 2022 00:00:00 +0000

This collaboration explored federated learning for musculoskeletal segmentation in CT imaging across institutions in Germany and Japan. Using labeled CT data from NAIST and an unlabeled whole-body CT cohort from University Hospital Bonn, we evaluated cross-site generalization and domain transfer under realistic scanner and population shifts. In parallel, a secure federated learning infrastructure was deployed and ethical approval obtained at University Hospital Bonn. While the project did not progress to publication-ready federated training due to annotation constraints and deployment complexity, it provided valuable methodological and organizational insights.

Perspective

The collaboration strengthened institutional expertise in federated learning under real-world clinical constraints and informed subsequent national and international projects with more harmonized data availability.

Federated disentangled representation learning for unsupervised brain anomaly detection

Sat, 01 Oct 2022 00:00:00 +0000

CTFSH: Full Head CT Anomaly Detection with Unsupervised Learning

Sat, 01 Jan 2022 00:00:00 +0000

Development of a Deep Learning Toolkit for MRI-Guided Online Adaptive Radiotherapy

Sat, 01 Jan 2022 00:00:00 +0000

In Germany close to 500,000 new patients are diagnosed with cancer every year. More than half of the patients undergo radiotherapy as part of their treatment. In-room imaging, in combination with modern external beam radiotherapy techniques, enables tight adjustment of the delivered dose to the target. However, the full potential of these techniques is currently not exploited. One of the reasons for this is the presence of anatomical changes. For the vast majority of patients in today’s radiotherapy, anatomical alterations are considered by introducing safety margins around the tumor volume during treatment planning, which increases the irradiated volume, the dose burden to organs at risk (OARs), and eventually limits the applicable dose.

An improved treatment can be realized by online adaptive radiation therapy (ART): instead of applying the same irradiation plan throughout the entire course of treatment, the treatment is optimized at each irradiation session on basis of the daily anatomy in treatment position, as inferred from in-room imaging. Great efforts have been made to realize the integration of magnetic resonance imaging (MRI) as in-room imaging modality, but only during the last few years integrated MR-linacs became clinically available at few academic institutions. The superior soft-tissue contrast allows for accurate visualization of targets and OARs, which has allowed basic online ART workflows to be implemented. The time overhead associated with the generation of a synthetic-CT from MRI and a re-delineation of the patient for treatment plan optimization greatly limits applicability to all patients.

The goal of this project is to tackle the main bottlenecks of online plan adaptation: synthetic CT generation and organ delineation. By combining our respective expertise in radiotherapy and artificial intelligence, we aim at proposing deep-learning-based solutions to build a toolkit to tackle both of these challenges. In the case of synthetic CT generation, we will use generative adversarial networks with hybrid paired-unpaired training paradigms, additionally exploiting novel embedding loss functions to ensure geometric fidelity. Alternatively, disentangled representations will be investigated. For segmentation we aim at exploiting the information from previous fractions by performing spatio-temporal segmentation using longitudinal spatial memory networks. Finally, we aim at evaluating the benefits from these approaches by performing an evaluation of their clinical impact in terms of plan optimization, including physician grading of delineations and robustness assessment. The time benefits, including potential corrections to segmentations and synthetic CT, will also be quantified. Further details can be found .

Outcomes

MRI-guided online adaptive radiotherapy requires fast and reliable image segmentation and synthesis to support treatment adaptation under strict clinical time constraints. In this DFG-funded project, we addressed key methodological challenges related to patient-specific segmentation and uncertainty-aware modeling under inter-observer and inter-institutional variability. Together with the partners, we curated a longitudinal multi-center MRI cohort comprising planning and fraction scans across multiple institutions, enabling systematic evaluation under realistic clinical heterogeneity. Methodologically, the project resulted in two complementary contributions: CloverNet, a patient-specific 3D segmentation architecture leveraging prior planning annotations, and ProSona, a prompt-guided personalization framework modeling inter-expert variability via a continuous latent space. The proposed methods achieved a significant improvement over standard baselines and received a Best Paper Award at a MICCAI workshop.

Perspective

The project established a strong national research network at the interface of medical imaging, radiation oncology, and AI, and contributed to doctoral and master’s training. Ongoing work extends these approaches toward scalable foundation models for personalized and uncertainty-aware radiotherapy workflows, with a focus on clinical integration and multi-center validation.

Distributed, collaborative, and federated learning, and affordable AI and healthcare for resource diverse global health

Sat, 01 Jan 2022 00:00:00 +0000

FedNorm: Modality-Based Normalization in Federated Learning for Multi-Modal Liver Segmentation

Sat, 01 Jan 2022 00:00:00 +0000

Flamby: Datasets and benchmarks for cross-silo federated learning in realistic healthcare settings

Sat, 01 Jan 2022 00:00:00 +0000

Large vision models for robust and trustworthy medical image understanding

Sat, 01 Jan 2022 00:00:00 +0000

This collaboration investigates the applicability of large vision models to medical imaging, addressing robustness, generalization, and trustworthiness under domain shift and limited supervision. The work examined the behavior of foundation-scale models transferred from natural-image pretraining to medical imaging tasks. Methodological contributions include empirical analyses of cross-domain generalization, sensitivity to spurious correlations, and limitations of naïve scaling, alongside proposed evaluation protocols and architectural adaptations to improve robustness and reliability in medical contexts.

Perspective

The collaboration supported doctoral mentorship and strengthened national expertise at the intersection of foundation models and medical imaging. The insights gained inform ongoing work on principled evaluation and responsible deployment of large vision models in medical applications.

Semi-Supervised Federated Peer Learning for Skin Lesion Classification

Sat, 01 Jan 2022 00:00:00 +0000

SPA: Shape-Prior Variational Autoencoders for Unsupervised Brain Pathology Segmentation

Sat, 01 Jan 2022 00:00:00 +0000

FLamby: Federated learning benchmarks for realistic healthcare settings

Tue, 01 Jun 2021 00:00:00 +0000

Collaborators: EPFL (Switzerland), INRIA (France), University of Southern California (USA), Institut Polytechnique de Paris (France), University of California, Berkeley (USA)

A central bottleneck in federated learning research is the lack of realistic, multi-center benchmarks reflecting the heterogeneity of clinical data. Within this international collaboration, I contributed to the development of FLamby, a benchmark suite comprising seven real-world medical datasets with natural institutional splits across histopathology, radiology, dermatology, and tabular clinical data. Published at NeurIPS 2022 (Datasets and Benchmarks Track), FLamby demonstrated that standard federated learning approaches often underperform pooled training and highlighted the need for heterogeneity-aware optimization methods.

Our role

Contributing collaborator with responsibility for dataset preparation, benchmark design, and evaluation of federated learning methods.

Perspective

FLamby established a widely used community reference for cross-silo federated learning in healthcare. Its open-source release continues to inform benchmarking practices and supports subsequent federated learning studies and methodological developments.

Autoencoders for Unsupervised Anomaly Segmentation in Brain MR Images: A Comparative Study

Fri, 01 Jan 2021 00:00:00 +0000

Butterfly-Net: Spatial-Temporal Architecture For Medical Image Segmentation

Fri, 01 Jan 2021 00:00:00 +0000

FedDis: Disentangled Federated Learning for Unsupervised Brain Pathology Segmentation

Fri, 01 Jan 2021 00:00:00 +0000

FedPerl: Semi-Supervised Peer Learning for Skin Lesion Classification

Fri, 01 Jan 2021 00:00:00 +0000

Modeling Healthy Anatomy with Artificial Intelligence for Unsupervised Anomaly Detection in Brain MRI

Fri, 01 Jan 2021 00:00:00 +0000

The Federated Tumor Segmentation (FeTS) Challenge

Fri, 01 Jan 2021 00:00:00 +0000

The OOD Blind Spot of Unsupervised Anomaly Detection

Fri, 01 Jan 2021 00:00:00 +0000

Unsupervised anomaly detection in full-head CT for emergency triage

Fri, 01 Jan 2021 00:00:00 +0000

In the emergency room, head CT scans must be read and prioritised quickly — and the findings range from bleeds to fractures. Supervised detectors need labelled examples of every pathology; unsupervised anomaly detection (UAD) instead learns what normal anatomy looks like and flags everything else, an inexpensive way to bring triage to the clinical workflow. Most UAD work, however, addressed 2D slices of brain MRI rather than volumetric CT.

What we did

UAD for volumetric head CT. We compared two widely used reconstruction-based baselines on full 3D head CT, and improved their reconstruction quality with two recent computer-vision contributions — raising anomaly-detection AUROC from 0.70 to 0.77 for one baseline.
Keep the skull. Testing on skull-stripped and unstripped volumes showed that leaving the skull in lets the model detect fractures too — to our knowledge the first comparison of this kind.
Open labels and code. Supplementary labels for the public CQ500 head-CT dataset and the were released to support further research.

Our role

Academic lead and supervision, together with Daniel Rueckert (TU Munich) and Mehmet Yigitsoy and Abhijeet Parida (deepc GmbH).

An Uncertainty-Driven GCN Refinement Strategy for Organ Segmentation

Tue, 01 Dec 2020 00:00:00 +0000

The Future of Digital Health with Federated Learning

Mon, 14 Sep 2020 00:00:00 +0000

Deep Federated Learning in Healthcare

Wed, 01 Apr 2020 00:00:00 +0000

Deep Learning (DL) has emerged as a leading technology for accomplishing many challenging tasks showing outstanding performance in a broad range of computer vision and medical applications. However, this success comes at the cost of collecting and processing a massive amount of data, which often are not accessible, in Healthcare, due to privacy issues. Federated Learning (FL) has been recently introduced to allow training DL models without sharing the data. Instead, DL models at local hubs, i.e. hospitals, share only the trained parameters with a centralized DL model, which is, in return, responsible for updating the local DL models as well.

Our golas in this project is to develop novel models and algorithms for a ground-breaking new generation of deep FL, which can distill the knowledge from local hubs, i.e. hospitals, and edges, i.e. wearable devices, to provide personalized healthcare services.

The principal challenges, to overcome, concern the nature of medical data, namely data heterogeneity; severe class-imbalance, few amounts of annotated data, inter-/intra-scanners variability (domain shift), inter-/intra-observer variability (noisy annotations), system heterogeneity, and privacy issues (see the example below).

Collaboration:

Uncertainty-based graph convolutional networks for organ segmentation refinement

Wed, 01 Apr 2020 00:00:00 +0000

Bayesian Skip-Autoencoders for Unsupervised Hyperintense Anomaly Detection in High Resolution Brain Mri

Wed, 01 Jan 2020 00:00:00 +0000

Benefit of dual energy CT for lesion localization and classification with convolutional neural networks

Wed, 01 Jan 2020 00:00:00 +0000

Image-to-Images Translation for Multi-Task Organ Segmentation and Bone Suppression in Chest X-Ray Radiography

Wed, 01 Jan 2020 00:00:00 +0000

Inverse Distance Aggregation for Federated Learning with Non-IID Data

Wed, 01 Jan 2020 00:00:00 +0000

Liver lesion localisation and classification with convolutional neural networks: a comparison between conventional and spectral computed tomography

Wed, 01 Jan 2020 00:00:00 +0000

Modelling Uncertainty in Deep Learning for Medical Applications

Wed, 01 Jan 2020 00:00:00 +0000

Deep Learning has emerged as a leading technology for accomplishing many challenging tasks showing outstanding performance in a broad range of applications in computer vision and medical applications. Despite its success and merit in recent state-of-the-art methods, DL tools still lack in robustness hindering its adoption in medical applications. Modeling uncertainty, through Bayesian Inference and Monte-Carlo dropout, has been successfully introduced to computer vision for better understanding the underlying deep learning models. In this proposal, we investigate modeling the uncertainty for medical applications given the well-known challenges in medical image analysis, namely severe class-imbalance, few amounts of labeled data, domain shift, and noisy annotations.

On the Fairness of Privacy-Preserving Representations in Medical Applications

Wed, 01 Jan 2020 00:00:00 +0000

Precise proximal femur fracture classification for interactive training and surgical planning.

Wed, 01 Jan 2020 00:00:00 +0000

Scale-Space Autoencoders for Unsupervised Anomaly Segmentation in Brain MRI

Wed, 01 Jan 2020 00:00:00 +0000

SteGANomaly: Inhibiting CycleGAN Steganography for Unsupervised Anomaly Detection in Brain MRI

Wed, 01 Jan 2020 00:00:00 +0000

The future of digital health with federated learning. NPJ Digit Med. 2020; 3: 119

Wed, 01 Jan 2020 00:00:00 +0000

Adaptive image-feature learning for disease classification using inductive graph networks

Tue, 01 Jan 2019 00:00:00 +0000

Investigation of Focal Loss in Deep Learning Models For Femur Fractures Classification

Tue, 01 Jan 2019 00:00:00 +0000

Learn to estimate labels uncertainty for quality assurance

Tue, 01 Jan 2019 00:00:00 +0000

Learning Interpretable Features via Adversarially Robust Optimization

Tue, 01 Jan 2019 00:00:00 +0000

Learning-based x-ray image denoising utilizing model-based image simulations

Tue, 01 Jan 2019 00:00:00 +0000

MRI to CT Translation with GANs

Tue, 01 Jan 2019 00:00:00 +0000

Preliminary results of DSA denoising based on a weighted low-rank approach using an advanced neurovascular replication system

Tue, 01 Jan 2019 00:00:00 +0000

Self-attention equipped graph convolutions for disease prediction

Tue, 01 Jan 2019 00:00:00 +0000

Towards an Interactive and Interpretable CAD System to Support Proximal Femur Fracture Classification

Tue, 01 Jan 2019 00:00:00 +0000

Determination of the pose of an x-ray unit relative to an object using a digital model of the object

Mon, 01 Oct 2018 00:00:00 +0000

A photon recycling approach to the denoising of ultra-low dose X-ray sequences

Mon, 01 Jan 2018 00:00:00 +0000

Deep learning with synthetic data for free water elimination in diffusion MRI

Mon, 01 Jan 2018 00:00:00 +0000

Imaging intelligence: AI is transforming medical imaging across the imaging spectrum

Mon, 01 Jan 2018 00:00:00 +0000

Intraoperative stent segmentation in X-ray fluoroscopy for endovascular aortic repair

Mon, 01 Jan 2018 00:00:00 +0000

Multiple device segmentation for fluoroscopic imaging using multi-task learning

Mon, 01 Jan 2018 00:00:00 +0000

Weakly-supervised localization and classification of proximal femur fractures

Mon, 01 Jan 2018 00:00:00 +0000

X-ray Depthmaps: Revealing the Hidden Structures

Mon, 01 Jan 2018 00:00:00 +0000

Automatic classification of proximal femur fractures based on attention models

Sun, 01 Jan 2017 00:00:00 +0000

X-ray in-depth decomposition: Revealing the latent structures

Sun, 01 Jan 2017 00:00:00 +0000

X-Ray PoseNet: 6 DoF pose estimation for mobile X-Ray devices

Sun, 01 Jan 2017 00:00:00 +0000

Cathnets: detection and single-view depth prediction of catheter electrodes

Fri, 01 Jan 2016 00:00:00 +0000

Optimisation of X-ray images with respect to image quality and dose

Fri, 01 Jan 2016 00:00:00 +0000

X-ray guidance is at the heart of interventional procedures, but it exposes patients — and clinical staff, over a whole career — to ionising radiation. Following the “as low as reasonably achievable” principle, the dose can only be lowered if the image quality is kept, yet lower dose means fewer photons and noisier images, and at very low dose conventional denoisers oversmooth or create artefacts. Supervised deep learning would need aligned low- and high-dose image pairs, which cannot ethically be acquired during an intervention.

What we did

Photon recycling for ultra-low-dose sequences. A patch-based spatio-temporal denoiser with a properly parameterised Poisson–Gaussian noise model exploits non-local self-similarity across frames and low-rank approximation, without needing an initial estimate — better denoising at lower computational cost ( ).
Low-dose digital subtraction angiography. Spatially aligned DSA frames vary mainly through the flow of contrast agent, so a constrained, weighted rank-1 approximation of the temporal stack denoises each frame — evaluated with an advanced neurovascular replication system towards reducing both X-ray dose and contrast agent ( ).
Learning to denoise without clean images. A model-based simulation of realistic low-dose X-ray images — accounting for system gain, electronic noise and detector blur — generates training data from routine acquisitions, and the same imaging model normalises the input through noise-variance stabilisation. The resulting real-time network outperformed BM3D and WNNM on phantom and clinical data and was preferred by experts for sharp, artefact-free and realistic results ( ).

Our role

Scientific lead of the project as postdoc at TUM CAMP under Nassir Navab (PI and main supervisor), and mentoring of Sai Gokul Hariharan’s industrial PhD, run jointly with Norbert Strobel and Markus Kowarschik at Siemens Healthineers and completed at TU Munich in 2023.

Perspective

The thesis concluded that the X-ray dose can be reduced substantially without sacrificing clinically relevant information; the work continued at Siemens Healthineers in patents on X-ray noise reduction.

Single-view X-ray depth recovery: toward a novel concept for image-guided interventions

Fri, 01 Jan 2016 00:00:00 +0000