NOVA: A Benchmark for Anomaly Localization and Clinical Reasoning in Brain MRI

An open-world generalization benchmark under clinical distribution shift

Dataset on 🤗 Hugging Face
For academic, non-commercial use only

🔖 Citation

If you find this dataset useful in your work, please consider citing it:

📖 Click to show citation (⚠️ **Double-blind review warning**: If you are a reviewer, please **do not expand** this section if anonymity must be preserved.)

@article{bercea2025nova,
  title={NOVA: A Benchmark for Anomaly Localization and Clinical Reasoning in Brain MRI},
  author={Bercea, Cosmin I. and Li, Jun and Raffler, Philipp and Riedel, Evamaria O. and Schmitzer, Lena and Kurz, Angela and Bitzer, Felix and Roßmüller, Paula and Canisius, Julian and Beyrle, Mirjam L. and others},
  journal={arXiv preprint arxiv:2505.14064},
  year={2025},
  note={Preprint. Under review.}
}

Paper

📦 Usage

from datasets import load_dataset
dataset = load_dataset("Ano-2090/Nova")

🧪 Try it out in Colab

You can explore the NOVA dataset directly in your browser using the interactive notebook below:

Or open the notebook directly: 👉 example.ipynb

💡 Motivation

Machine learning models in real-world clinical settings must detect and reason about anomalies they have never seen during training. Current benchmarks mostly focus on known, curated categories—collapsing evaluation back into a closed-set problem and overstating model robustness.

NOVA is the first benchmark designed as a zero-shot, evaluation-only setting for assessing how well models:

Detect rare, real-world anomalies
Generalize across diverse MRI protocols and acquisition settings
Perform multimodal reasoning from image, text, and clinical context

It challenges foundation models and vision-language systems with what they were not trained for: the unexpected.

🧠 Dataset Overview

906 brain MRI slices
281 rare neurological conditions, spanning neoplastic, vascular, metabolic, congenital, and other pathologies
Real-world clinical heterogeneity (unprocessed, long-tailed distribution)
Radiologist-written captions and double-blinded bounding boxes
Clinical histories and diagnostics for reasoning tasks

🛠️ All cases are 2D PNG slices, sized 480×480, and are available under CC BY-NC-SA 4.0.

📊 Benchmark Tasks

NOVA captures the clinical diagnostic workflow through three open-world tasks:

🔍 1. Anomaly Localization

Detect abnormal regions via bounding box prediction. Evaluated with:

mAP@30, mAP@50, mAP@[50:95]
True/false positive counts per case

📝 2. Image Captioning

Generate structured radiology-style descriptions.

Evaluated with Clinical/Modality F1, BLEU, METEOR
Also assesses normal/abnormal classification

🧩 3. Diagnostic Reasoning

Predict the correct diagnosis based on clinical history + image caption.

Evaluated with Top-1, Top-5 accuracy
Label coverage and entropy analysis for long-tail reasoning

🧪 Model Performance (Stress Test)

Task	Top Model (2025)	Top Metric	Score
Anomaly Localization	Qwen2.5-VL-72B	mAP@50	24.5%
Image Captioning	Gemini 2.0 Flash	Clinical Term F1	19.8%
Diagnostic Reasoning	GPT-4o	Top-1 Accuracy	24.2%

Even top-tier foundation models fail under this open-world generalization benchmark.

⚠️ Intended Use

NOVA is intended strictly for evaluation. Each case has a unique diagnosis, preventing leakage and forcing true zero-shot testing.

Do not fine-tune on this dataset.
Ideal for:

Vision-language model benchmarking
Zero-shot anomaly detection
Rare disease generalization

📬 Contact

Stay tuned for the public leaderboard coming soon.