CardioLens is available for approved non-commercial research. Please download the Data Access Form, complete and sign it, and email the signed form to zixian12138@163.com with your Hugging Face username.

CardioLens

CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations
NeurIPS 2026 — Evaluations & Datasets Track

Paper · Code & Evaluation · Dataset Files

Overview

CardioLens is a clinically grounded benchmark for evaluating multimodal large language models (MLLMs) on multi-sequence cardiovascular magnetic resonance (CMR). Unlike conventional medical visual question answering based primarily on isolated images, CardioLens examines whether models can interpret findings distributed across anatomical views, cardiac phases, and complementary imaging sequences.

The benchmark covers Cine, Late Gadolinium Enhancement (LGE), Perfusion, and T2-weighted CMR, with three independently evaluated tiers of increasing clinical demands:

Image Understanding: recognize clinically defined findings within individual CMR sequences.
Report Generation: synthesize findings into structured, sequence-level clinical reports.
Disease Diagnosis: predict patient-level diagnostic categories from multi-sequence CMR images.

These are evaluation tiers, not a mandatory chained inference pipeline. Each tier is evaluated on its corresponding input and targets; the diagnosis task does not require a model-generated report from Tier 2.

Dataset Statistics

The hosted release contains 100 patients from the 585-patient research benchmark evaluated in the paper.

	Full research benchmark (paper)	Hosted release (this repository)
Patients	585	100
CMR image slices	473,896	78,707
Verified QA items	13,494	2,297
Sequences	Cine, LGE, Perfusion, T2	Cine, LGE, Perfusion, T2

The hosted release includes 1,805 image-understanding QA items, 361 report-generation items, and 131 diagnosis QA items (2,297 in total). The same three task categories and four CMR sequence types are represented in the release.

Available Files

File	Contents
`CardioLens_NIfTI.zip`	Released CMR studies in NIfTI format
`CardioLens_VQA.zip`	VQA / benchmark release materials
`Reports.xlsx`	Released report references
`Diagnosis_Label.xlsx`	Released patient-level diagnosis labels

For data preprocessing, prompts, slice-selection protocols, model inference, and evaluation scripts, see the GitHub repository.

Request Access and Download

CardioLens is a gated research dataset. The dataset page and documentation are public, but downloading the data requires individual Hugging Face approval.

To request access:

Download and complete the CardioLens Data Access Form (PDF), then sign it.
Sign in to Hugging Face and submit the access request.
Email the signed PDF to zixian12138@163.com with the subject CardioLens Access - [HF username] - [Institution]. Include the same institution and Hugging Face username on the form.
Your request will be reviewed after both submissions are received.

Hugging Face does not support PDF attachments in gated-access requests; please submit the signed form by email. Do not upload completed forms to public repositories.

After approval, download the release using a logged-in account:

pip install -U huggingface_hub
hf auth login
hf download kiwi12138/CardioLens --repo-type dataset --local-dir ./CardioLens-data

Or download only the NIfTI archive:

hf download kiwi12138/CardioLens CardioLens_NIfTI.zip --repo-type dataset --local-dir ./CardioLens-data

Access is granted per user, not to an entire organization. Downloaded files remain subject to the published dataset license and applicable privacy and institutional requirements.

Benchmark Tasks

Tier	Input	Evaluation target
1 · Image Understanding	Selected images/frames from an individual sequence	Single-label or multi-label findings including ventricular function, myocardial enhancement, edema, perfusion abnormalities, and anatomical localization
2 · Report Generation	Selected views/frames from an individual sequence	A coherent clinical report covering the sequence-specific findings
3 · Disease Diagnosis	Selected images/frames from multiple CMR sequences	Etiology (normal heart, ischemic heart disease, non-ischemic cardiomyopathy) and non-ischemic subtypes

The diagnostic subtypes in the benchmark include hypertrophic, dilated, restrictive, and arrhythmogenic cardiomyopathy, as well as myocarditis. Please refer to the paper for the complete task taxonomy, prompts, metrics, and protocol definitions.

Data Construction and Quality Control

The benchmark was constructed from non-public clinical CMR archives. Original clinical reports were reviewed by experienced radiologists, converted into structured task labels, and checked against source reports; ambiguous cases were reviewed further. Patient-level diagnosis labels were based on clinical discharge records. CMR studies were de-identified before release.

For the main experiments, the paper evaluates 24 MLLMs and compares random, clinically motivated heuristic, and data-driven MIL slice-selection protocols. The results expose a gap between performance on conventional image-level medical benchmarks and clinically grounded, study-level CMR interpretation.

Intended Use and Limitations

Intended use: non-commercial research and benchmarking of medical multimodal models, including image understanding, clinical report generation, and disease classification from CMR.

Not intended for: direct clinical diagnosis, treatment decisions, patient management, or re-identification of individuals.

Important limitations include variation in CMR acquisition, sequence availability, and cohort composition; uneven representation of rare diseases; and the need to select a limited number of frames or slices when evaluating MLLMs that cannot process full 3D/4D studies. The hosted 100-patient subset is smaller than the full research cohort, especially for rare diagnosis categories. Results on the subset should not be assumed to exactly reproduce every full-cohort metric.

Use of this repository is subject to the CC BY-NC-SA 4.0 license shown on the Hub and to the applicable dataset access conditions. Users are responsible for ensuring that their intended use complies with those terms.

Paper and Code

Paper: CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations
Official evaluation code: github.com/kiwi12138/CardioLens

If you use CardioLens in research, please cite the accompanying paper.

Downloads last month: 61

Total file size:

3.72 GB