Darya Taratynova

PhD Student, TUM · CAMP Lab · MCML

I care about AI × medical imaging

Interested in —

  • - foundation models
  • - cardiac imaging
  • - machine unlearning
  • - ultrasound

Contact —

-

Darya Taratynova

A little about me

I'm a Ph.D. student at the Technical University of Munich, School of Computation, Information and Technology, supervised by Prof. Nassir Navab in the Computer Aided Medical Procedures and Augmented Reality (CAMP) lab, and a fellow of the Munich Center for Machine Learning (MCML). Previously, I received my M.Sc. in Machine Learning from MBZUAI in the UAE, where I worked with Prof. Mohammad Yaqub in the BioMedIA lab.

I work on AI for ultrasound imaging. I primarily focus on foundation models, multimodal learning, and video understanding. With my work, I aim to build models that learn general representations of ultrasound across organs, devices, and tasks.

Outside research, I like long runs, cycling, and swimming. And I have two cats, , who I found on the streets back in the UAE and who moved with me to Munich.

News

  • Started my Ph.D. at the Technical University of Munich in the CAMP lab, supervised by Prof. Nassir Navab, as a fellow of the Munich Center for Machine Learning (MCML).
  • Graduated with an M.Sc. in Machine Learning from MBZUAI and received the Outstanding Master's Research Award!
  • Our paper EchoSonar-R, a multi-view reasoning model for echocardiography, is out on arXiv!
  • Our paper CardioBench was accepted by IJCAI-ECAI 2026!
  • Co-taught the tutorial Foundation Models for Cardiac Ultrasound at the RISE-MICCAI Winter School, MBZUAI, Abu Dhabi.
  • Presented our paper TPA at BMVC 2025 in Sheffield, UK.
  • Our team won 1st place in the FLARE 2025 Challenge (Task 5) at MICCAI 2025!
  • Our team won 2nd place in the LISA Challenge at MICCAI 2025!
  • Presented our paper Forget-MI at MICCAI 2025 in Daejeon, South Korea.
  • Started my M.Sc. in Machine Learning at MBZUAI, supervised by Prof. Mohammad Yaqub in the BioMedIA lab.
  • Graduated with distinction from Nazarbayev University with a B.Sc. in Mathematics and a minor in Computer Science.

Publications

  • EchoSonar-R overview: multi-view visual and structure encoding followed by cross-modal projection of video and structure tokens

    EchoSonar-R: A Multi-View Reasoning-Enabled Model for Disease Classification and Report Generation in Echocardiography

    Darya Taratynova, Ahmed Aly, Numan Saeed, Mohammad Yaqub

    arXiv preprint, 2026

    A vision-language model that reads several echo views at once to classify multiple heart diseases and write the report, grounded in cardiac anatomy and trained with supervised fine-tuning followed by GRPO reinforcement learning.

    [paper][arXiv]
  • CardioBench overview: classification and regression tasks across eight public echocardiography datasets

    CardioBench: Do Echocardiography Foundation Models Generalize Beyond the Lab?

    Darya Taratynova*, Ahmed Aly*, Numan Saeed, Mohammad Yaqub

    IJCAI-ECAI 2026

    A benchmark that unifies eight public echocardiography datasets into nine tasks, showing where current echo foundation models transfer well and where they still fail, such as view classification and subtle pathology.

    [paper][arXiv][code]
  • TPA architecture: fetal ultrasound frames and normal and abnormal text prompts encoded and projected into a shared space

    TPA: Temporal Prompt Alignment for Fetal Congenital Heart Defect Classification

    Darya Taratynova, Alya Almsouti, Beknur Kalmakhanbet, Numan Saeed, Mohammad Yaqub

    BMVC 2025

    Classifies fetal heart defects from ultrasound video by aligning temporal video features with text prompts in a vision-language foundation model, with an uncertainty module that improves calibration.

    [paper][arXiv][code]
  • Forget-MI framework: forget, noisy and retain sets fed to the frozen original model and the tuned unlearning model

    Forget-MI: Machine Unlearning for Forgetting Multimodal Information in Healthcare Settings

    Shahad Hardan*, Darya Taratynova*, Abdelmajid Essofi, Karthik Nandakumar, Mohammad Yaqub

    MICCAI 2025

    Removes a patient's image and report data, together and separately, from a trained multimodal medical model, coming close to the behaviour of a model retrained without that data.

    [paper][arXiv][code]

Services

Teaching

Tutorial, Foundation Models for Cardiac Ultrasound, RISE-MICCAI Winter School, MBZUAI,

Mentoring

Mentor, MBZUAI NextGen Summer Program,

Talks and Presentations

Talk, Foundation Models and Their Adaptation to Echocardiography, BioMedIA Lab, MBZUAI,