CORTEX

A Structured Reasoning Benchmark for Trustworthy 3D Chest CT MLLMs

CORTEX (Clinically Organized Reasoning and sTructured EXplanation) is a structured reasoning benchmark for multimodal large language models working with 3D chest CT. It converts chest CT question-answering and report-generation tasks into auditable, radiologist-inspired reasoning traces.

This Hugging Face repository is the data release for the paper:

Hashmat Shadab Malik, Anees Ur Rehman Hashmi, Numan Saeed, Muzammal Naseer, Salman Khan, and Christoph Lippert. “CORTEX: A Structured Reasoning Benchmark for Trustworthy 3D Chest CT MLLMs.”

• Paper: arXiv:2606.27264
• Venue: SAFER 2026, MICCAI 2026
• Github repo: CORTEX GitHub repository

Dataset summary

The release contains three task families. CORTEX was constructed from the validation split of CT-RATE.

Configuration	File	Examples in this release	Description
closedended	data/closedended.json	8,910	Multiple-choice chest CT VQA with structured reasoning
openended	data/openended.json	64,211	Open-ended chest CT VQA with structured reasoning
report	data/report.json	3,039	Chest CT report-generation examples with structured reasoning
Total	—	76,160	—

Dataset structure

Each configuration is stored as a JSON array. Records contain:

Field	Type	Description
id	string	Sample identifier
ct_rate_id	string	Identifier inherited from the source VQA in CT-RATE
type	string	closedended, openended, or report
image	string	Referenced CT-RATE volume filename
conversations	list	Ordered human/model messages

The conversations field follows the common conversational format used by multimodal instruction datasets. Model messages contain a structured reasoning trace with four stages:

1. <task> — task and clinical-context understanding
2. <observation> — visual assessment organized by anatomy
3. <reason> — option or hypothesis evaluation
4. <answer> — final answer and rationale, or report synthesis

Prompts

The task-specific prompts used to generated the reasoning traces in this dataset are included for reproducibility:

prompts/
├── generation/
│   ├── closedended.md
│   ├── openended.md
│   └── report.md
└── scoring/
    ├── closedended.md
    ├── openended.md
    └── report.md

Generation prompts produce the four-stage CORTEX reasoning traces. Scoring prompts apply the corresponding automated evaluation.

Download

from huggingface_hub import snapshot_download

local_path = snapshot_download(
    repo_id="aneesurhashmi/cortex",
    repo_type="dataset",
    local_dir="./cortex",
)

print(f"Downloaded to: {local_path}")

Intended use

CORTEX is intended for research on multimodal clinical question answering, radiology report generation, structured diagnostic reasoning, and stage-level evaluation of reasoning traces.

This dataset is not a clinical decision-support system. It must not be used as a substitute for qualified radiologist interpretation or patient care. Users should review the source dataset’s license, privacy documentation, data-use terms, and institutional requirements before downloading, using, or redistributing the data.

Data provenance

CORTEX was constructed from the validation split of CT-RATE, a chest CT question-answering dataset. The reasoning traces were generated with task-specific prompts and selected using expert-in-the-loop + rubric-based verification. Please read the paper for the full generation and validation methodology.

Limitations

• The benchmark is limited to chest CT and English-language text.
• Reasoning traces are structured model-generated annotations and should not be treated as independent expert reports.
• The data snapshot and paper totals differ as documented above.
• Dataset access, licensing, de-identification details, and CT-volume hosting should be confirmed from the final release configuration.

Citation

@article{malik2026cortex,
  title   = {CORTEX: A Structured Reasoning Benchmark for Trustworthy 3D Chest CT MLLMs},
  author  = {Malik, Hashmat Shadab and Hashmi, Anees Ur Rehman and Saeed, Numan and Naseer, Muzammal and Khan, Salman and Lippert, Christoph},
  journal = {arXiv preprint arXiv:2606.27264},
  year    = {2026}
}