OWL Challenge | Interpretable AI in Pathology, WACV 2027

From prediction to explanation

Nearly 30% of patients experience recurrence after radical prostatectomy, and deep learning can already predict that risk directly from digitised slides, often beyond what current clinical grading captures. What those models will not tell you is which morphology drives the prediction, or whether they lean on a clinically meaningful structure or on a spurious correlation.

OWL builds upon earlier challenges from the same group. PANDA showed that models can reproduce a pathologist's task, Gleason grading, at expert level, but only within the ceiling of existing human-defined patterns. LEOPARD dropped the annotations and predicted recurrence from outcome data alone, with strong predictive accuracy across international cohorts, but the winning models stayed black boxes. OWL keeps the discovery setting and adds the missing question: is the explanation any good?

The scientific question behind the challenge: what kind of tissue patterns do highly performant AI recurrence-risk models actually base their predictions on?

The task

Interpretable prediction of recurrence risk

For every case, a submitted algorithm must produce four things — a prediction, where it looked, and why.

1. Time to BCR in years

A predicted time-to-biochemical recurrence (BCR) risk score for the case, learned under weak supervision from patient-level outcomes only.

2. Risk score

A predicted biochemical recurrence (BCR) risk score ("low", "medium", "high") for the case, learned under weak supervision from patient-level outcomes only. This score should be derived from the participant's algorithm's own time-to-recurrence predictions.

3. Attribution map

A spatial attribution map highlighting the regions relevant to the prediction, as an 8-bit unsigned integer TIFF.

4. Interpretability report

A free-form PDF of at most 5 pages with a human-readable interpretation of the decision for that case: concept attributions, textual explanation, and so on.

The difficulty is real: whole-slide images are gigapixel-scale and heterogeneous, the predictive patterns can be sparse and spread across scales, and in genuine biomarker discovery there is by definition no interpretability ground truth to compare against.

Data

H&E-stained prostatectomy whole-slide images with biochemical recurrence status and time-to-event or censoring information, from three medical centres on two continents.

Split	Cases	Availability
Training	508	Public, via the AWS Open Data Initiative under CC BY-NC-SA
Test	861	Private, internal and external cohorts, hidden from participants

Radboudumc, Netherlands

University Hospital Cologne, Germany

Instituto de Medicina Patológica, Brazil

All patient data are fully anonymised, and collection, sharing and research use were approved by the institutional review boards and ethics committees of the participating institutions. Validation and test data and labels stay hidden, and the private cohorts have not been used to develop publicly available foundation models — which keeps contamination and memorisation out of the benchmark.

Evaluation

Three axes, in two stages

No single metric can decide whether an explanation is faithful, clinically meaningful and useful for discovery, so OWL combines automatic evaluation with a structured expert reader study.

Debug phase

Participants can submit their containers for a sanity check on training cases to see if the output on the platform matches the local one.

Test phase, qualification

Models are ranked by C-index on the hidden test cohort. Those above C-index 0.65 advance to the interpretability stage; below that threshold the prognostic signal is too weak for the explanation to mean much.

Automatic faithfulness

Perturbation analysis: regions the attribution map marks as highly relevant are progressively removed, and the change in the model's output is compared against perturbing random regions. A larger gap means a more faithful explanation.

AI expert panel

Manuscripts, code, metrics, predictions, attribution maps and reports reviewed by at least two experts each, on Likert-scale questionnaires covering clarity, scientific plausibility, meaningful novelty, usefulness for understanding model behaviour, transferability to other pathology problems, and risk of hallucination.

Pathology expert panel

Practising pathologists and trainees independently assess a subset of cases per submission, at least two experts each, on clinical meaningfulness, alignment with the image evidence, usefulness for understanding and questioning the prediction, and support for clinical understanding.

Final ranking

A weighted sum of the aggregated scores.

Reader studies run through the Grand Challenge reader-study functionality; all algorithms are executed in a controlled environment for reproducibility. Confidence intervals come from bootstrapping over patient cases, and significance from a two-sided permutation test.

Taking part

How to participate

Join on Grand Challenge
Create a verified account, then join the challenge — joining confirms that you accept the challenge rules.
Download the public training data
508 training cases are available through the AWS Open Data Initiative.
Train locally, then containerise
Models are developed on your own hardware. Wrap the trained algorithm and its weights into a Docker container: the platform runs it in an offline AWS environment, so nothing can be downloaded at inference time.
Submit through the two phases
Debug (3 training cases, several submissions per day, logs visible) to check that cloud and local predictions match; Test (1 submission per team).
Present at the workshop
Top-ranking teams present their methods during the one-hour OWL session in the morning program, and take part in a joint benchmark publication with up to 5 co-authors per submission.

Compute limits. During evaluation each algorithm gets an NVIDIA T4 with 16 GB VRAM or an A10G with 24 GB VRAM, and must finish in under 59 minutes per case. The platform enforces both.

Papers are optional. Challenge participants are invited, but not required, to submit a full paper to the CVPath track. Submissions by challenge participants and by organizers are handled by independent chairs, so that nobody reviews their own work.

Rules

Rules in brief

The full and authoritative rules live on the challenge website; this is the short version.

Every participant belongs to exactly one team; a team may consist of a single participant.
Anonymous participation is not allowed: to appear on the leaderboards, real names and affiliations must be shown on a verified Grand Challenge profile.
Members of the organizing and sponsoring centres may take part, but are not eligible for prizes or for the final ranking.
This is a code-execution challenge: you submit a Docker container, not prediction files. Containers run offline, with no internet access, so all weights and resources must be inside the image.
External data and pre-trained models are allowed when freely and publicly available under a permissive open-source licence, and must be declared in the algorithm name, the algorithm page or a supporting publication.
The final submission must be accompanied by a methodology manuscript (template provided), a URL to the source code under a permissive licence, and a copy of the submitted container.
Participants must respect the compute limits and the publication embargo period. The organizers may disqualify unfair or dishonest practice at any point.

Timeline

Challenge schedule

Separate from the CVPath paper deadlines, which are listed on the CVPath page.

Phase	What happens
27.08.2026 - 20.11.2026 Launch	Starter kit, baseline, data loaders, Docker templates and evaluation scripts released; registration opens on Grand Challenge
20.09.2026 - 20.11.2026 Debug	Sanity check of submissions.
20.10.2026 - 20.11.2026 Test	One submission per team on the hidden test cohort; models above C-index 0.60 qualify for the interpretability stage
Expert reader study	AI and pathology panels assess the qualifying submissions
Results	Results announcement, ranking and presentations by the top-ranking teams in the OWL session, on the CVPATH workshop day

The phase dates for this edition are published on the challenge website, together with any change to the schedule.

owl.grand-challenge.org Starter kit on GitHub

Who runs it

Challenge organizing team

Bringing together computational pathology, explainable AI and clinical uropathology. The panel of clinical evaluators is larger than the list below and is finalised before the manual evaluation phase.