Pathologist REasoning-Guided REport Generation Challenge¶

Important dates¶

🕐 Timezone: Times are given in Korea Standard Time (KST) and Coordinated Universal Time (UTC) for the same instant. KST = UTC+9 year-round.

📎 Additional deadlines (workshop, proceedings, camera-ready) — TBD on the official challenge pages.

The Top-3 ranked teams will be officially announced during the MICCAI 2026 conference event (27 September – 1 October 2026). Certificates of Merit will be awarded to the top five performing teams. Additional monetary awards — pending confirmation from societies, industrial partners, and MICCAI SIGs.

Participation¶

REG² participation is allowed only through a verified Grand Challenge account.

To verify your account, follow Grand Challenge’s official guide: Account verification (complete your profile first, then use Request verification from your profile as described there).

Notice¶

Official notices (policy, data access, deadlines, or platform changes). Newest first.

📌 1 June 2026 — Revised CoT annotations¶

Issue: During CoT data preprocessing, some longer entries in train_CoT.json were truncated when converting to JSON. Eight training cases had incomplete chain-of-thought annotations.

Resolution: Complete CoT annotations are now available for these cases. You do not need to re-download the full training set. Use the command in Data download to fetch only the revised files. No other training files were changed in this update.

Access link, affected slide IDs, and full guidance are on Data download.

📌 31 May 2026 — Revised training slides¶

Issue: Eighteen training WSI files have been revised.

Resolution: Updated files are now available. You do not need to re-download the full training set. Use the command in Data download to fetch only these 18 cases, then replace the previous copies in your local training slide directory. No other training files were changed in this update.

Access link, folder contents, and full guidance are on Data download.

📌 25 May 2026 — Visual Grounding samples¶

Visual Grounding (Metric B) example cases are now available. They show how ROI images, identifiers, and ROI–question pairs are organized — samples for structure only; you are expected to build your own scenarios (ROIs, background regions, and questions) for training and development.

Access link, folder contents, and full guidance are on Data download.

📌 20 May 2026 — Revised training slides¶

Issue: Eleven training WSI files had incomplete TIFF metadata/index information in the original conversion.

Resolution: Re-converted files are now available. You do not need to re-download the full training set. Use the command in Data download to fetch only these 11 cases, then replace the faulty files in your local training slide directory. No other training files were changed.

📌 15 April 2026 — Save the dates¶

REG² will follow the timeline in Important dates above. Workshop, proceedings, and camera-ready deadlines remain TBD and will be posted when confirmed.

Updates¶

Short news and release announcements. Newest first (latest at the top).

📢 30 May 2026 — Debug phase Submissions are live¶

Debug phase Submissions are now open on Grand Challenge. Use this window to validate your Docker submission pipeline, fix container and output-format issues, and confirm that your algorithm produces the required Workflow Reasoning and Visual Grounding JSON artifacts before test phases begin.

Getting started: The public REG2026 GitHub repository includes a sample container and a step-by-step submission guide. Start in algorithm_submission_template/ and read algorithm_submission_template/README.md thoroughly — it is the baseline reference for how inference, both task interfaces, local tests, and upload to Grand Challenge are expected to work. Adapt your own model from that template rather than building the pipeline from scratch.

Reminders: Debug submission limits and per-case runtime requirements are in Rules & regulations (sections Development & evaluation phases and Resource constraints).

📢 15 April 2026 — Welcome to REG²¶

Pathologist Reasoning-Guided Report Generation Challenge (REG²) builds on REG2025 by benchmarking diagnostic reasoning alongside report generation. Follow Notice and Updates for releases and rule changes.

Challenge overview¶

REG² is a benchmark for generating pathology reports from gigapixel-scale whole slide images (WSIs) while making the diagnostic reasoning process explicit. The challenge builds on REG2025, which measured report generation quality but relied on metrics that only partially reflect clinical needs. REG² asks participants to produce not only reports but structured reasoning aligned with pathologist exploration and reportability decisions. An agenda that connects naturally to agentic AI (planning, tool use, and multi-step decision-making) while addressing the scarcity of rigorous medical reasoning evaluation data.

Motivation¶

Pathology examination is often the final confirmation and gold standard in diagnosis, and the written report is central to downstream clinical decisions. Producing that report is difficult: pathologists must synthesize enormous WSIs and attend to fine-grained, clinically decisive features. Recent vision–language models have shown strong results on automated report generation despite gigapixel complexity. REG2025 established a useful baseline where finalists achieved encouraging performance under that setting.

However, REG2025 evaluation leaned on surface-level lexical scores (e.g., ROUGE-L, BLEU-4) and shallow semantic tools (e.g., SciSpacy, OpenBioLLM). Those metrics do not capture what matters clinically: clinical correctness, negation and omission of safety-critical findings. REG² is designed to close that gap by making diagnostic reasoning central to data and evaluation. It focuses on how pathologists explore evidence, compare diagnoses, and decide what belongs in the report.

From REG2025 to REG²¶

REG2025 focused on how well models matched reference reports. REG² extends that goal by integrating a reasoning trajectory that precedes and supports the report: the same kind of structured thinking pathologists use to justify conclusions and explain them to colleagues. Modern agentic systems are well suited to decompose such workflows, but the community still lacks standardized, clinically grounded benchmarks for reasoning on WSIs. REG² aims to supply that reasoning benchmark and compare agentic vs non-agentic approaches on a common footing.

Publication and use of challenge data¶

Participants may not use challenge data to prepare or submit conference papers, preprints, or journal articles until the official challenge meta paper and proceedings papers are publicly available. Prior written permission from the organizers is required for any exception.