Computational Ophthalmology |

Histopathologische Klassifikation konjunktivaler melanozytärer intraepithelialer Läsionen (CMIL) mit einem Deep-Learning-Algorithmus

Sat, 26 Sep 2026 00:00:00 +0000

Eyelid Tumor Triage using Uncertainty-Aware and Explainable Artificial Intelligence

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

Entwicklung einer künstlichen Intelligenz zur ultraschallgestützten Detektion der Riesenzellarteriitis

Wed, 09 Sep 2026 00:00:00 +0000

Artificial intelligence for ultrasound-based detection of giant cell arteritis

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

INTERACT-CMIL: Multi-Task Shared Learning and Inter-Task Consistency for Conjunctival Melanocytic Intraepithelial Lesion Grading

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

OP0343 The ClaudicAItio Model: Ultrasound Image Classification in Giant Cell Arteritis by Application of Artificial Intelligence

Sun, 01 Jun 2025 00:00:00 +0000

Artery-Fusion Deep Learning for Enhanced Ultrasound Diagnosis of Giant Cell Arteritis

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

Conjunctival Melanocytic Intraepithelial Lesion (CMIL) Histopathology Dataset

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

Curated with Prof. Martina C. Herwig-Carl (Ophthalmic Pathology, UKB) together with University Hospitals Tübingen and Erlangen.

Exploring the limit of image resolution for human expert classification of vascular ultrasound images in giant cell arteritis and healthy subjects: the GCA-US-AI project

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

Giant Cell Arteritis Ultrasound Dataset

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

Curated with the Center for Rare Rheumatological Diseases (Claus-Jürgen Bauer, Prof. Valentin Schäfer), University Hospital Bonn. One of the largest clinically validated datasets for ultrasound-based GCA diagnosis; being extended towards a multi-center German cohort.

Our lab led the ML-readiness: quality control, cleaning, standardised preprocessing, annotation workflow and training/validation splits.

Investigation of the image resolution limit for human interpretation of ultrasound images in giant cell arteritis and healthy subjects-the GCA-US-AI project

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

KI-gestützte Diagnostik von Tumoren am Auge: Dignitätseinordnung von Lidtumoren mittels KI zur optimierten Patienteneinbestellung

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

Ein Team unter Leitung von Prof. Dr. Martina Christina Herwig-Carl (Klinik für Augenheilkunde, Universitätsklinikum Bonn) entwickelt eine App, die bei der Einschätzung und Klassifizierung von Augenlidtumoren unterstützt. Dies soll eine schnellere Diagnostik und eine bessere operative Versorgung ermöglichen.

Eyelid lesions are common, and telling benign from malignant tumours early is essential — yet many are first seen outside specialised centres, where subtle clinical and morphological features make the call difficult. The question there is less a definitive diagnosis than who needs to be referred, and how urgently.

What we do

Referral-oriented malignancy risk from routine photographs. A face-landmarker-guided pipeline standardises the periocular region, and a fine-tuned ResNet-50 estimates malignancy risk with patient-level evaluation on 956 patients from the University Eye Hospital Bonn archive (2020–2025), histopathology as ground truth — AUC 86% and sensitivity 84% ( ).
Knowing when to look again. Monte Carlo dropout gives each prediction an uncertainty that is highest for misclassified — especially missed malignant — cases, turning it into an alert to re-check those images; Grad-CAM shows that the model attends to the visible lesion.
From model to app. Funded as a prototyping grant by the Transfer Center enaCom of the University of Bonn, the work is being turned into an app that supports the assessment and triage of eyelid tumours in clinical practice.

Our role

Co-PI: methodological lead on uncertainty-aware and explainable AI, with Ashkan Shakarami (postdoctoral fellow, Department of Ophthalmology) developing the models.

Retinal Fundus Image Dataset

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

International collaboration with Dr. Said El Shamieh (Beirut Arab University, Lebanon). Harmonisation, label mapping and bias/generalisation evaluation protocols led by our lab.

Affordable Deep Learning for Diagnosing Inherited and Common Retinal Diseases via Color Fundus Photography

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

INTERACT-CMIL: Interpretable multi-task grading of conjunctival melanocytic lesions

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

Collaborators: University Hospital Tübingen, University Hospital Erlangen

This project addresses the grading of conjunctival melanocytic intraepithelial lesions (CMIL), rare precursors of conjunctival melanoma characterized by complex, interdependent histopathological criteria and substantial inter-observer variability. Together with clinical partners, we curated a multi-center, expert-annotated histopathology dataset comprising 486 conjunctival biopsy image patches labeled along five clinically defined grading axes. Methodologically, we developed INTERACT-CMIL, a multi-task deep learning framework modeling inter-criterion dependencies via selective multi-head supervision and an inter-task consistency loss.

The proposed approach outperformed CNN and pathology foundation-model baselines (macro-F1 up to 85%), producing coherent and interpretable predictions aligned with expert grading standards. Methodological development and evaluation were coordinated within my group in close collaboration with the clinical partners.

Perspective

The project established a reproducible computational benchmark for CMIL grading and contributed to the standardization of digital ocular pathology. A related follow-up project on AI-assisted eyelid tumor analysis received a University of Bonn Prototype Grant, supporting translational development and clinical feasibility studies.

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

Affordable deep learning for inherited and common retinal diseases

Wed, 01 Jun 2022 00:00:00 +0000

This international collaboration addresses the diagnosis of inherited retinal diseases (IRDs) and common retinal disorders using affordable color fundus photography, motivated by limited access to genetic testing and advanced imaging in low-resource settings. Together with partners in Germany and Lebanon, we curated and analyzed heterogeneous multi-center retinal imaging datasets combining in-house cohorts and public data, spanning diverse demographics and disease spectra. Methodologically, we developed lightweight deep learning models for multi-class retinal disease classification covering 21 retinal conditions, including 13 IRDs. The approach achieved strong diagnostic performance for prevalent IRDs (e.g., F1 score of 0.86 for retinitis pigmentosa) while remaining computationally efficient and interpretable.

Perspective

The collaboration established a sustainable international research link between ophthalmology groups in Germany and Lebanon and contributed to interdisciplinary student training. Ongoing work explores privacy-preserving and federated learning strategies to improve robustness and equitable deployment across populations further.

GCA-US-AI: Deep learning–based ultrasound diagnosis of Giant Cell Arteritis

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

In close interdisciplinary cooperation with clinical rheumatology, this project addresses the highly operator-dependent and time-critical ultrasound diagnosis of Giant Cell Arteritis (GCA), a vasculitis associated with irreversible vision loss. Together with our partners, we curated and annotated an in-house dataset of 3,800 ultrasound examinations from 344 patients, constituting one of the largest clinically validated datasets for ultrasound-based GCA diagnosis. Methodologically, we developed artery-aware deep learning models that integrate anatomical context into ultrasound image analysis, achieving robust performance improvements over artery-agnostic baselines (AUROC up to ~0.9 for larger arteries). Uncertainty estimation and visual explainability were incorporated to support reliable clinical interpretation.

Perspective

The collaboration is ongoing and continues to serve as a basis for joint publications, supervised theses, and continued dataset extension (in-house already ~15,000 images) toward a multi-center cohort. Current work explores vision–language models integrating imaging and clinical data and forms the basis of a planned joint DFG proposal coordinated together with the clinical partners.

Eine computergestützte automatische Polypencharakterisierung von Hyperplasten, Adenomen und Serratierten Adenomen im Kolorektum-Ergebnisse der CASSANDRA Studie

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

Fairness by Learning Orthogonal Disentangled Representations

Sat, 01 Aug 2020 00:00:00 +0000

A learning without forgetting approach to incorporate artifact knowledge in polyp localization tasks

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

Microaneurysms segmentation and diabetic retinopathy detection by learning discriminative representations

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

On the Fairness of Privacy-Preserving Representations in Medical Applications

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

Retinal Layer Segmentation Reformulated as OCT Language Processing

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

Understanding the effects of artifacts on automated polyp detection and incorporating that knowledge via learning without forgetting

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

Learning interpretable disentangled representations using adversarial vaes

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

Multi-scale Microaneurysms Segmentation Using Embedding Triplet Loss

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

Capsule networks against medical imaging data challenges

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

Disentangled, fair and discriminative representations for ophthalmic imaging

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

Data-driven models only reach clinical practice if what they learn is discriminative enough for subtle findings, compact and interpretable enough to be inspected, and free of information it should not rely on. Together with Carl Zeiss Meditec, the project addressed all three through representation learning, starting from retinal imaging for diabetic-retinopathy screening.

What we did

Finding the smallest lesions. Microaneurysms are the earliest visible sign of diabetic retinopathy and occupy only a few pixels of a fundus image. A two-stage approach segments them on multiple scales with a region-proposal fully convolutional network, then refines the candidates with a patch-wise network trained with an embedding triplet loss and selective sampling of hard patches — a 30% relative improvement over a plain fully convolutional network ( ; extended in ).
Interpretable latent spaces. An adversarial variational autoencoder with a total-correlation constraint enforces independent latent factors while keeping reconstruction fidelity, improving disentanglement by 81.5% and making the representation useful for clustering and for classification with few labels ( ).
Fairness by orthogonal disentanglement. Splitting the representation into a target code and a sensitive code, enforcing orthogonality between them and maximising the entropy of the target code with respect to the sensitive attribute yields representations that are invariant to it — state of the art on three of five benchmarks ( ) — and carries over to privacy-preserving medical applications such as autism detection and healthcare insurance ( ).

Our role

Scientific lead of the project as postdoc at TUM CAMP under Nassir Navab, and co-supervision — with Abouzar Eslami (Carl Zeiss Meditec) — of Hasan Sarhan’s industrial PhD, completed at TU Munich in 2024.

Perspective

The work on disentangled and orthogonal representations became a building block of our later research on fairness, privacy and federated learning in medical imaging.