News & Highlights

DIRECT v2.2.0 — PyPI, new docs, and Hugging Face weights

After a long stretch of behind-the-scenes work, we released DIRECT v2.2.0. The toolkit is now installable from PyPI, ships with a redesigned documentation site, and hosts pretrained weights on Hugging Face — bringing code, experiments, docs, and models into one place.

This release also lands code and configs from several of our recent papers: adaptive k-space sampling, static and dynamic MRI reconstruction, acquisition-conditioned models, and motion registration.

Install with:

pip install direct-recon

New preprint: MIRAGE for MRI contrast enhancement

Our new preprint MIRAGE: Multi-scale Lesion-Informed Representation with Auxiliary Guidance for MRI Contrast Enhancement is now on arXiv (arXiv:2607.19137).

With Andrea Borghesi, Xin Wang, and Jonas Teuwen, we introduce a residual 2D U-Net that combines global reconstruction and perceptual losses with lesion-aware supervision for inferring contrast enhancement from pre-contrast breast MRI — improving lesion localization over pix2pix, conditional diffusion, and latent bridge-matching baselines on the multi-centre MAMA-SYNTH dataset.

MIDL 2026 in Taipei — two poster presentations

I am attending the 9th International Conference on Medical Imaging with Deep Learning (MIDL 2026) in Taipei, Taiwan, 8–10 July at the Chientan Youth Activity Center — the first MIDL edition held in Asia.

I am presenting two works as posters: End-to-End Co-Optimization of Adaptive k-Space Sampling and Reconstruction for Dynamic MRI and Conditional Learned Reconstruction for Medical Imaging (with Nikita Moriakov, Jan-Jakob Sonke, and Jonas Teuwen).

fomo.fo team wins CVPR CT foundation models challenge

Team fomofo from the Foundation Models for Oncology Lab placed first on Task 1 (All-data Track) of the CVPR 2026 Challenge on Foundation Models for General CT Image Diagnosis, evaluating frozen pretrained 3D CT representations with linear probing across chest and abdomen.

Our submission using TAP-CT achieved the top balanced accuracy (0.654) and AUROC (0.714) on the public leaderboard — part of the foundation model work I contribute to through AIFOFOMO at NKI.

Digital poster at ISMRM 2026, Cape Town

Our group is at ISMRM 2026 in Cape Town, where I present the digital poster Joint Optimization of Acquisition, Reconstruction, and Registration for Maximizing Motion Estimation in Dynamic Cardiac MRI, together with Dr Jonas Teuwen and Dr Jan-Jakob Sonke.

The study introduces an end-to-end framework that jointly optimizes adaptive k-space sampling, image reconstruction, and deformable registration to improve motion estimation in highly undersampled dynamic cardiac MRI — shifting the focus from image fidelity alone toward motion estimation accuracy.

Registration, Atlases, and Motion session · Abstract via ISMRM portal (membership required).

PhD completed at UvA & NKI

I obtained my PhD from the Faculty of Medicine at the University of Amsterdam on 1 April 2026, becoming the first doctoral graduate from the AI for Oncology Lab. My thesis, What We Do Sample, We Must Learn to Reconstruct: From Missing k-Space Data to Meaningful Images — Deep Learning in MRI Reconstruction and Beyond, explored deep learning for accelerated MRI reconstruction, adaptive sampling, self-supervised learning, cross-domain generalization, and clinical validation.

Supervised by Prof. Dr. Ir. J.J. Sonke, Prof. Dr. Ir. C.I. Sanchez Gutierrez, and Dr. Ir. J.J.B. Teuwen. I continue as a postdoctoral researcher at NKI, working on foundation models and AI for oncology.

FastSlide open source released

Together with colleagues at NKI, I released FastSlide - a C++20 whole-slide image reader with native Python bindings for digital pathology and computational imaging. We built it because no existing Python library handled both multiplex and histopathology formats efficiently for deep learning workflows where fast random access matters more than pre-tiling.

FastSlide reads Vectra multiplex QPTIFF, SVS, MRXS, OME-TIFF, OME-Zarr, Philips iSyntax, Zeiss CZI, Ventana, Olympus VSI, and more - with native support for multi-channel fluorescence, Z-stacks, and T-series. Version 0.7.0 added Ventana and Olympus VSI support plus full focal-plane and time-point navigation. Thread-safe and PyTorch DataLoader-ready, released under Apache 2.0.