EyeAssist: Radiologist Eye-Tracking Datasets and Evaluation Protocols

EyeAssist is a collection of two multimodal medical imaging datasets paired with radiologist eye-tracking (gaze) data, together with reference evaluation code that demonstrates how the datasets are used in benchmarking studies.

The release contains two datasets:

• EyeAssist-PE — chest CT volumes for lung-cancer prognosis, paired with gaze recordings from 7 radiologists across 2 reading sessions.
• EyeAssist-Neo — neonatal chest X-rays paired with gaze recordings under multiple experimental conditions (expert vs. generalist; with vs. without clinical context).

Both are designed to support research on gaze-guided deep learning, explainability in medical AI, and modeling of radiologist visual behavior.

Repository layout

EyeAssist/
├── Dataset/
│   ├── EyeAssist-PE/                  # chest CT + gaze (~4.9 GB)
│   │   ├── CT/                        # 40 CT volumes (NIfTI)
│   │   ├── Gaze/                      # gaze recordings, 7 readers × 2 sessions
│   │   ├── Saliency/                  # gaze-derived saliency maps + figures
│   │   ├── Clinical context.docx
│   │   ├── Expert Prognosis Decision.xlsx
│   │   └── README.md
│   └── EyeAssist-Neo/                 # neonatal X-ray + gaze (~462 MB)
│       ├── Xrays/                     # X-ray images (JPEG)
│       ├── Gaze&Saliency/             # gaze recordings under different conditions
│       └── clinical context.csv
└── Evaluation Protocol Code/
    ├── protocol1/                     # tabular feature baselines (per dataset)
    ├── protocol2/                     # deep saliency / transfer experiments
    └── protocol3/                     # gaze-weighted feature pooling (PE)

Each subdirectory contains its own README with dataset-specific schema and column definitions.

Datasets at a glance

	EyeAssist-PE	EyeAssist-Neo
Modality	Chest CT (NIfTI)	Chest X-ray (JPEG)
Cases	40 (20 survival / 20 death; 20 central / 20 peripheral)	100+
Readers	7 radiologists (R1–R7)	Experts, generalists, residents
Sessions	2 reading sessions	Session 2 with multiple conditions
Conditions	Blind / Context	Expert vs Generalist; With vs Without Clinical Context
Gaze format	per-frame CSV (Trial.csv)	per-frame CSV (fixations.csv)
Labels	Survival outcome + expert prognosis	Diagnosis, gestational age, clinical context

Quick start

import nibabel as nib
import pandas as pd

# EyeAssist-PE: load a CT volume + radiologist gaze
ct = nib.load("Dataset/EyeAssist-PE/CT/ca_42_1_diecentral.nii.gz").get_fdata()
gaze = pd.read_csv("Dataset/EyeAssist-PE/Gaze/Session 1/R1/Trial.csv")

# EyeAssist-Neo: load X-ray fixations
fix = pd.read_csv(
    "Dataset/EyeAssist-Neo/Gaze&Saliency/Session2 expert vs generalist/Expert/expert1/csv/fixations.csv"
)

Evaluation Protocols

Three reference protocols are included under Evaluation Protocol Code/:

• Protocol 1 — tabular gaze-feature baselines: extracts summary gaze statistics (fixation count, dwell time, entropy, ROI revisit rate, etc.) and trains classical classifiers. Runs on both datasets.
• Protocol 2 — deep learning experiments: saliency-map prediction with the EML-NET backbone (vendored under protocol2/EyeAssist-PE/models/EML-NET-Saliency/), and transfer-learning experiments on the Neo dataset.
• Protocol 3 — gaze-weighted feature pooling for PE: extracts CT features with U-Net / nnU-Net / SwinUNETR backbones and pools them by per-reader gaze density, comparing no_gaze / blind / context conditions.

Each protocol directory has its own README with run instructions and config defaults.

Setup

pip install numpy pandas nibabel scipy scikit-learn scikit-image matplotlib torch torchvision pillow monai

Backbone-specific models (Models Genesis pretrained weights, MONAI SwinUNETR weights) are not bundled — see Evaluation Protocol Code/protocol3/main.py for the expected paths and download links.

File formats

Type	Format	Tools
CT volumes	.nii, .nii.gz	nibabel, SimpleITK
X-ray images	.jpg, .jpeg	PIL, cv2
Gaze data	.csv	pandas
Clinical context	.csv, .docx	pandas, python-docx
Prognosis labels	.xlsx	pandas (openpyxl)
Saliency maps	.png, .npy	PIL, numpy
Pretrained weights	.pt, .pth	torch

Reader anonymization

All radiologists, experts, and readers are referred to by anonymous identifiers (e.g. R1–R7, expert1–expert5, generalist1–generalist5, reader1–reader3). No personally identifying information is included.

License

Released under CC-BY-4.0. See LICENSE for details.

Acknowledgements

The Protocol 2 saliency experiments use a vendored copy of the EML-NET-Saliency codebase by Sen Jia and Neil D. B. Bruce; see the upstream repository for license and citation.