Using masked autoencoders and DINOv2, we developed self-supervised SCN foundation models that learn transferable representations from unlabelled images. These models were adapted to diverse downstream tasks, including atopic dermatitis severity, contact dermatitis patch-test reactions, ultraviolet radiation exposure, demographic characteristics, geographic origin, and anatomical site classification.

The results show that SCN representations capture clinically and biologically meaningful variation, with the strongest performance observed for atopic dermatitis severity and occupational ultraviolet radiation exposure. Reduced performance in cross-site transfer experiments also revealed substantial site-related domain shift, underscoring the need for domain adaptation and prospective validation before clinical deployment. The broader goal is to establish scalable, objective methods for skin barrier assessment, disease monitoring, and digital biomarker discovery.

The project develops uncertainty-aware generative methods that learn entire conditional distributions rather than producing a single average prediction. These models can be used to predict gene expression from tissue images and, more broadly, to link or translate between imaging and molecular modalities while preserving biologically meaningful variation.

A central part of the research concerns how such models should be evaluated. Conventional point-prediction metrics may reward average-like outputs without revealing whether a model captures heterogeneity, represents multiple plausible outcomes, or generalizes across experimental settings. The project therefore investigates evaluation strategies that assess distributional fidelity, uncertainty calibration, biological structure, and robustness to technical and biological variation. The broader goal is to make generative models for spatial biology both more expressive and more trustworthy.

While current spatial transcriptomics technologies provide detailed molecular information from two-dimensional tissue sections, our goal is to use AI to reconstruct and analyze the three-dimensional organization of tumors and their microenvironments.

By integrating cutting-edge AI with state-of-the-art spatial omics technologies, this interdisciplinary collaboration aims to establish a new computational framework for studying cancer biology in three dimensions, opening new opportunities for precision oncology and data-driven biomedical discovery.

We thank Digital Futures for their support and look forward to sharing the scientific advances that will emerge from this exciting initiative.

The funded project, Linking Omics and Images with Deep Learning to Decode Cellular Responses, combines the complementary expertise of the two groups to better understand how genetic perturbations shape cellular behavior. The project will integrate image-based profiling and omics-based modeling through deep learning, creating a framework that links molecular signaling networks with cellular phenotypes. By combining high-content microscopy data with molecular measurements, the project aims to uncover mechanistic relationships between cellular state and morphology in PDAC.

We are grateful to PALS and the Knut and Alice Wallenberg Foundation for their support and look forward to strengthening this exciting collaboration between groups at KTH and KI.

The project brings together expertise in advanced microscopy, artificial intelligence, and kidney pathology to develop a next-generation precision medicine pipeline for quantitative disease assessment. A central part of the project will be carried out by David Unnersjö Jess and Jen-Hung Wang, who will work on the implementation and development of the imaging and machine learning components of the pipeline.

The project will explore state-of-the-art self-supervised learning approaches to automatically identify, segment, and quantify disease-relevant structures from large-scale 3D microscopy data. By combining high-resolution imaging of intact kidney biopsies with AI-assisted image analysis, the project aims to improve diagnostic precision while reducing the time and effort required for pathology workflows.

We thank Digital Futures for their support and look forward to sharing the progress and outcomes of this exciting initiative.

Project title: Microscopy Foundation Model & AI-Ready Data Hub for Smart Microscopy

The initiative is led by our collaborator Wei Ouyang and brings together researchers from across the SciLifeLab community to advance the next generation of AI-driven microscopy.

Gisele Miranda is one of the consortium PIs. She will contribute with expertise in machine learning and image analysis as part of this interdisciplinary effort.

The project aims to build an AI-ready microscopy data hub as a foundation for training robust models capable of interpreting samples in real time and guiding microscopy acquisition on the fly.

Max will contribute to the implementation, development, and dissemination of state-of-the-art image analysis workflows for the life science community. His work will support researchers across a wide range of imaging modalities and biological applications, helping to advance reproducible and scalable image-based research.

We are excited to have Maximilian join the team and look forward to his contributions to the facility and the broader imaging community.

Jen-Hung’s research focuses on image-based profiling and computational approaches to understanding cellular phenotypes, while Felipe works on machine learning methods for integrating multimodal biological data.