CohortLens · Fraunhofer MEVIS

Quantitative readouts from small cohorts.

CohortLens builds segmentation and stratification models from a handful of annotated examples, so that cohorts too small for conventional deep learning still yield reproducible image readouts. We develop it at Fraunhofer MEVIS, and it is available for feasibility studies, licensing, and product development under an ISO 13485-certified quality management system.

Talk to us about your cohort Read how it works
Prostatectomy slide with regions weighted by a recurrence-risk model; red marks high attention
Regions weighted by a recurrence-risk model on a prostatectomy slide. Red marks high attention. Results from the BMBF-funded project ProSurvival.

What CohortLens does

Rapid segmentation of medical images

From as few as three or four annotated example regions per structure, CohortLens trains a segmentation model for your cohort and applies it to every image. Typical targets are tumor area, invasive margin, inflammatory infiltrate, collagen and fibrosis, cell boundaries, and marker-positive regions in IHC.

The interactive annotation and segmentation tool in CohortLens, predicting glands on a colon biopsy from a handful of examples.

You receive quantification of the segmented image features, and exportable annotations for further analysis.

Cohort stratification from study endpoints

Given an endpoint per subject, such as subtype, treatment response, mutation status, or survival, CohortLens stratifies your cohort and returns a score for each subject. It also highlights the tissue regions that carry the signal, uncovering the morphological patterns behind an endpoint and pointing toward the mechanisms that produce them.

Prostatectomy slide, H and E stained, shown as a pathologist sees it

The attention map shows the weight the model gave each region when estimating recurrence risk, from blue through to red.

You receive cross-validated endpoint scores and the morphological patterns associated with the endpoint as a starting point for understanding disease mechanisms.

CohortLens is a research tool. It is not a medical device and its outputs are not intended for diagnostic use.

How it works

1You bring slides and a question

For segmentation, a handful of marked example regions. For stratification, an endpoint per subject such as recurrence, response, or mutation status.

2CohortLens adapts a foundation model to your cohort

The image encoder is already pretrained across many tissues, stains and tasks. Only the task-specific part is trained on your data, which is why small cohorts suffice and why nothing about your data enters a shared model.

3You receive readouts you can check

The result are cross-validated per-subject scores or measurements and the slide regions that carried the signal. For unusual applications, further pretraining builds your own custom foundation model, matched to your tissue and imaging modality.

Evidence

1st

UNICORN 2025 benchmark

In the MICCAI 2025 UNICORN lighthouse challenge across 20 pathology, radiology and language tasks, the MEVIS submission ranked first overall (score 0.456 against a baseline of 0.378) and first in the pathology and radiology vision tracks.

Results at grand-challenge.org

6 %

of the data of self-supervised models

Tissue Concepts, an earlier version of the model, was trained on 912 157 patches from 7 042 whole-slide images across 16 tasks and reached an AUC of 0.990 on five-fold prostate grading, with comparable results in breast and colorectal tasks.

Nicke et al., Computers in Biology and Medicine, 2025

1 %

of the training data for in-domain tasks

UMedPT, the first model using our multi-task pretraining approach, was pretrained on 17 tasks from 15 openly available datasets in histology, tomography and X-ray. The frozen encoder matched ImageNet pretraining with 1 % of a task's training data and reached 99.7 % AUC in cross-center colorectal cancer classification.

Schäfer et al., Nature Computational Science, 2024

5 %

of the compute, one slide-level model

Whole Slide Concepts, the slide-level model in the CohortLens family, combines subtyping, risk estimation and mutation prediction in one supervised model, outperforms self-supervised models on seven benchmark tasks, and was trained on openly available data only. Code and weights are public.

Nicke et al., arXiv 2507.05742, 2025

Example application

Recurrence risk in prostate cancer: the ProSurvival project.

Question
After surgery for prostate cancer, some patients relapse and many do not. Grading the tissue under the microscope separates the two groups only in part. Can the slide itself say more?
Data
The model was trained on slides from around 1 300 patients and then tested on three independent cohorts it had never seen, from TCGA hold-out centers, the Charité in Berlin and University Hospital Frankfurt.
Result
On all three cohorts the model ranked patients by relapse risk more accurately than the established ISUP grading. It also separated risk within a single grade group, where grading offers no distinction at all. Eight pathologists reviewed the regions the model identified as most relevant to its risk predictions. They identified cribriform growth, associated with relapse, and inflammatory patterns, associated with lower risk—findings which is consistent with current prostate cancer research.
Status
Nicke et al., manuscript under review. Joint work with Charité – Universitätsmedizin Berlin and the Dr. Senckenberg Institute of Pathology at University Hospital Frankfurt, Goethe University.

How we work together

CohortLens is developed at Fraunhofer MEVIS and is available to pharmaceutical and diagnostics companies, research groups, CROs and laboratories that need quantitative readouts from tissue images, in preclinical and clinical studies alike.

Feasibility study

You bring a question, slides, and where relevant an endpoint per subject. Scope and duration are agreed per project. A segmentation model can be built from a few slides; endpoint prediction usually needs on the order of a hundred subjects with known endpoints, depending on the endpoint and the strength of the signal.

You receive a report with per-subject analysis results, and a recommendation on how to move on.

License

Models and pipeline run in your environment, on your infrastructure, and we provide support and updates. Terms are set per project and cover research use; a route to clinical use is available through a product development partnership.

You receive the software, its documentation, and a long-term partner.

Product development partnership

From validated prototype to software under IVDR, developed within the MEVIS quality management system certified to EN ISO 13485 and with an IEC 62304 software lifecycle. MEVIS acts as development partner; you hold the product.

You receive development under an ISO 13485-certified quality management system, documented for the route into routine use.

We are a research group. Our technology advances through joint projects with application partners and we love ambitious research projects. We are more than open to joint grant proposals.

Your data stays yours

You keep ownership of your data. Analysis can run on your infrastructure or on ours, and nothing you send us enters a shared model.

Path to clinical use

Fraunhofer MEVIS develops software under a quality management system certified to EN ISO 13485 and follows IEC 62304 for the software lifecycle. This allows CohortLens components to be used in clinical studies and, where a partner intends to bring a readout to market, to be developed further into a regulated product.

Who you will work with

You work directly with the group that builds CohortLens: researchers and software engineers in computational pathology at Fraunhofer MEVIS, the same people who publish the methods behind it.

Open science. The research code and baseline weights behind our publications are available on GitHub.

Talk to us about your cohort

Tell us what you want to measure and roughly how many subjects or slides you have. We usually reply within a few days.

Prefer a conversation? Write to cohortlens@mevis.fraunhofer.de or schedule a call for 30 minutes.

The images shown here are in whole or part based upon data generated by the TCGA Research Network: cancer.gov/tcga. CohortLens is a research tool and not a medical device.