BrainTRACE

Tracing Longitudinal, Multimodal, and Volumetric Evidence in Brain MRI Clinical Reasoning

Accepted at NeurIPS 2026, Evaluations & Datasets Track (Poster).

Maintained by Qizhen Lan. Contact: qizhen.lan@uth.tmc.edu.

Paper / OpenReview · Code · Scoring

Dataset overview

BrainTRACE evaluates vision-language models on longitudinal brain MRI, spanning recognition, within-study reasoning, interval comparison, multi-timepoint trajectories, and clinical synthesis. Its source cohort contains 1,778 patients and 7,299 studies from MR-RATE.

The release has 6,923 source rows. Each of the 70 case-reasoning rows expands into six steps, giving 7,273 scored VQA instances in total.

Track (track value)	Source rows	Scored instances
Broad-compatible (broadQA)	5,653	5,653
3D-evidence (3D)	1,200	1,200
Case reasoning (clinical_reasoning_QA)	70 cases	420 steps
Total	6,923	7,273

The 30 single-turn task templates contain 5,176 closed-form and 1,677 open-ended items. Case reasoning adds one six-step template. The corresponding answer_type values in the Parquet are closed_form, open_ended, and chain.

This repository releases questions, targets, rubrics, evidence pointers, and cohort metadata. MRI images, source radiology reports, rendered PNGs, and volumes are not included. To recover visual inputs, obtain MR-RATE from its upstream maintainers under their Data Use Agreement and render the evidence locally. See ATTRIBUTION.md.

Load the benchmark

from datasets import load_dataset

dataset = load_dataset("lanqz7766/BrainTRACE", split="test")
print(len(dataset))  # 6923 source rows
print(dataset[0]["question"])

This is the formal release maintained by Qizhen Lan. The original anonymous review release remains available for existing links.

The only split is test; this is an evaluation benchmark with no training or development split. Each row contains model inputs and hidden scoring references. Pass only the permitted visual inputs and questions to the evaluated model. Keep targets, source reports, masks, measurements, and rubrics out of its input.

Data fields (24 columns)

Field	Contents
item_id	Stable item identifier
track, subset	Evaluation track; v1 or v2 for the 3D track
level, sub_category, template	Reasoning level (1–5), task family, and template identifier
answer_type	closed_form, open_ended, or chain
question, options	Model-facing question and multiple-choice options, when applicable
gt_value, reference_answer	Closed-form target or reference answer
rubric_json, gt_format	JSON-encoded scoring reference and answer format
patient_id_anon, study_uids	Upstream pseudonymous identifiers for recovering evidence
shown_tp_labels, n_shown_tps, shown_relative_days	Timepoint order, count, and relative timing
input_modalities, input_planes, slice_selection_rule	Rendering specifications
image_relpaths, volume_relpath	Expected local evidence paths; the files must be rendered separately
chain_steps	JSON-encoded list of six case-reasoning prompts; null for single-turn items

String fields containing JSON must be decoded before use. The Parquet schema defines the column types; inapplicable fields can be null or empty.

Recover visual inputs

1. Obtain MR-RATE and accept its upstream access terms and DUA.
2. Download this repository's metadata to ./braintrace_dataset.
3. Clone the public code repository and run the renderer against your local MR-RATE download:

git clone https://github.com/lanqz7766/BrainTRACE.git
cd BrainTRACE
python -m pip install -r requirements.txt

python reproduction/render_images.py \
  --dataset /path/to/braintrace_dataset \
  --mr-rate-root /path/to/MR-RATE \
  --out-root /path/to/braintrace_dataset

The rendering guide describes the required MR-RATE directory layout and local PNG/volume outputs. Pseudonymous patient and study identifiers link each item to the upstream evidence.

The cohort/ directory contains manifests for all 1,778 source patients and 7,299 studies, aggregate demographics and scanner metadata, and schema documentation. These manifests can help scope an upstream download. They do not supply images, reports, or patient-level pathology subgroup labels.

Evaluation and scoring

Use the inference guide to produce model outputs. The scoring entry point is:

python scoring/score.py \
  --dataset /path/to/braintrace_dataset \
  --predictions /path/to/predictions.jsonl \
  --out-dir /path/to/scores \
  --judge-model gpt-4o-mini-2024-07-18

Closed-form scoring matches predictions to the released targets. Open-ended scoring uses the item rubrics and a pinned LLM judge. Case reasoning reports step-level performance and the fraction of cases that pass all six steps. Add --mode closed to score closed-form items without judge calls.

The released scorer defaults to gpt-4o-mini-2024-07-18; changing the judge changes the evaluation conditions. Check prediction coverage before comparing runs, because rows without a prediction are excluded by the released scorer. Record model, renderer, decoding, and judge settings alongside scores. See the scorer guide, rubric protocol, and paper results.

Intended use and limitations

BrainTRACE is intended for research on visual and longitudinal reasoning in medical VLMs. Its scores have not been validated for clinical decision support or patient-facing use.

• The benchmark uses a selected longitudinal MR-RATE cohort. Its population, scanner mix, and follow-up patterns limit generalization to other settings; cohort summaries are available in cohort/cohort_summary.json.
• Targets are grounded in source reports, structured metadata, and segmentation-derived measurements, which can contain errors. Documented clinical item review covers the 70 case trajectories and 420 steps; it does not establish an independent second clinical read of every benchmark item.
• Open-ended and case scores depend on the judge and rubric. The paper reports human and cross-judge validation within their stated sample sizes.
• The release contains evaluation targets. Training on these items would compromise subsequent evaluation on this benchmark.

License and privacy

BrainTRACE task definitions, rubrics, dataset card, and released metadata use CC-BY-NC-SA 4.0. Companion code uses Apache-2.0. MR-RATE images, reports, and imagery derivatives remain governed by the upstream DUA; the BrainTRACE license does not override it.

The cohort files contain upstream pseudonymous identifiers and demographic/scanner metadata, including date-shifted study dates. Source reports and images are not redistributed. Pseudonymization does not eliminate re-identification risk; follow the upstream DUA and do not attempt to identify participants. See ATTRIBUTION.md and the cohort schema.

Citation

Please cite both BrainTRACE and the upstream MR-RATE dataset when using this benchmark.

@inproceedings{lan2026braintrace,
  title     = {BrainTRACE: Tracing Longitudinal, Multimodal, and Volumetric Evidence in Brain MRI Clinical Reasoning},
  author    = {Lan, Qizhen and Fan, Mengchen and Zhang, Hang and Duan, Jingwei and Lin, Moule and Chen, Jialin and Geng, Baocheng and Jiang, Xiaoqian},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  note      = {Evaluations and Datasets Track},
  url       = {https://openreview.net/forum?id=hXktXOmwp6}
}

Dataset version: 1.0.0. Report data or documentation issues through GitHub Issues, or contact Qizhen Lan.