ERDES: A Benchmark Video Dataset for Retinal Detachment and Macular Status Classification in Ocular Ultrasound

This paper introduces ERDES, the first open-access benchmark dataset of ocular ultrasound videos labeled for retinal detachment presence and macular status, alongside baseline performance evaluations of 40 deep learning models across eight architectures to advance automated diagnosis in resource-limited settings.

The Gist

Imagine your eye is like a high-tech camera. Inside this camera, there's a delicate film called the retina. If this film peels away from the back of the camera, it's called a Retinal Detachment (RD). This is a medical emergency because, just like a camera film, if it stays peeled too long, the picture (your vision) can be ruined forever.

Here is the tricky part: Not all peels are the same.

• The "Good" Emergency: The film is peeling, but the most important part (the center, called the macula) is still stuck down. If doctors fix this immediately (within 24 hours), the patient usually keeps their perfect vision.
• The "Bad" Emergency: The film has peeled all the way to the center. Even if doctors fix it, the vision might never fully come back.

Doctors also have to tell the difference between a real emergency (RD) and a harmless annoyance called Posterior Vitreous Detachment (PVD). Think of PVD like a piece of dust floating in the camera lens. It looks a bit like the film peeling, but it's actually just a harmless bubble that moves around. Telling the difference is hard, even for experts.

The Problem: The "Expert" Bottleneck

To find these problems, doctors use a special ultrasound camera (POCUS). It's like a flashlight that lets them see inside the eye without cutting it open. However, looking at these ultrasound videos is like trying to read a blurry, moving map. You need a highly trained expert to spot the difference between a "dust bubble" (PVD) and a "peeling film" (RD), and then to check if the center is safe.

In many places, especially busy emergency rooms or remote areas, there aren't enough experts to look at every single video. This means some emergencies get missed, or people get sent to the wrong specialists.

The Solution: ERDES (The "Training Gym" for AI)

This paper introduces ERDES (Eye Retinal DEtachment ultraSound). Think of ERDES as the first-ever "training gym" for artificial intelligence (AI) to learn how to be an eye doctor.

Before this, AI had no good practice material. It was like trying to teach a student to drive without ever letting them sit in a car. Now, the researchers have released a massive library of 5,381 real eye ultrasound videos (about 5 hours of footage) that are perfectly labeled.

What makes ERDES special?

1. It knows the difference: The videos are tagged not just with "Is there a problem?" but also "Is the center (macula) safe or peeled?" This is crucial because it tells the AI how urgent the situation is.
2. It's open source: Just like a public library, anyone can borrow these videos to build their own AI doctors.
3. It's realistic: The videos come from real patients in emergency rooms, including the messy parts like moving eyes and different machine settings, so the AI learns to handle real-world chaos.

How the AI Learned (The "Coach" and the "Student")

The researchers didn't just dump the videos on the AI; they gave it a structured workout plan:

• Step 1: The Filter (Stage 1): They trained an AI to act like a security guard. Its only job is to look at a video and say, "Is there a peeling film (RD) or is everything normal?"
• Step 2: The Triage Nurse (Stage 2): If the security guard says "Yes, there's a problem," the video gets passed to a second AI. This one acts like a triage nurse. It looks only at the peeling films and asks, "Is the center safe (Macula-Intact) or is it gone (Macula-Detached)?"

The Result:
They tested 40 different "student" AI models (using different brain architectures like 3D ResNet and Transformers). The best ones learned to spot the peeling film with 94% accuracy and could tell if the center was safe with 88% accuracy.

Why This Matters

Think of this like giving every emergency room a super-powered assistant.

• Speed: The AI can look at a video in seconds, flagging the dangerous cases for the human doctor to review immediately.
• Safety: It helps ensure that patients with "Macula-Intact" detachments get surgery today to save their sight, rather than waiting days.
• Accessibility: Even in a small clinic without a specialist on staff, a doctor with a basic ultrasound machine could use this AI tool to know exactly how urgent the situation is.

In short, ERDES is the foundation for a future where AI helps doctors save vision faster, ensuring that the right patients get the right help at the right time.

Technical Summary

Here is a detailed technical summary of the paper "ERDES: A Benchmark Video Dataset for Retinal Detachment and Macular Status Classification in Ocular Ultrasound."

1. Problem Statement

Retinal Detachment (RD) is a vision-threatening ocular emergency requiring immediate surgical intervention. The urgency and visual prognosis depend critically on macular status: whether the macula (the central retina responsible for high-acuity vision) is intact or detached.

• Clinical Challenge: Point-of-care ultrasound (POCUS) is a rapid, non-invasive tool for detecting RD, but its utility is limited by the need for expert interpretation. Distinguishing RD from benign conditions like Posterior Vitreous Detachment (PVD) and determining macular involvement are difficult tasks, especially in resource-limited settings.
• Research Gap: While deep learning has been applied to medical imaging, existing models for RD detection generally ignore macular status. Furthermore, no publicly available dataset exists that supports video-based classification of RD with specific labels for macular involvement. Existing datasets are often image-based (static frames) or lack macular annotations.

2. Methodology

A. Dataset Creation (ERDES)

The authors introduced ERDES (Eye Retinal DEtachment ultraSound), the first open-access dataset of ocular ultrasound video clips.

• Data Source: 5,381 video clips collected from 2010–2022 at the University of Arizona Emergency Department.
• Demographics: Diverse adult population (various races/ethnicities).
• Acquisition: High-frequency linear probes (5–12 MHz) using various devices (Mindray, Philips, GE, Sonosite).
• Labeling Workflow:
  • Adjudication: Three clinical experts independently reviewed clips, with a fourth expert resolving discrepancies.
  • Classes:
    1. Non-RD: Normal eyes and Posterior Vitreous Detachment (PVD).
    2. RD: Subdivided into Macula-Intact and Macula-Detached.
  • Granularity: Labels are assigned at the clip level (entire video) rather than frame-level to reflect clinical interpretation and account for temporal variations (eye motion, probe movement).
• Preprocessing:
  • De-identification: Automated removal of Protected Health Information (PHI) and device text overlays.
  • ROI Extraction: A YOLOv8 model was trained to detect the ocular globe and crop the Region of Interest (ROI), removing peripheral noise.
  • Standardization: Clips were converted to grayscale, zero-padded to square, and resized to 96×128×128 (Depth×Height×Width) via trilinear interpolation.

B. Benchmarking Framework

The authors trained and evaluated 40 deep learning models across 8 spatiotemporal architectures using the MONAI framework.

• Architectures: 3D ResNet, 3D U-Net, V-Net, UNet++, SENet, Swin-UNETR, UNETR, and ViT-based classifiers.
• Tasks: Five binary classification tasks:
  1. Non-RD vs. RD (Primary)
  2. Macula-Detached vs. Macula-Intact (Primary)
  3. Normal vs. PVD
  4. Normal vs. RD
  5. PVD vs. RD
• Training Strategy: Models were trained from scratch (no external pre-training) on stratified splits (72% train, 8% val, 20% test) to reflect real-world class imbalance.
• Pooling Strategy: Standard 3D global average pooling was used. For the challenging Normal vs. PVD task (where PVD is subtle and mobile), a selective temporal pooling strategy was tested, retaining only the top $k$ most informative temporal segments.

C. Proposed Diagnostic Pipeline

A two-stage cascaded classification system mimicking clinical workflow:

1. Stage 1: Detects presence of RD (Non-RD vs. RD).
2. Stage 2: If RD is detected, classifies macular status (Macula-Intact vs. Macula-Detached).

• Model Selection: 3D ResNet for Stage 1 and 3D U-Net for Stage 2 were selected based on superior sensitivity in benchmarks.

3. Key Contributions

1. ERDES Dataset: The first open-access (CC-BY 4.0) dataset of ocular ultrasound video clips labeled for both RD presence and macular status.
2. Comprehensive Benchmarking: Extensive evaluation of 40 models across 8 architectures, establishing baseline performance metrics (Accuracy, Precision, Sensitivity, Specificity, F1-score) for five distinct diagnostic tasks.
3. Two-Stage Pipeline: A clinically aligned cascaded approach that separates RD detection from macular status assessment, optimizing for the specific urgency of macula-intact cases.
4. Technical Innovations:
   • Automated ROI localization using YOLOv8 to handle variable probe positions.
   • Investigation of selective temporal pooling to improve detection of subtle, mobile pathologies (PVD) that appear in fewer frames.
5. Open Source: Public release of the dataset, pre-trained models, and training scripts on Zenodo and GitHub.

4. Key Results

• RD Detection (Non-RD vs. RD):
  • The 3D U-Net achieved the best performance with 99.1% Accuracy, 95.0% Sensitivity, and 99.6% Specificity.
  • The 3D ResNet also performed strongly (97.7% Accuracy, 93.9% Sensitivity).
• Macular Status (Macula-Detached vs. Intact):
  • 3D U-Net again led with 88.2% Accuracy, 89.9% Sensitivity, and 87.0% Specificity.
  • 3D ResNet followed closely with 86.2% Accuracy.
• Challenging Tasks:
  • Normal vs. PVD: This was the most difficult task due to the subtle nature of PVD. Global pooling yielded lower sensitivity (~45-63%). Selective pooling improved performance for most architectures (e.g., 3D U-Net sensitivity increased from 78.4% to 80.7% with $r=0.3$).
• End-to-End Pipeline Performance:
  • The cascaded system achieved an end-to-end sensitivity of 84.4% for detecting Macula-Intact RD and 81.6% for Macula-Detached RD.
  • Error propagation was minimal due to the high sensitivity and specificity of the Stage 1 model.

5. Significance

• Clinical Impact: By automating the distinction between macula-intact and macula-detached RD, ERDES enables AI systems to prioritize surgical cases. Macula-intact cases require same-day surgery to prevent permanent vision loss, whereas macula-detached cases have a different urgency profile.
• Resource Accessibility: The dataset and models aim to democratize RD diagnosis, allowing non-experts in resource-limited settings to utilize POCUS with AI assistance.
• Research Foundation: ERDES fills a critical gap in medical AI by providing a standardized, video-based benchmark for spatiotemporal analysis of ocular ultrasound, moving beyond static image analysis.
• Reproducibility: The open release of data, code, and pre-trained models accelerates further research in ophthalmic ultrasound interpretation.