Dataset Card: GazeIntent = RadSeq & RadExplore & RadHybrid

Dataset Name: phamtrongthang/GazeIntent
Repository: UARK‑AICV/RadGazeIntent
License: CC BY-NC-SA 4.0

---

1. Dataset Summary

GazeIntent is the first intention-labeled eye-tracking dataset for radiological interpretation, capturing radiologist's diagnostic intentions during chest X-ray analysis. It includes:

• 3,562 chest X-ray samples with expert radiologist eye-tracking data
• Fine-grained intention labels for each fixation point
• Three distinct intention modeling paradigms representing different visual search behaviors
• Multi-label annotations for 13 radiological findings

This dataset supports research in intention interpretation, gaze-informed diagnosis, cognitive modeling, and explainable AI in medical imaging.

🏅 This work was accepted at ACM MM 2025 - A top-tier international conference on multimedia research.

---

2. Dataset Structure

Attribute	Description
Total Samples	3,562 chest X-rays
Sources	EGD (1,079) + REFLACX (2,483)
Modality	Chest X-ray images
Gaze Data	2D coordinates + fixation duration + intention labels
Intention Classes	13 radiological findings
Radiologists	Multiple expert radiologists

---

3. Three Intention Paradigms

RadSeq (Systematic Sequential Search)

• Models radiologists following a structured diagnostic checklist
• One finding examined at a time in sequential order
• Reflects systematic, methodical visual search patterns

RadExplore (Uncertainty-driven Exploration)

• Captures opportunistic visual search behavior
• Radiologists consider multiple findings simultaneously
• Represents exploratory, uncertainty-driven attention

RadHybrid (Hybrid Pattern)

• Combines initial broad scanning with focused examination
• Two-phase approach: overview → targeted search
• Reflects real-world diagnostic behavior patterns

---

4. Intended Uses

• Radiologist intention interpretation and prediction
• Gaze-informed medical diagnosis systems
• Cognitive modeling of expert visual reasoning
• Medical education and training assessment
• Explainable AI for radiology applications
• Human-AI collaboration in medical imaging

---

5. Tasks and Benchmarks

Primary Task: Fixation-based Intention Classification

• Baseline: RadGazeIntent (transformer-based architecture)
• Input: Fixation sequences + chest X-ray images
• Output: Intention confidence scores for 13 findings

Evaluation Metrics:

• Classification: Accuracy, F1-score, Precision, Recall
• Multi-label: Per-class and macro-averaged metrics

Findings Covered: Atelectasis, Cardiomegaly, Consolidation, Edema, Enlarged Cardiomediastinum, Fracture, Lung Lesion, Lung Opacity, Pleural Effusion, Pleural Other, Pneumonia, Pneumothorax, Support Devices

---

6. Data Availability

The processed intention-labeled datasets are publicly available via Hugging Face under CC BY-NC-SA 4.0 license.

Access Requirements: Users must agree to share contact information and accept the license terms to access the dataset files.

---

7. Technical Details

Data Processing: Three datasets derived from existing eye-tracking sources (EGD, REFLACX) using different intention modeling assumptions:

• Uncertainty Filtering: Assigns labels based on temporal alignment with radiologist transcripts
• Sequential Constraints: Applies GazeSearch methodology for systematic search modeling
• Hybrid Integration: Combines initial scanning phase with focused examination periods

---

8. Citation

Please cite this dataset using the following BibTeX entry:

@article{pham2025interpreting,
  title={Interpreting Radiologist's Intention from Eye Movements in Chest X-ray Diagnosis},
  author={Pham, Trong-Thang and Nguyen, Anh and Deng, Zhigang and Wu, Carol C and Nguyen, Hien and Le, Ngan},
  journal={arXiv preprint arXiv:2507.12461},
  year={2025}
}

---

9. Acknowledgments

This work is supported by:

• National Science Foundation (NSF) Award No OIA-1946391, NSF 2223793 EFRI BRAID
• National Institutes of Health (NIH) 1R01CA277739-01
• Built upon EGD and REFLACX eye-tracking datasets

Contact: Trong Thang Pham ([email protected])