publications

2026

1. arXiv

   

   Thumbnail adapted from the corresponding paper.

   Earth System Foundation Model (ESFM): A unified framework for heterogeneous data integration and forecasting

   Firat Ozdemir, Yun Cheng, Salman Mohebi, and 11 more authors

   arXiv preprint arXiv:2605.00850, 2026

   Abs arXiv Bib HTML Code

   

   An open foundation model extending Aurora’s 3D Swin UNet to integrate heterogeneous Earth system data — satellite observations, station data, and gridded climate data — under a single backbone. Adds axial attention, individual-variable tokenization, and adaptive layer-norm ensembles for probabilistic forecasting; demonstrated on Super Typhoon Doksuri and stratospheric warming cases.

   @article{ozdemir2026esfm,
     title = {Earth System Foundation Model ({ESFM}): A unified framework for heterogeneous data integration and forecasting},
     author = {Ozdemir, Firat and Cheng, Yun and Mohebi, Salman and Lehmann, Fanny and Adamov, Simon and Zhang, Zhenyi and Trentini, Leonardo and Grund, Dana and Fuhrer, Oliver and Hoefler, Torsten and Mishra, Siddhartha and Schemm, Sebastian and Soja, Benedikt and Salzmann, Mathieu},
     journal = {arXiv preprint arXiv:2605.00850},
     year = {2026},
     preview_alt = {CMIP6, ERA5, MODIS, and station data flowing through one Earth-system foundation model.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2. JGR: MLC

   

   Thumbnail adapted from the corresponding paper.

   Fine-Tuning a Weather Foundation Model With Lightweight Decoders for Unseen Physical Processes

   Fanny Lehmann, Firat Ozdemir, Benedikt Soja, and 3 more authors

   Journal of Geophysical Research: Machine Learning and Computation, 2026

   e2025JH000837 2025JH000837

   Abs DOI

   

   Recent advances in AI weather forecasting have led to the emergence of so-called “foundation models”, typically defined by expensive pretraining and minimal fine-tuning for downstream tasks. However, in the natural sciences, a desirable foundation model should also encode meaningful statistical relationships between the underlying physical variables. This study evaluates the performance of the state-of-the-art Aurora foundation model in predicting hydrological variables, which were not included during pretraining. We introduce a lightweight approach using shallow decoders trained on the latent representations of the pretrained model to predict these new variables. As a baseline, we compare this to fine-tuning the full model, which allows further optimization of the latent space while incorporating new variables into both inputs and outputs. The decoder-based approach requires 43% less training time and 53% less memory, while achieving strong accuracy across various hydrological variables and preserving desirable properties of the foundation model, such as autoregressive stability. Notably, decoder accuracy depends on the physical correlation between the new variables and those used during pretraining, indicating that Aurora’s latent space captures meaningful physical relationships. The decoders outperform UNet, climatology, and persistence baselines on variables related to pretraining, showing the benefits of the latent space. In this sense, we argue that an important quality metric for foundation models in Earth sciences is their ability to be extended to new variables without a full fine-tuning. This provides a new perspective for making foundation models more accessible to communities with limited computational resources, while supporting broader adoption in Earth sciences.

3. CPAL

   

   Thumbnail adapted from the corresponding paper.

   FocusDC: Real-World Scene Infusion for Robust Dataset Condensation

   Youbing Hu, Yun Cheng, Olga Saukh, and 4 more authors

   In The Third Conference on Parsimony and Learning, 2026

   Abs arXiv Bib HTML

   

   Dataset condensation that synthesizes training data using attention focused on target object locations, improving generalization across ViTs and extending dataset distillation to dense tasks (object detection). First such application to COCO with YOLOv11.

   @inproceedings{hu2026focusdc,
     title = {{FocusDC}: Real-World Scene Infusion for Robust Dataset Condensation},
     author = {Hu, Youbing and Cheng, Yun and Saukh, Olga and Ozdemir, Firat and Lu, Anqi and Cao, Zhiqiang and Li, Zhijun},
     booktitle = {The Third Conference on Parsimony and Learning},
     year = {2026},
     preview_alt = {Key image patches and low-resolution images reconstructed with a soft class label.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

4. arXiv

   

   Thumbnail adapted from the corresponding paper.

   Can AI Weather Models Predict Beyond Two Weeks? A Quantitative Benchmark and Analysis of Long Rollouts

   Fanny Lehmann, Firat Ozdemir, Yun Cheng, and 4 more authors

   2026

   Abs arXiv Bib HTML

   

   While AI weather models excel at short-to-medium range forecasts (up to 15 days), they frequently suffer from ill-defined “instabilities” when rolled out over longer horizons. This work addresses the lack of a formal taxonomy by categorizing these failures into three distinct regimes — blow-up, drift, and loss of seasonality — through year-long rollouts of nine state-of-the-art AI weather models. Our analysis reveals that stability hinges on the treatment of small spatio-temporal scales: unstable models amplify high-frequency energy, while stable models act as denoisers when noise is added to their inputs. Far from reducing these models to mere stochastic parrots, our findings highlight that stable models generate unique weather trajectories, conditioned on the initial state. We verify our findings through ablation studies on architectural design choices, conducted using state-of-the-art Vision Transformer (ViT) AI weather model architectures.

   @misc{lehmann2026aiweathermodelspredict,
     title = {Can AI Weather Models Predict Beyond Two Weeks? A Quantitative Benchmark and Analysis of Long Rollouts},
     author = {Lehmann, Fanny and Ozdemir, Firat and Cheng, Yun and Hoefler, Torsten and Schemm, Sebastian and Soja, Benedikt and Mishra, Siddhartha},
     year = {2026},
     archiveprefix = {arXiv},
     primaryclass = {cs.LG},
     url = {https://arxiv.org/abs/2605.30184},
     preview_alt = {Long-range AI weather-model forecasts illustrating drift and loss of seasonality.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

5. arXiv

   

   Thumbnail adapted from the corresponding paper.

   A Latent ODE Approach to Spatiotemporal Modeling of Cine Cardiac MRI

   David Brüggemann, Ekaterina Krymova, Firat Ozdemir, and 6 more authors

   2026

   Bib

   

   @misc{bruggemann2026latent,
     title = {A Latent ODE Approach to Spatiotemporal Modeling of Cine Cardiac MRI},
     author = {Brüggemann, David and Krymova, Ekaterina and Ozdemir, Firat and von Spiczak, Jochen and Kozerke, Sebastian and Mora, Samia and Manka, Robert and Salzmann, Mathieu and Demler, Olga V.},
     year = {2026},
     archiveprefix = {arXiv},
     primaryclass = {cs.AI},
     url = {https://arxiv.org/abs/2606.26718},
     preview_alt = {Cardiac mesh sequences and patient covariates encoded into a latent neural ODE for reconstruction and personalized heart-failure risk scoring.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

6. EGU

   

   Extending foundation models from weather to climate: challenges and promises

   Fanny Lehmann, Riccardo Neumarker, Gabriele Scorrano, and 10 more authors

   In EGU General Assembly, 2026

   Abs DOI Bib

   

   AI weather models and weather-based foundation models have demonstrated impressive skills in short- to medium-range forecasts. While most weather models become unstable on longer time scales, a wide variety of AI climate emulators have been proposed, raising questions about the fundamental differences between these approaches. In this work, we compare state-of-the-art models when producing rollouts on annual time scales. We quantify and characterize different types of instability: smoothing, visual artifacts, drift, and loss of seasonality. This analysis highlights the previously unreported stability of the Aurora foundation model and the Earth System Foundation Model (ESFM) for rollouts longer than 35 years. To encompass more diverse representations of possible states of the Earth, ESFM is pretrained on a variety of CMIP6 datasets from the historical period, in addition to the ERA5 reanalysis commonly used in AI models. ESFM also includes climate forcings for physically driven long rollouts. We demonstrate the benefits of CMIP6 pretraining when finetuning on new CMIP6 datasets, including datasets with higher resolution, unseen physical processes, and climate change scenarios. Overall, this work opens perspectives to adapt large-scale pretrained foundation models to the specific challenges of climate projections.

   @inproceedings{Lehmann2026ESFMClimate,
     author = {Lehmann, Fanny and Neumarker, Riccardo and Scorrano, Gabriele and Cheng, Yun and Mohebi, Salman and Ozdemir, Firat and Gou, Junyang and Fuhrer, Oliver and Hoefler, Torsten and Mishra, Siddhartha and Salzmann, Mathieu and Schemm, Sebastian and Soja, Benedikt},
     title = {Extending foundation models from weather to climate: challenges and promises},
     booktitle = {EGU General Assembly},
     year = {2026},
     address = {Vienna, Austria},
     publisher = {Copernicus Meetings},
     doi = {10.5194/egusphere-egu26-19872},
     url = {https://doi.org/10.5194/egusphere-egu26-19872},
     preview_alt = {A globe, atmospheric fields, and a shared neural representation connecting weather and climate.}
   }

2025

1. JAMES

   

   Thumbnail adapted from the corresponding paper.

   Revisiting machine learning approaches for short- and longwave radiation inference in weather and climate models

   Guillaume Bertoli, Salman Mohebi, Firat Ozdemir, and 5 more authors

   Journal of Advances in Modeling Earth Systems, 2025

   Abs DOI Bib HTML

   

   Compares five ML radiation emulators coupled to the ICON weather model on GPUs. A BiLSTM with physics-informed normalization is most accurate and enables stable aquaplanet simulations at  80 km resolution, with performance comparable to GPU-optimized classical physics parameterizations.

   @article{bertoli2025revisiting,
     title = {Revisiting machine learning approaches for short- and longwave radiation inference in weather and climate models},
     author = {Bertoli, Guillaume and Mohebi, Salman and Ozdemir, Firat and Jucker, Jonas and R{\"u}dis{\"u}hli, Stefan and Perez-Cruz, Fernando and Salzmann, Mathieu and Schemm, Sebastian},
     journal = {Journal of Advances in Modeling Earth Systems},
     volume = {17},
     number = {9},
     pages = {e2025MS004956},
     year = {2025},
     publisher = {American Geophysical Union},
     doi = {10.1029/2025MS004956},
     preview_alt = {BiLSTM and ecRad net-heating-rate cross-sections at two and four simulation days.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2. IEEE RA-L

   

   Thumbnail adapted from the corresponding paper.

   PokeFlex: A Real-World Dataset of Volumetric Deformable Objects for Robotics

   Jan Obrist, Miguel Zamora, Hehui Zheng, and 5 more authors

   IEEE Robotics and Automation Letters, 2025

   Abs DOI arXiv Bib HTML Code

   

   A real-world multimodal robotics dataset of 18 deformable objects with varying stiffness, captured via 360-degree volumetric capture while a robot pokes each object. Modalities include 3D textured meshes, point clouds, RGB and depth at 30 fps; targets online 3D mesh reconstruction and mesh-simulation-based control.

   @article{obrist2025pokeflex,
     title = {{PokeFlex}: A Real-World Dataset of Volumetric Deformable Objects for Robotics},
     author = {Obrist, Jan and Zamora, Miguel and Zheng, Hehui and Hinchet, Ronan and Ozdemir, Firat and Zarate, Juan and Katzschmann, Robert K. and Coros, Stelian},
     journal = {IEEE Robotics and Automation Letters},
     year = {2025},
     publisher = {IEEE},
     doi = {10.1109/LRA.2025.3600148},
     preview_alt = {Animated deformation of a red octopus-shaped object.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2024

1. Med. Image Anal.

   

   Thumbnail adapted from the corresponding paper.

   Signal domain adaptation network for limited-view optoacoustic tomography

   Anna Klimovskaia Susmelj, Berkan Lafci, Firat Ozdemir, and 4 more authors

   Medical Image Analysis, 2024

   Abs DOI Bib HTML Code

   

   Extends prior signal-domain MIDL work into a full domain-adaptation framework for limited-view optoacoustic tomography, combining training on both time-resolved signals and tomographic reconstructions to enhance image quality.

   @article{susmelj2024signal,
     title = {Signal domain adaptation network for limited-view optoacoustic tomography},
     author = {Susmelj, Anna Klimovskaia and Lafci, Berkan and Ozdemir, Firat and Davoudi, Neda and De{\'a}n-Ben, Xos{\'e} Lu{\'i}s and Perez-Cruz, Fernando and Razansky, Daniel},
     journal = {Medical Image Analysis},
     volume = {91},
     pages = {103012},
     year = {2024},
     publisher = {Elsevier},
     doi = {10.1016/j.media.2023.103012},
     preview_alt = {Grid comparing optoacoustic reconstructions from baseline and SDAN methods.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2. npj Digit. Med.

   

   Thumbnail adapted from the corresponding paper.

   Modeling multiple sclerosis using mobile and wearable sensor data

   Shkurta Gashi, Pietro Oldrati, Max Moebus, and 8 more authors

   npj Digital Medicine, 2024

   Abs DOI Bib HTML

   

   Using a free-living dataset of 55 people with MS and 24 healthy controls over 489 days, derives features from mobile and wearable devices (notably physical activity and heart rate) to distinguish MS from controls and estimate disability and fatigue levels, demonstrating feasibility of continuous remote MS monitoring.

   @article{gashi2024modeling,
     title = {Modeling multiple sclerosis using mobile and wearable sensor data},
     author = {Gashi, Shkurta and Oldrati, Pietro and Moebus, Max and Hilty, Marc and Barrios, Liliana and Ozdemir, Firat and {PHRT Consortium} and Kana, Veronika and Lutterotti, Andreas and R{\"a}tsch, Gunnar and Holz, Christian},
     journal = {npj Digital Medicine},
     volume = {7},
     number = {1},
     pages = {64},
     year = {2024},
     doi = {10.1038/s41746-024-01025-8},
     publisher = {Nature Publishing Group},
     preview_alt = {Study design linking clinical visits, free-living sensors, feature extraction, and multiple-sclerosis modeling.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

3. Turbo Expo

   

   Data-driven inverse design method for turbomachinery

   Kwok Kai So, Luis Salamanca, Firat Ozdemir, and 1 more author

   In Turbo Expo: Power for Land, Sea, and Air, 2024

   Abs DOI Bib HTML

   

   Neural-network inverse design approach for turbomachinery: once trained with design parameters as inputs and performance metrics as outputs, the model is inverted to map target performance to design parameters. Generalizes beyond training ranges and is extensible to multi-component systems.

   @inproceedings{so2024data,
     title = {Data-driven inverse design method for turbomachinery},
     author = {So, Kwok Kai and Salamanca, Luis and Ozdemir, Firat and Perez-Cruz, Fernando},
     booktitle = {Turbo Expo: Power for Land, Sea, and Air},
     volume = {88087},
     pages = {V12DT34A004},
     year = {2024},
     organization = {American Society of Mechanical Engineers},
     doi = {10.1115/GT2024-121752},
     preview_alt = {Turbomachinery variants flowing through a neural network toward an optimized component.}
   }

2023

1. EGU

   

   Thumbnail adapted from the corresponding paper.

   Building a physics-constrained, fast and stable machine learning-based radiation emulator

   Guillaume Bertoli, Sebastian Schemm, Firat Ozdemir, and 2 more authors

   2023

   Abs DOI Bib HTML

   

   Combines machine learning with the ecRad radiative transfer solver using Gaussian processes and random forests, aiming for a physically consistent radiation emulator for climate modeling.

   @techreport{bertoli2022building,
     title = {Building a physics-constrained, fast and stable machine learning-based radiation emulator},
     author = {Bertoli, Guillaume and Schemm, Sebastian and Ozdemir, Firat and Szekely, Eniko and Cruz, Fernando Perez},
     year = {2023},
     institution = {Copernicus Meetings},
     doi = {10.5194/egusphere-egu23-3418},
     preview_alt = {Convolutional radiation-emulator architecture with neighboring grid cells and an energy-conservation loss.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2. AISTATS

   

   Thumbnail adapted from the corresponding paper.

   Retrospective Uncertainties for Deep Models using Vine Copulas

   Nataša Tagasovska, Firat Ozdemir, and Axel Brando

   In International Conference on Artificial Intelligence and Statistics, 2023

   Abs arXiv Bib HTML Code

   

   Vine-Copula Neural Networks (VCNN): given any pretrained network, fits a vine copula model to retrospectively produce calibrated uncertainty estimates without modifying the loss or architecture. Task- and architecture-agnostic across regression, classification, recurrent, and fully connected models.

   @inproceedings{tagasovska2023retrospective,
     title = {Retrospective Uncertainties for Deep Models using Vine Copulas},
     author = {Tagasovska, Nata{\v{s}}a and Ozdemir, Firat and Brando, Axel},
     booktitle = {International Conference on Artificial Intelligence and Statistics},
     year = {2023},
     preview_alt = {Vine-copula neural network adding calibrated uncertainty intervals to predictions.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

3. TMLR

   

   Thumbnail adapted from the corresponding paper.

   OADAT: Experimental and Synthetic Clinical Optoacoustic Data for Standardized Image Processing

   Firat Ozdemir, Berkan Lafci, Xose Luis Dean-Ben, and 2 more authors

   Transactions on Machine Learning Research, 2023

   Abs arXiv Bib HTML Code

   

   Provides standardized experimental clinical optoacoustic data of human forearms and forward-model simulations, plus 44 benchmark experiments covering sparse acquisition, limited view, and segmentation tasks.

   @article{ozdemir2023oadat,
     title = {OADAT: Experimental and Synthetic Clinical Optoacoustic Data for Standardized Image Processing},
     author = {Ozdemir, Firat and Lafci, Berkan and Dean-Ben, Xose Luis and Razansky, Daniel and Perez-Cruz, Fernando},
     journal = {Transactions on Machine Learning Research},
     number = {2835-8856},
     year = {2023},
     preview_alt = {Forearm acquisition setup and semi-circle and multisegment optoacoustic transducer arrays.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

4. NeurIPS

   

   Thumbnail adapted from the corresponding paper.

   Renku: a platform for sustainable data science

   Rok Roškar, Chandrasekhar Ramakrishnan, Michele Volpi, and 23 more authors

   In Advances in Neural Information Processing Systems, 2023

   Abs DOI Bib HTML Code

   

   Renku is a platform for reproducible and collaborative data analysis, recording the full chain from data import through outputs as a knowledge graph. Supports cloud and on-prem data connectors, GitHub/GitLab integration, and preconfigured compute environments.

   @inproceedings{roskar2023renku,
     title = {Renku: a platform for sustainable data science},
     author = {Ro{\v{s}}kar, Rok and Ramakrishnan, Chandrasekhar and Volpi, Michele and Perez-Cruz, Fernando and Gasser, Lilian and Ozdemir, Firat and Paitz, Patrick and Alisafaee, Mohammad and Fischer, Philipp and Grubenmann, Ralf and Harris, Eliza and Olevski, Tasko and Remlinger, Carl and Salamanca, Luis and Capon Garcia, Elisabet and Cavazzi, Lorenzo and Chrobasik, Jakub and Cordoba Osnas, Darlin and Degano, Alessandro and Dupre, Jimena and Johnson, Wesley and Kettner, Eike and Kinkead, Laura and Murphy, Sean D. and Thiebaut, Flora and Verscheure, Olivier},
     booktitle = {Advances in Neural Information Processing Systems},
     volume = {36},
     pages = {42161--42173},
     year = {2023},
     doi = {10.52202/075280-1827},
     preview_alt = {Puzzle-piece diagram of versioned code, reusable data, workflows, environments, and a knowledge graph.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2022

1. SciPost

   

   Thumbnail adapted from the corresponding paper.

   Learning summary statistics for bayesian inference with autoencoders

   Carlo Albert, Simone Ulzega, Firat Ozdemir, and 2 more authors

   SciPost Physics Core, 2022

   Abs DOI arXiv Bib HTML

   

   For approximate Bayesian computation with intractable likelihoods, proposes using latent representations from autoencoders as learned summary statistics. The decoder receives explicit/implicit noise so the encoder isolates parameter-relevant signal; validated on two stochastic model classes.

   @article{albert2022learning,
     title = {Learning summary statistics for bayesian inference with autoencoders},
     author = {Albert, Carlo and Ulzega, Simone and Ozdemir, Firat and Perez-Cruz, Fernando and Mira, Antonietta},
     journal = {SciPost Physics Core},
     volume = {5},
     number = {3},
     pages = {043},
     year = {2022},
     doi = {10.21468/SciPostPhysCore.5.3.043},
     preview_alt = {ENCA and INCA architectures above latent summary-statistic distributions.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2. GMD

   

   Thumbnail adapted from the corresponding paper.

   A Bayesian data assimilation framework for lake 3D hydrodynamic models with a physics-preserving particle filtering method using SPUX-MITgcm v1

   Artur Safin, Damien Bouffard, Firat Ozdemir, and 5 more authors

   Geoscientific Model Development, 2022

   Abs DOI Bib HTML

   

   Couples SPUX Bayesian inference with the MITgcm hydrodynamic solver for Lake Geneva via smoothed particle MCMC; includes a BiLSTM mapping bulk to skin lake temperature using atmospheric forcing.

   @article{safin2022bayesian,
     title = {A Bayesian data assimilation framework for lake 3D hydrodynamic models with a physics-preserving particle filtering method using SPUX-MITgcm v1},
     author = {Safin, Artur and Bouffard, Damien and Ozdemir, Firat and Ram{\'o}n, Cintia L and Runnalls, James and Georgatos, Fotis and Minaudo, Camille and {\v{S}}ukys, Jonas},
     journal = {Geoscientific Model Development},
     volume = {15},
     number = {20},
     pages = {7715--7730},
     year = {2022},
     publisher = {Copernicus GmbH},
     doi = {10.5194/gmd-15-7715-2022},
     preview_alt = {Observed, data-assimilated, and difference profiles of lake temperature across depth.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

3. MIDL

   

   Thumbnail adapted from the corresponding paper.

   Signal Domain Learning Approach for Optoacoustic Image Reconstruction from Limited View Data

   Anna Klimovskaia Susmelj, Berkan Lafci, Firat Ozdemir, and 4 more authors

   In International Conference on Medical Imaging with Deep Learning, 2022

   Abs Bib HTML Code

   

   Multi-spectral optoacoustic tomography reconstruction from limited tomographic view. Proposes a Style Network to bridge the simulated-vs-experimental domain gap and a Side Network trained on simulated data to remove limited-view artifacts.

   @inproceedings{susmelj2022signal,
     title = {Signal Domain Learning Approach for Optoacoustic Image Reconstruction from Limited View Data},
     author = {Susmelj, Anna Klimovskaia and Lafci, Berkan and Ozdemir, Firat and Davoudi, Neda and Dean-Ben, Xose Luis and Perez-Cruz, Fernando and Razansky, Daniel},
     booktitle = {International Conference on Medical Imaging with Deep Learning},
     pages = {1173--1191},
     year = {2022},
     organization = {PMLR},
     preview_alt = {Signal- and image-domain comparisons for limited-view optoacoustic reconstruction.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2021

1. arXiv

   

   Thumbnail adapted from the corresponding paper.

   Probabilistic modeling of lake surface water temperature using a Bayesian spatio-temporal graph convolutional neural network

   Michael Stalder, Firat Ozdemir, Artur Safin, and 3 more authors

   arXiv preprint arXiv:2109.13235, 2021

   Abs arXiv Bib HTML

   

   A spatio-temporal model combining Bayesian recurrent layers with graph convolutional networks to estimate lake surface temperature from depth-resolved simulations and meteorological data, producing probabilistic predictions over sparse spatial inputs.

   @article{stalder2021probabilistic,
     title = {Probabilistic modeling of lake surface water temperature using a Bayesian spatio-temporal graph convolutional neural network},
     author = {Stalder, Michael and Ozdemir, Firat and Safin, Artur and Sukys, Jonas and Bouffard, Damien and Perez-Cruz, Fernando},
     journal = {arXiv preprint arXiv:2109.13235},
     year = {2021},
     preview_alt = {Lake maps comparing prediction error and fit for two probabilistic models.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2. KBS

   

   Thumbnail adapted from the corresponding paper.

   Active learning for segmentation based on Bayesian sample queries

   Firat Ozdemir, Zixuan Peng, Philipp Fuernstahl, and 2 more authors

   Knowledge-Based Systems, 2021

   Abs DOI arXiv Bib HTML Code

   

   Proposes an active learning method combining representativeness (via information-maximizing autoencoders) with uncertainty to iteratively select optimal samples for annotation in medical image segmentation. With 10% of a shoulder MRI dataset annotated, reaches within 5% Dice of the fully-annotated upper bound; the gap drops to 2% at under 20% annotated.

   @article{ozdemir2021active,
     title = {Active learning for segmentation based on Bayesian sample queries},
     author = {Ozdemir, Firat and Peng, Zixuan and Fuernstahl, Philipp and Tanner, Christine and Goksel, Orcun},
     journal = {Knowledge-Based Systems},
     volume = {214},
     pages = {106531},
     year = {2021},
     publisher = {Elsevier},
     doi = {10.1016/j.knosys.2020.106531},
     preview_alt = {Dice-score curves versus annotated-data percentage for five sampling strategies.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2020

1. arXiv

   

   Thumbnail adapted from the corresponding paper.

   Delineating Bone Surfaces in B-Mode Images Constrained by Physics of Ultrasound Propagation

   Firat Ozdemir, Christine Tanner, and Orcun Goksel

   arXiv preprint arXiv:2001.02001, 2020

   Abs arXiv Bib HTML

   

   Detects bone surfaces in B-mode ultrasound via a factor-graph model with unary potentials encoding local soft-tissue vs acoustic-shadow likelihood and pairwise constraints encoding ultrasound propagation physics. Achieves RMSE 0.28 mm, symmetric Hausdorff 1.78 mm, 99.9% detection rate.

   @article{ozdemir2020delineating,
     title = {Delineating Bone Surfaces in B-Mode Images Constrained by Physics of Ultrasound Propagation},
     author = {Ozdemir, Firat and Tanner, Christine and Goksel, Orcun},
     journal = {arXiv preprint arXiv:2001.02001},
     year = {2020},
     preview_alt = {Best and median ultrasound cases comparing constrained bone graphs with phase symmetry.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2. PhD Thesis

   

   Medical Image Segmentation with Efficient Use of Annotations and Resources

   Firat Özdemir

   ETH Zurich, 2020

   Abs Bib HTML

   

   Addresses medical image segmentation with efficient use of annotations and resources. Proposes interactive segmentation via random walker on multi-scale 3D graphs, active learning metrics for optimal sample selection, class-incremental learning for extending to new anatomical structures with minimal annotations, and automatic segmentation leveraging imaging physics with graphical models.

   @phdthesis{ozdemir2020medical,
     title = {Medical Image Segmentation with Efficient Use of Annotations and Resources},
     author = {{\"O}zdemir, Firat},
     year = {2020},
     school = {ETH Zurich},
     preview_alt = {Medical scans, anatomical contours, and a neural network representing efficient image segmentation.}
   }

3. EGU

   

   Data assimilation in lake Geneva using the SPUX framework

   Artur Safin, Damien Bouffard, James Runnalls, and 5 more authors

   In EGU General Assembly Conference Abstracts, 2020

   Abs DOI Bib HTML

   

   Bayesian data assimilation framework for Lake Geneva using the SPUX package with ensemble weather forecasts, parallelized on the CSCS cluster. Combines in-situ and remote sensing measurements (including satellite lake surface temperature) with a 3D hydrodynamic model via particle filtering, incorporating measurement uncertainties and multi-level model calibration.

   @inproceedings{safin2020data,
     title = {Data assimilation in lake Geneva using the SPUX framework},
     author = {Safin, Artur and Bouffard, Damien and Runnalls, James and Georgatos, Fotis and Bouillet, Eric and Ozdemir, Firat and Perez Cruz, Fernando and {\v{S}}ukys, Jonas},
     booktitle = {EGU General Assembly Conference Abstracts},
     pages = {19564},
     year = {2020},
     doi = {10.5194/egusphere-egu2020-19564},
     preview_alt = {Lake Geneva, observation sensors, a satellite, and a data-assimilation flow.}
   }

2019

1. IJCARS

   

   Thumbnail adapted from the corresponding paper.

   Extending pretrained segmentation networks with additional anatomical structures

   Firat Ozdemir and Orcun Goksel

   International journal of computer assisted radiology and surgery, 2019

   Abs DOI arXiv Bib HTML

   

   Class-incremental learning for medical image segmentation: extends networks to new anatomical structures without catastrophic forgetting of previously learned ones.

   @article{ozdemir2019extending,
     title = {Extending pretrained segmentation networks with additional anatomical structures},
     author = {Ozdemir, Firat and Goksel, Orcun},
     journal = {International journal of computer assisted radiology and surgery},
     volume = {14},
     pages = {1187--1195},
     year = {2019},
     publisher = {Springer International Publishing},
     doi = {10.1007/s11548-019-01984-4},
     preview_alt = {Knee MRI showing added tibia and retained femur segmentations.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2. SPIE

   

   Thumbnail adapted from the corresponding paper.

   Deep learning techniques for bone surface delineation in ultrasound

   Matija Ciganovic, Firat Özdemir, Mazda Farshad, and 1 more author

   In Medical Imaging 2019: Ultrasonic Imaging and Tomography, 2019

   Abs DOI Bib HTML

   

   Examines deep learning for automatic bone surface extraction from intraoperative ultrasound to support computer-assisted orthopedic interventions. Evaluates two popular segmentation networks on 1,800 manually annotated ultrasound frames, investigating preprocessing effects. Demonstrates excellent qualitative and quantitative segmentation performance.

   @inproceedings{ciganovic2019deep,
     title = {Deep learning techniques for bone surface delineation in ultrasound},
     author = {Ciganovic, Matija and {\"O}zdemir, Firat and Farshad, Mazda and G{\"o}ksel, Orcun},
     booktitle = {Medical Imaging 2019: Ultrasonic Imaging and Tomography},
     volume = {10955},
     pages = {224--230},
     year = {2019},
     organization = {SPIE},
     doi = {10.1117/12.2512997},
     preview_alt = {Ultrasound frame with predicted and manually annotated bone-surface contours.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2018

1. IJCARS

   

   Thumbnail adapted from the corresponding paper.

   Registration of 3D freehand ultrasound to a bone model for orthopedic procedures of the forearm

   Matija Ciganovic, Firat Ozdemir, Fabien Pean, and 3 more authors

   International journal of computer assisted radiology and surgery, 2018

   Abs DOI Bib HTML

   

   3D freehand ultrasound to preoperative bone model registration for orthopedic surgical guidance. Extracts partial bone surfaces via phase symmetry and factor graphs, then aligns to 3D bone models using a multi-initialization surface-based scheme robust to partial views. Achieves clinically feasible registration accuracy on 36 forearm ultrasound volumes.

   @article{ciganovic2018registration,
     title = {Registration of 3D freehand ultrasound to a bone model for orthopedic procedures of the forearm},
     author = {Ciganovic, Matija and Ozdemir, Firat and Pean, Fabien and Fuernstahl, Philipp and Tanner, Christine and Goksel, Orcun},
     journal = {International journal of computer assisted radiology and surgery},
     volume = {13},
     pages = {827--836},
     year = {2018},
     publisher = {Springer International Publishing},
     doi = {10.1007/s11548-018-1756-0},
     preview_alt = {Ultrasound, MRI, and overlay views showing cross-modal forearm bone registration.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2. MICCAI

   

   Thumbnail adapted from the corresponding paper.

   Learn the new, keep the old: Extending pretrained models with new anatomy and images

   Firat Ozdemir, Philipp Fuernstahl, and Orcun Goksel

   In Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part IV 11, 2018

   Abs DOI arXiv Bib HTML

   

   Incremental learning for medical image segmentation: extends pretrained networks to new anatomy by retaining a small representative set of prior-task images and annotations while learning new structures, avoiding catastrophic forgetting.

   @inproceedings{ozdemir2018learn,
     title = {Learn the new, keep the old: Extending pretrained models with new anatomy and images},
     author = {Ozdemir, Firat and Fuernstahl, Philipp and Goksel, Orcun},
     booktitle = {Medical Image Computing and Computer Assisted Intervention--MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part IV 11},
     pages = {361--369},
     year = {2018},
     organization = {Springer International Publishing},
     doi = {10.1007/978-3-030-00937-3_42},
     preview_alt = {Shoulder MRI views showing retained and newly learned anatomical segmentations.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

3. arXiv

   

   Thumbnail adapted from the corresponding paper.

   Generative adversarial networks for MR-CT deformable image registration

   Christine Tanner, Firat Ozdemir, Romy Profanter, and 3 more authors

   arXiv preprint arXiv:1807.07349, 2018

   Abs arXiv Bib HTML

   

   Evaluates cycle-GANs for MR-CT deformable image registration. Cycle-GAN inadvertently synthesizes geometric changes alongside intensity mappings; reducing discriminator receptive field mitigates this. NMI+MIND achieves 76.9% segmentation overlap on abdominal registration.

   @article{tanner2018generative,
     title = {Generative adversarial networks for MR-CT deformable image registration},
     author = {Tanner, Christine and Ozdemir, Firat and Profanter, Romy and Vishnevsky, Valeriy and Konukoglu, Ender and Goksel, Orcun},
     journal = {arXiv preprint arXiv:1807.07349},
     year = {2018},
     preview_alt = {Thoracic CT and MRI registration examples with anatomical contour overlays.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

4. DLMIA

   

   Thumbnail adapted from the corresponding paper.

   Active learning for segmentation by optimizing content information for maximal entropy

   Firat Ozdemir, Zixuan Peng, Christine Tanner, and 2 more authors

   In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, 2018

   DOI arXiv Bib HTML

   

   @inproceedings{ozdemir2018active,
     title = {Active learning for segmentation by optimizing content information for maximal entropy},
     author = {Ozdemir, Firat and Peng, Zixuan and Tanner, Christine and Fuernstahl, Philipp and Goksel, Orcun},
     booktitle = {Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4},
     pages = {183--191},
     year = {2018},
     organization = {Springer International Publishing},
     doi = {10.1007/978-3-030-00889-5_21},
     preview_alt = {Ground-truth, baseline, and proposed active-learning segmentation examples.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2017

1. MICCAI WS

   

   Thumbnail adapted from the corresponding paper.

   Automatic segmentation of abdominal MRI using selective sampling and random walker

   Janine Thoma, Firat Ozdemir, and Orcun Goksel

   In Medical Computer Vision and Bayesian and Graphical Models for Biomedical Imaging: MICCAI 2016 International Workshops, MCV and BAMBI, Athens, Greece, October 21, 2016, Revised Selected Papers 8, 2017

   Abs DOI Bib HTML

   

   MRI segmentation combining voxel-wise random forest classification with Random Walker spatial regularization. Proposes selective sampling of training patches near anatomical boundaries to improve joint segmentation performance. Shows Dice improvements of up to 12% on 20 full-body MR T1 scans from the VISCERAL dataset.

   @inproceedings{thoma2017automatic,
     title = {Automatic segmentation of abdominal MRI using selective sampling and random walker},
     author = {Thoma, Janine and Ozdemir, Firat and Goksel, Orcun},
     booktitle = {Medical Computer Vision and Bayesian and Graphical Models for Biomedical Imaging: MICCAI 2016 International Workshops, MCV and BAMBI, Athens, Greece, October 21, 2016, Revised Selected Papers 8},
     pages = {83--93},
     year = {2017},
     organization = {Springer International Publishing},
     doi = {10.1007/978-3-319-61188-4_8},
     preview_alt = {Two abdominal MRI slices comparing automatic multi-organ segmentation with ground truth.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2. IJCARS

   

   Thumbnail adapted from the corresponding paper.

   Interactive segmentation in MRI for orthopedic surgery planning: bone tissue

   Firat Ozdemir, Neerav Karani, Philipp Fürnstahl, and 1 more author

   International journal of computer assisted radiology and surgery, 2017

   Abs DOI Bib HTML

   

   Interactive MRI bone segmentation using random forest classification and random walker regularization, enabling surgeon corrections and enforcing anatomically connected structures. Interactive humerus segmentation takes  150 s with 3.5× time gain over manual annotation; a multi-resolution random walker further reduces run time 20× for high-resolution data.

   @article{ozdemir2017interactive,
     title = {Interactive segmentation in MRI for orthopedic surgery planning: bone tissue},
     author = {Ozdemir, Firat and Karani, Neerav and F{\"u}rnstahl, Philipp and Goksel, Orcun},
     journal = {International journal of computer assisted radiology and surgery},
     volume = {12},
     pages = {1031--1039},
     year = {2017},
     publisher = {Springer International Publishing},
     doi = {10.1007/s11548-017-1570-0},
     preview_alt = {Four humerus MRI overlays comparing interactive bone-segmentation protocols.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2016

1. MICCAI

   

   Thumbnail adapted from the corresponding paper.

   Graphical modeling of ultrasound propagation in tissue for automatic bone segmentation

   Firat Ozdemir, Ece Ozkan, and Orcun Goksel

   In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19, 2016

   Abs DOI Bib HTML

   

   Physics-based factor graph formulation for automatic bone surface segmentation in ultrasound, encoding ultrasound propagation physics without manual interaction or complex parameterization. Evaluated on in-vivo multi-anatomical-location ultrasound images. Achieves RMSE 0.59 mm, significantly outperforming state-of-the-art.

   @inproceedings{ozdemir2016graphical,
     title = {Graphical modeling of ultrasound propagation in tissue for automatic bone segmentation},
     author = {Ozdemir, Firat and Ozkan, Ece and Goksel, Orcun},
     booktitle = {Medical Image Computing and Computer-Assisted Intervention--MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19},
     pages = {256--264},
     year = {2016},
     organization = {Springer International Publishing},
     doi = {10.1007/978-3-319-46723-8_30},
     preview_alt = {Ultrasound bone-surface examples comparing reference, phase symmetry, and factor-graph delineations.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }

2013

1. ICIAR

   

   Thumbnail adapted from the corresponding paper.

   A decision forest based feature selection framework for action recognition from rgb-depth cameras

   Farhood Negin, Fırat Özdemir, Ceyhun Burak Akgül, and 2 more authors

   In Image Analysis and Recognition: 10th International Conference, ICIAR 2013, Póvoa do Varzim, Portugal, June 26-28, 2013. Proceedings 10, 2013

   Abs DOI Bib HTML

   

   Random decision forest framework for action recognition using kinematic features from skeletal body representations. Discriminatively selects a spatiotemporally localized feature subset (1/10th of original), then trains an SVM classifier. Achieves 94% on MSRC-12 (12 classes) and 98% on WorkoutSU-10 (10 exercise classes).

   @inproceedings{negin2013decision,
     title = {A decision forest based feature selection framework for action recognition from rgb-depth cameras},
     author = {Negin, Farhood and {\"O}zdemir, F{\i}rat and Akg{\"u}l, Ceyhun Burak and Y{\"u}ksel, Kamer Ali and Er{\c{c}}il, Ayt{\"u}l},
     booktitle = {Image Analysis and Recognition: 10th International Conference, ICIAR 2013, P{\'o}voa do Varzim, Portugal, June 26-28, 2013. Proceedings 10},
     pages = {648--657},
     year = {2013},
     organization = {Springer Berlin Heidelberg},
     doi = {10.1007/978-3-642-39094-4_74},
     preview_alt = {Color-coded 3D skeleton with joint groups and coordinate axes.},
     preview_note = {Thumbnail adapted from the corresponding paper.}
   }