MM-ODIR-129: Multi-Modal Optic Disc Image Dataset

MM-ODIR-129 is a curated fundus photography dataset of 129 optic-disc-centered retinal images, designed for glaucoma detection, optic-disc/cup segmentation, and multimodal research that combines medical imaging with clinical textual descriptions. Each image is annotated by ophthalmologists with a binary diagnosis (Normal / Pathological) and, when pathological, graded for seven glaucomatous signs. A subset of 69 pathological images includes pixel-level optic cup and optic disc masks.

Disclaimer: This dataset is intended for research and educational purposes only. It is not a medical device and must not be used for clinical diagnosis or treatment decisions without appropriate validation and regulatory approval.

---

Dataset Overview

Property	Value
Total images	129
Image modality	Color fundus photography (optic disc centered)
Image size	Variable resolution (RGB); masks match each image's dimensions
Masks	Binary optic-cup and optic-disc masks (PNG + NumPy .npy)
Diagnosis labels	Normal, Pathological
Primary condition	Glaucoma (for pathological cases)
Splits	train: 77, validation: 26, test: 26

---

Split Distribution

Split	Total	Normal	Pathological
train	77	36	41
validation	26	12	14
test	26	12	14
Total	129	60	69

---

File Structure

MM-ODIR-129/
├── annotations.json      # Full clinical annotations for all 129 images
├── split.json            # Train / validation / test partition
├── train/                # Training split files
│   ├── <id>_<side>.jpg           # Fundus image
│   ├── <id>_<side>_cup.png       # Optic cup mask (PNG)
│   ├── <id>_<side>_cup.npy       # Optic cup mask (NumPy)
│   ├── <id>_<side>_disc.png      # Optic disc mask (PNG)
│   └── <id>_<side>_disc.npy      # Optic disc mask (NumPy)
├── validation/           # Validation split files (same naming)
└── test/                 # Test split files (same naming)

Naming Convention

Files are named as {patient_id}_{eye}.jpg, where {eye} is left or right. Masks append _cup or _disc before the extension.

---

Annotations (annotations.json)

Each entry is a JSON object with the following fields:

Field	Description
id	Internal annotation identifier
image_filename	Filename of the fundus image
label	Binary label: Normal or Pathological
transcription	Free-text clinical description written by the annotating ophthalmologist
locs_data	Structured clinical findings (see below)

locs_data.glaucoma Grading

For pathological cases, seven structural signs of glaucoma are scored. Each sign is a single-select ordinal variable. Higher scores indicate greater abnormality or severity. The definitions below are the ones used during annotation.

1. Cup-to-Disc Ratio (CDR) — cup_to_disc_ratio

Score	Definition
0	≤ 0.3 (normal)
1	0.4 – 0.5 (borderline)
2	0.6 – 0.7 (suspicious)
3	0.8 – 0.9 (advanced cupping)
4	1.0 (total cupping)

2. Neuroretinal Rim — neuroretinal_rim

Score	Definition
0	Normal (ISNT rule preserved: Inferior ≥ Superior ≥ Nasal ≥ Temporal)
1	ISNT rule violation (thinning not following normal pattern)
2	Focal notching (localized rim loss)
3	Diffuse thinning
4	Near-total or total rim loss

3. Disc Hemorrhage (Drance Hemorrhage) — disc_hemorrhage

Score	Definition
0	None
1	Present (splinter hemorrhage at or near disc margin)

4. Peripapillary Atrophy (PPA) — peripapillary_atrophy

Score	Definition
0	None
1	Beta-zone PPA only (atrophy with visible sclera/choroidal vessels)
2	Large or progressive beta-zone PPA

5. Retinal Nerve Fiber Layer (RNFL) Defect — rnfl_defect

Score	Definition
0	No visible RNFL defect
1	Wedge-shaped RNFL defect (localized)
2	Diffuse RNFL thinning
3	Both wedge defect and diffuse thinning

6. Optic Disc Pallor — disc_pallor

Score	Definition
0	Normal color
1	Mild pallor
2	Significant pallor (consider other optic neuropathies)

7. Vessel Changes — vessel_changes

Score	Definition
0	Normal vessel pattern
1	Bayoneting (vessels deflect at rim edge)
2	Nasalization of vessels
3	Both bayoneting and nasalization

Clinical note: Glaucoma diagnosis from fundus photography alone has limitations. IOP measurement, visual field testing (perimetry), and OCT of the RNFL/ganglion cell layer are typically required for definitive diagnosis and staging. These labels capture structural findings visible on fundus images only.

For Normal cases, locs_data is typically empty ({}), although a few records may contain glaucoma fields with all-zero or low scores.

---

Masks

• PNG masks: 8-bit grayscale (L mode), same spatial dimensions as the corresponding fundus image. Foreground pixels (255) denote the region of interest; background pixels (0) denote non-region.
• NumPy masks: Arrays of uint8 values {0, 1} with the same height and width as the corresponding fundus image.
• Masks are provided only for pathological images within each split.

Split	Images with masks
train	41
validation	14
test	14

---

Use Cases

This dataset supports several research tasks:

1. Binary classification: Predict Normal vs. Pathological from fundus images.
2. Multi-label grading: Predict the seven glaucoma severity scores.
3. Optic disc / cup segmentation: Segment the optic disc and cup from fundus images.
4. Multimodal learning: Combine fundus images with clinical transcriptions for vision-language modeling.
5. Report generation: Generate clinical descriptions from retinal images.

---

Loading the Data with 🤗 Datasets

If you upload this folder as a Hugging Face dataset repository, you can load it with:

from datasets import load_dataset

dataset = load_dataset("TheBug95/MM-ODIR-129")

For a custom loader that reads the JSON annotations and masks together, a minimal example is:

import json
from pathlib import Path
from PIL import Image
import numpy as np

base = Path("MM-ODIR-129")

with open(base / "annotations.json") as f:
    annotations = {a["image_filename"]: a for a in json.load(f)}

with open(base / "split.json") as f:
    splits = json.load(f)

for split, items in splits.items():
    for ann in items:
        img_name = ann["image_filename"]
        stem = Path(img_name).stem
        img = Image.open(base / split / img_name)

        cup_path = base / split / f"{stem}_cup.png"
        disc_path = base / split / f"{stem}_disc.png"
        cup_mask = Image.open(cup_path) if cup_path.exists() else None
        disc_mask = Image.open(disc_path) if disc_path.exists() else None

        label = ann["label"]
        transcription = ann.get("transcription", "")
        glaucoma_scores = ann.get("locs_data", {}).get("glaucoma", {})

---

Citation

A formal citation for this dataset is not yet available. If you use MM-ODIR-129 in your research, please reference the dataset URL until a publication or BibTeX entry is released:

https://huggingface.co/datasets/TheBug95/MM-ODIR-129

---

Acknowledgments

The clinical annotations were provided by ophthalmologists from Costa Rica and include detailed free-text transcriptions describing the optic nerve head, neuroretinal rim, cup-to-disc ratio, retinal nerve fiber layer defects, and vascular changes.

---

License

This dataset is released under the MIT License.

---

Limitations

• The dataset is relatively small (129 images) and is best suited for proof-of-concept research, few-shot learning, or as a fine-tuning set for larger pre-trained models.
• All pathological cases in this release are annotated as glaucoma; other optic neuropathies are not represented.
• Patient metadata is not provided, so demographic analyses are not possible.