OPTED: Open Preprocessed Trachoma Eye Dataset Using Zero-Shot SAM 3 Segmentation

This paper introduces OPTED, an open-source preprocessed trachoma eye dataset derived from 2,832 images using a zero-shot SAM 3 pipeline to automatically extract and standardize regions of interest, thereby addressing the scarcity of high-quality data for automated trachoma classification in Sub-Saharan Africa.

The Gist

Imagine you are trying to teach a robot to spot a specific type of weed in a giant, messy garden. The problem is, the photos you have of the garden are taken from far away, and they include the gardener's gloved hands, the sky, dirt, and random leaves. If you feed these messy photos directly to the robot, it gets confused and can't learn the difference between the weed and the rest of the garden.

This paper is about building a super-clean, organized garden specifically for a robot to learn how to spot Trachoma, a disease that causes blindness.

Here is the story of how they did it, broken down into simple steps:

1. The Problem: The "Messy Garden"

Trachoma is a big problem, especially in places like Ethiopia and Sub-Saharan Africa. It's the leading cause of preventable blindness. Doctors usually look at the inside of a patient's eyelid to diagnose it. But taking photos of this is tricky.

• The Mess: The photos often show gloved fingers, skin, and bad lighting.
• The Goal: We need to teach computers to look only at the pink, fleshy part inside the eyelid (the "tarsal conjunctiva") to see if it's inflamed or has little bumps (follicles).
• The Gap: Until now, there was no clean, ready-to-use collection of these specific photos for computer scientists to study.

2. The Solution: The "Smart Robot Butler" (SAM 3)

The authors created a new dataset called OPTED. To make it, they didn't hire a team of humans to cut out the eyelids from thousands of photos (that would take forever!). Instead, they used a "Smart Robot Butler" called SAM 3 (Segment Anything Model 3).

Think of SAM 3 as a very smart assistant who has seen a billion pictures of the world. You don't need to teach it what an eyelid is; you just have to ask it nicely in plain English.

3. The Secret Recipe: Finding the Right "Magic Words"

The researchers realized that if they told the robot, "Find the conjunctiva," it would get confused because the robot doesn't speak "medical jargon." It speaks "visual descriptions."

So, they ran a test with five different "magic phrases" to see which one made the robot do the best job:

• Phrase A: "Red tissue inside eye"
• Phrase B: "Membrane under eyelid"
• Phrase C: "Inner surface of eyelid with red tissue" (The Winner!)

It turns out, the robot understood the visual description best. When they used the winning phrase, the robot successfully found the eyelid in 99.5% of the photos. For the few it missed, they had a backup plan to try the other phrases.

4. The Assembly Line: The 4-Step Pipeline

Once the robot found the eyelid, the paper describes a four-step assembly line to turn a messy photo into a perfect, standardized image:

1. The Cut (Segmentation): The robot draws a digital outline around just the eyelid tissue.
2. The Cleanup (Background Removal): Everything outside that outline (fingers, sky, shadows) is painted black and thrown away.
3. The Straightening (Alignment): If the photo is sideways, the robot rotates it so the eyelid is always horizontal, like a book on a shelf.
4. The Resize (Lanczos Interpolation): The image is shrunk down to a perfect square (224x224 pixels). They used a special resizing technique (Lanczos) that is like using a high-quality photo editor to shrink the image without making it look blurry or pixelated. This ensures the fine details (like tiny bumps on the eyelid) stay sharp.

5. The Result: A Ready-to-Use Library

The final product is OPTED:

• 2,832 Clean Photos: All processed and ready for computers to study.
• Three Categories: The photos are labeled as Normal (healthy), TF (mild inflammation), or TI (severe inflammation).
• Open Source: The best part? The authors gave away the "recipe" (the code) and the "ingredients" (the photos) for free.

Why Does This Matter?

Imagine trying to bake a cake but you have to wash the flour, sift it, and measure it yourself every single time before you can even start. That's what researchers were doing before. Now, with OPTED, they have a pre-measured, pre-sifted bag of flour.

This allows scientists to focus on building better "bakers" (AI models) to diagnose eye diseases faster and more accurately. Since Trachoma is a huge problem in Africa, and this dataset comes from real field studies there, it's a massive step toward the goal of eliminating blindness by 2030.

In short: They built a digital factory that uses a smart AI to clean up messy eye photos, turning them into a perfect, standardized library so computers can learn to save people's sight.

Technical Summary

Here is a detailed technical summary of the paper "OPTED: Open Preprocessed Trachoma Eye Dataset Using Zero-Shot SAM 3 Segmentation."

1. Problem Statement

Trachoma, caused by Chlamydia trachomatis, is the leading infectious cause of blindness globally, with Sub-Saharan Africa (particularly Ethiopia) bearing over 85% of the global burden. While deep learning offers promise for automated screening, progress is hindered by two critical gaps:

• Data Scarcity: There are no publicly available, preprocessed trachoma datasets originating from Sub-Saharan Africa. Existing datasets are either proprietary, small, or lack standardized preprocessing.
• Preprocessing Challenges: Raw clinical eyelid photographs contain significant background noise (gloved fingers, skin, variable lighting) that hinders direct machine learning application. Previous methods relied on manual cropping or less robust automated techniques, and no reproducible pipeline for Region of Interest (ROI) extraction has been released.

2. Methodology: The OPTED Pipeline

The authors present OPTED, an open-source dataset and a fully reproducible four-step preprocessing pipeline built on SAM 3 (Segment Anything Model 3), a foundation model capable of zero-shot segmentation via text prompts.

A. Source Data

• Origin: Aggregated from 7 field studies across 6 countries (including Ethiopia, Tanzania, and The Gambia) provided by Tropical Data.
• Volume: 2,963 raw images; 2,832 were selected based on having known labels (Normal, TF, TI) assigned by WHO-certified graders.
• Classes:
  • Normal: Healthy eyes.
  • TF: Trachomatous inflammation, follicular.
  • TI: Trachomatous inflammation, intense.

B. The Four-Step Pipeline

1. Zero-Shot Text-Prompt Segmentation:
   • SAM 3 is applied without fine-tuning.
   • Prompt Selection: Five candidate text prompts were evaluated. The optimal prompt identified was "inner surface of eyelid with red tissue."
   • Mechanism: SAM 3 generates candidate masks with confidence scores. The highest-scoring mask is selected. A confidence threshold of 0.3 is used for quality control.
2. Background Removal and Cropping:
   • Pixels outside the mask are set to black (RGB 0,0,0) to maximize contrast.
   • An axis-aligned bounding box is computed around the mask.
   • The image is cropped with 5% padding on all sides.
3. Alignment:
   • Images are rotated so the longest axis is horizontal (landscape orientation) to ensure consistent spatial layout.
4. Resizing:
   • Images are resized to 224 × 224 pixels using Lanczos interpolation.
   • Justification: Lanczos was selected over nearest neighbor, bilinear, and bicubic methods because it achieved the highest PSNR (39.16 dB) and SSIM (0.9713), best preserving fine tissue details like follicles and vascular patterns.

C. Prompt Evaluation Strategy

The authors systematically compared five prompts across the entire dataset. The winning prompt ("inner surface of eyelid with red tissue") achieved:

• Mean Confidence: 0.873 (highest).
• Consistency: Lowest standard deviation (0.069).
• Coverage: Largest mask area ratio (29.8%).
• Detection Rate: 99.5% (13 images missed were recovered via a fallback mechanism using secondary prompts).

3. Key Contributions

1. First Open Preprocessed Dataset from Sub-Saharan Africa: OPTED is the first publicly available, preprocessed trachoma dataset derived from the region with the highest disease burden.
2. Systematic Prompt Engineering: A rigorous evaluation of text prompts for medical segmentation, demonstrating that visually descriptive prompts ("red tissue") outperform clinical terminology ("everted conjunctiva") when using vision-language foundation models.
3. Reproducible Open-Source Pipeline: The complete code, preprocessing scripts, prompt comparison artifacts, and the final dataset are released under open licenses (CC-BY-NC 4.0 for data, MIT for code).
4. Dual-Format Output: The dataset provides images in two formats:
   • Cropped/aligned images preserving original aspect ratios.
   • Standardized 224 × 224 images ready for pre-trained architectures (ResNet, ViT, etc.).

4. Results and Statistics

• Dataset Size: 2,832 processed images.
• Class Distribution:
  • Normal: 2,487 images (87.8%).
  • TF: 324 images (11.4%).
  • TI: 21 images (0.7%).
  • Note: The dataset exhibits significant class imbalance, particularly for the TI class.
• Performance Metrics:
  • Overall Mean Confidence: 0.872 (Std 0.070).
  • Class-Specific Confidence: TI (0.902) > TF (0.888) > Normal (0.870).
  • Processing Speed: ~1.3 seconds per image on an NVIDIA RTX 2000 Ada GPU (full dataset processed in ~1 hour).
• Quality Assurance: All outputs underwent manual review. The 13 images missed by the primary prompt were successfully recovered using fallback prompts.

5. Significance and Future Impact

• Reproducibility: By releasing the full pipeline and artifacts, the authors enable the research community to benchmark classification models on a standardized, high-quality dataset, eliminating the "black box" of previous preprocessing methods.
• Clinical Relevance: The dataset supports the WHO's 2030 goal of eliminating trachoma by providing scalable, technology-driven screening tools.
• Methodological Insight: The finding that natural language visual descriptions outperform medical jargon in zero-shot segmentation for specific anatomical structures offers a preliminary guideline for other medical imaging tasks using foundation models.
• Future Work: The authors plan to validate segmentation quality using IoU/Dice metrics against expert annotations and address the class imbalance (specifically the scarcity of TI images) through community contributions and cross-validation strategies.

Availability: The dataset, code, and documentation are publicly available to facilitate reproducible trachoma classification research.