[Paper Review] OPTED: Open Preprocessed Trachoma Eye Dataset Using Zero-Shot SAM 3 Segmentation

Kibrom Gebremedhin, Hadush Hailu+1 more|arXiv (Cornell University)|Mar 6, 2026

Reproductive tract infections research0 citations

TL;DR

OPTED presents an open-source preprocessed trachoma eye dataset derived with SAM 3 zero-shot segmentation, detailing a four-step pipeline and prompt evaluation to produce two output formats for ML ready images.

ABSTRACT

Trachoma remains the leading infectious cause of blindness worldwide, with Sub-Saharan Africa bearing over 85% of the global burden and Ethiopia alone accounting for more than half of all cases. Yet publicly available preprocessed datasets for automated trachoma classification are scarce, and none originate from the most affected region. Raw clinical photographs of eyelids contain significant background noise that hinders direct use in machine learning pipelines. We present OPTED, an open-source preprocessed trachoma eye dataset constructed using the Segment Anything Model 3 (SAM 3) for automated region-of-interest extraction. We describe a reproducible four-step pipeline: (1) text-prompt-based zero-shot segmentation of the tarsal conjunctiva using SAM 3, (2) background removal and bounding-box cropping with alignment, (3) quality filtering based on confidence scores, and (4) Lanczos resizing to 224x224 pixels. A separate prompt-selection stage identifies the optimal text prompt, and manual quality assurance verifies outputs. Through comparison of five candidate prompts on all 2,832 known-label images, we identify "inner surface of eyelid with red tissue" as optimal, achieving a mean confidence of 0.872 (std 0.070) and 99.5% detection rate (the remaining 13 images are recovered via fallback prompts). The pipeline produces outputs in two formats: cropped and aligned images preserving the original aspect ratio, and standardized 224x224 images ready for pre-trained architectures. The OPTED dataset, preprocessing code, and all experimental artifacts are released as open source to facilitate reproducible trachoma classification research.

Motivation & Objective

• Provide a publicly available preprocessed trachoma image dataset originating from Sub-Saharan Africa.
• Develop a reproducible four-step preprocessing pipeline using SAM 3 for ROI extraction in eyelid images.
• Systematically evaluate text prompts to identify the optimal SAM 3 prompt for tarsal conjunctiva segmentation.
• Release dataset, code, and artifacts to enable reproducible trachoma classification research.

Proposed method

• Use SAM 3 in a zero-shot setting to segment the tarsal conjunctiva from eyelid photographs using five candidate text prompts.
• Apply background removal, bounding-box cropping with 5% padding, and horizontal alignment to obtain cropped images.
• Resize images to 224x224 pixels using Lanczos interpolation for compatibility with standard architectures.
• Perform a prompt-selection stage to pick the optimal prompt based on detection rate, mean confidence, and mask coverage, with a fallback mechanism for misses.
• Provide outputs in two formats: cropped/aligned images and standardized 224x224 images, along with full open-source preprocessing code.

Figure 2: Overview of the OPTED preprocessing pipeline. Raw eyelid photographs are processed through SAM 3 text-prompt segmentation, background removal, bounding-box cropping with 5% padding, horizontal alignment, and Lanczos resizing to $224\times 224$ pixels.

Experimental results

Research questions

• RQ1 Which SAM 3 text prompt yields the best segmentation performance for the tarsal conjunctiva in trachoma eyelid images?
• RQ2 Can a fully reproducible four-step preprocessing pipeline produce reliable 224x224 ML-ready images from raw trachoma photos?
• RQ3 What is the impact of prompt choice on detection rate, confidence, and mask coverage across a multi-country dataset?
• RQ4 How well does the open-source OPTED pipeline generalize to a diverse set of trachoma images from Sub-Saharan Africa and other regions?

Key findings

• The prompt “inner surface of eyelid with red tissue” achieves the highest mean confidence (0.873) and largest mask area (29.8%).
• Overall mean SAM 3 confidence across the dataset is 0.872 with std 0.070.
• Detection rate is 99.5% with 13 misses recovered via fallback prompts.
• The OPTED pipeline produces 2 formats: cropped/aligned and 224x224 PNG images, with a stratified train/validation/test split (70/15/15).
• Lanczos resizing yields the best PSNR (39.16 dB) and SSIM (0.9713) for 224x224 outputs.

Figure 3: Visual comparison of SAM 3 masks from the five candidate prompts on three sample images (two Normal, one Trachoma). Blue overlay indicates the predicted mask; yellow contours delineate boundaries. The selected prompt P5 (green border) provides the most complete coverage of the tarsal conju