📢 Domain & Email Migration Notice From May 30th, 2026, Fundusnap will transition to new domains as fundusnap.com will not be renewed: 🌐 Website: fundusnap.faizath.com (formerly fundusnap.com) ⚙️ API: fundusnap-api.faizath.com (formerly api.fundusnap.com) 📧 Email: contact@fundusnap.faizath.com (formerly contact@fundusnap.com) 🛰️ CDN: fundusnap-cdn.faizath.com (formerly cdn.fundusnap.com) 📈 Status Pages: https://status.faizath.com/status/fundusnap (formerly status.fundusnap.com)

AI-assisted diabetic retinopathy screening — showing where on the retina the findings are.

🤗 Hugging Face  •  🐙 GitHub

fundusnap-v1-lesiondet-yolo11m-20m

Retinal lesion detector for colour fundus (retinal) photographs. Given a single fundus image it returns axis-aligned bounding boxes for twelve findings — ten pathological (microaneurysms, haemorrhages, exudates, IRMA, atrophy, scars, capture artefacts) and two anatomical landmarks (optic disc, fovea) — each with a class label and a confidence score. The model is an Ultralytics YOLO11m detector fine-tuned from the COCO-pretrained yolo11m.pt checkpoint at 640×640 — roughly 20M parameters, hence the name: lesiondet (lesion detection) + yolo11m (YOLO11 medium) + 20m (parameter count).

It ships as a stripped Ultralytics best.pt checkpoint, ready for inference or further fine-tuning, together with the FastAPI service Fundusnap uses to serve it.

Its companion model is fundusnap-v1-severitycls-rn34-22m, which grades overall DR severity. This one answers the follow-up question: where does the grade come from?

⚠️ Intended use — not for clinical use

This model is released for research and engineering use, and at most as a visualisation and triage aid inside a workflow that a qualified clinician supervises.

• It is not a medical device and has no regulatory clearance (FDA, CE/MDR, or otherwise).
• It must never be the sole basis for a diagnosis, referral, or treatment decision.
• Inside Fundusnap it exists for explainability — it draws boxes over an image that has already been graded elsewhere. It is not a grader, and the number of boxes it returns is not a severity score.
• Detection quality is modest (mAP@50-95 ≈ 0.28, recall ≈ 0.53). Roughly half of the annotated findings in its own validation split are missed. An empty result means "nothing detected", never "nothing there".
• It has not been validated prospectively, on any specific camera or population, or against a reference annotation standard beyond the dataset described below.

Anyone deploying it in a screening context is responsible for their own validation and for keeping a human grader in the loop.

Label scheme

Twelve classes, in model.names order:

Index	Class	What it is
0	Artefact	Image capture artefact (dust, glare, reflection) — not retinal pathology
1	Atrophic Scar	Healed atrophic scarring of the retina/choroid
2	Atrophy	Retinal or chorioretinal atrophy
3	Disc	Anatomical landmark — the optic disc
4	Flame R-hemorrhage	Flame-shaped retinal haemorrhage (nerve fibre layer)
5	Fovea	Anatomical landmark — the fovea / macular centre
6	H-exudate	Hard exudate (lipid deposit)
7	IRMA	Intraretinal microvascular abnormality
8	Laser Scar	Photocoagulation scar from prior laser treatment
9	Microaneurysm	Microaneurysm — the earliest visible DR lesion
10	R-hemorrhage	Retinal haemorrhage (dot/blot)
11	S-exudate	Soft exudate / cotton-wool spot

Note that Disc, Fovea, and Artefact are not lesions. Disc and Fovea are normal anatomy present in essentially every gradable image, and Artefact marks image-quality problems. Any downstream logic that counts detections as evidence of disease must exclude these three.

Files

Path	What it is
models/fundus_artifacts.pt	The model. Ultralytics best.pt, stripped of optimiser/EMA state (~40 MB, stored via Git LFS).
main.py	FastAPI service that loads the checkpoint and exposes JSON and annotated-image endpoints.
Dockerfile	Container build for that service (python:3.10-slim, port 8000).
requirements.txt	Runtime dependencies for the service.

Usage

Ultralytics (recommended)

pip install ultralytics

from ultralytics import YOLO

model = YOLO("models/fundus_artifacts.pt")

# conf/iou default to 0.25 / 0.7 — lower conf if you would rather over-detect than miss lesions.
results = model("fundus.jpg", conf=0.25, iou=0.7, imgsz=640)

for r in results:
    for box in r.boxes:
        x1, y1, x2, y2 = map(int, box.xyxy[0])
        cls_id = int(box.cls[0])
        conf = float(box.conf[0])
        print(f"{model.names[cls_id]:20s} {conf:.2f}  [{x1}, {y1}, {x2}, {y2}]")

# Annotated image as a BGR numpy array (what /visualize/fundus-artifacts/ returns):
annotated = results[0].plot()

Images are letterboxed to 640×640 internally; coordinates come back in the original image's pixel space, so no rescaling is needed. Batch by passing a list of paths or an (N, H, W, 3) array.

Exporting

model.export(format="onnx", opset=14, dynamic=True)   # also: torchscript, tflite, coreml, engine

Fine-tuning

The checkpoint is stripped (epoch: -1, no optimiser or EMA state), so it resumes as a starting point, not as a paused run:

model = YOLO("models/fundus_artifacts.pt")
model.train(data="your_dataset/data.yaml", epochs=50, imgsz=640, batch=16)

Your data.yaml must keep the same twelve class names in the same order, or retrain the head.

Training

Data — an internal 12-class fundus detection set exported in Roboflow layout, referenced in the run as FUNDUS-3/data.yaml. The dataset itself is not published with this model, and its provenance was not recorded in the checkpoint — image counts, split sizes, source cameras, annotator count, and licensing terms are all unknown from the artefacts in this repository. Treat every number below as a description of that run, not as an evaluation on any known public benchmark.

The wall-clock time is informative about scale: 35 epochs completed in 258 seconds (≈7.4 s/epoch) at batch 16 on GPU, which implies a small training set — on the order of a few hundred to a couple of thousand images.

Setup (from the checkpoint's train_args)

Base model	yolo11m.pt, COCO-pretrained
Architecture	yolo11m.yaml, scale m (depth 0.5, width 1.0), nc=12, anchor-free Detect head
Input	640×640, letterboxed
Epochs	35 (patience=100, so no early stop)
Batch size	16 (nbs=64 nominal)
Optimizer	auto, lr0=0.01, lrf=0.01, momentum=0.937, weight_decay=0.0005
Warmup	3 epochs, warmup_momentum=0.8
Loss weights	box 7.5, cls 0.5, dfl 1.5
Augmentation	mosaic 1.0 (disabled for the last 10 epochs), fliplr=0.5, scale=0.5, translate=0.1, HSV (0.015/0.7/0.4), erasing=0.4, RandAugment
Not used	flipud, degrees, shear, perspective, mixup, copy_paste
Precision	AMP
Seed	0, deterministic=True
Ultralytics	8.3.165 (run object_detection_model_v1, 2025-07-13)

Run log (abridged — validation metrics per epoch)

Epoch	box_loss	cls_loss	dfl_loss	Precision	Recall	mAP@50	mAP@50-95
1	2.3819	3.0624	1.4673	0.2534	0.2641	0.2048	0.0955
5	1.9664	1.4739	1.1881	0.3955	0.4270	0.4079	0.1925
10	1.8575	1.3348	1.1310	0.4918	0.4582	0.4587	0.2325
15	1.7731	1.2301	1.1132	0.5443	0.4606	0.4961	0.2575
20	1.7162	1.1354	1.0861	0.5211	0.5088	0.5289	0.2639
27	—	—	—	0.5354	0.5252	0.5343	0.2819
30	1.5557	0.9660	1.0399	0.5279	0.5054	0.5132	0.2633
35	1.4925	0.8980	1.0153	0.5573	0.5154	0.5300	0.2744

Training losses fall steadily to the last epoch while validation mAP plateaus after ~epoch 20 — the run was starting to overfit. Epoch 27 is the best checkpoint by Ultralytics' fitness metric (0.3071), and that is the epoch shipped as models/fundus_artifacts.pt.

Evaluation

Measured by Ultralytics on the run's own validation split, at the best epoch (27):

Metric	Value
Precision	0.5354
Recall	0.5252
mAP@50	0.5343
mAP@50-95	0.2819
Fitness (0.1·mAP50 + 0.9·mAP50-95)	0.3071

Per-class metrics are not available. The run's results.csv, confusion matrix, and PR curves were not retained, and the checkpoint stores aggregates only. Since the twelve classes are very unevenly difficult — the optic disc is a large, high-contrast, always-present object, while a microaneurysm is a handful of pixels — the aggregate almost certainly hides a wide spread, with the landmark classes propping the average up and the small-lesion classes well below it. Do not read mAP@50 ≈ 0.53 as "roughly half-right on lesions".

These figures are self-reported — produced by the training run itself, not verified by Hugging Face — and are mirrored in the model-index metadata at the top of this card, which is why the Hub labels them as such.

Limitations and bias

• Absolute performance is modest. mAP@50-95 of 0.28 and recall of 0.53 are far below what a clinical detection tool would need. Roughly half the annotated findings are missed even in-distribution.
• Aggregate metrics are flattered by the easy classes. Two of twelve classes (Disc, Fovea) are large, high-contrast anatomy present in nearly every image, and detecting them is close to trivial. They are pooled into the same mAP as microaneurysms.
• Small lesions are the weak point, and they matter most. Microaneurysms and dot haemorrhages are a few pixels wide; at 640×640 with letterboxing, much of that detail is gone before the model sees it. Those are exactly the lesions that define early, referable DR.
• The run is small. 35 epochs in ~4 minutes of wall clock implies a small dataset; validation mAP had already plateaued while training loss kept falling, i.e. the model was overfitting rather than running out of schedule.
• Unknown data provenance. Dataset size, sources, cameras, populations, and annotation protocol were not recorded. No domain-shift analysis and no subgroup fairness audit is possible from what survives, and none has been done. Expect degradation on different hardware, fields of view, illumination, or demographics.
• Class definitions are non-standard. The label set mixes pathology, anatomy, and image-quality artefacts, and terms like Atrophy vs Atrophic Scar or R-hemorrhage vs Flame R-hemorrhage depend on the annotator's convention rather than a published grading standard.
• No ungradable option beyond Artefact. A blurred, over-exposed, or non-fundus image still yields plausible-looking boxes; the Artefact class flags local capture problems, not whole-image ungradability.
• Confidence is not calibrated. Scores are usable for ranking and thresholding, not as probabilities that a finding is real.

Serving

The FastAPI service in main.py is what the Fundusnap backend calls (FUNDUSNAP_AI_HOST).

Endpoints

Method	Path	Returns
POST	/inspect/fundus-artifacts/	JSON: filename, model_used, and detections[] with class_name, confidence, and integer box.{x1,y1,x2,y2}
POST	/visualize/fundus-artifacts/	image/jpeg — the input with red boxes and labels drawn on
GET	/	Health check with the list of loaded models

Both POST endpoints take a multipart/form-data upload under the field name file.

# JSON detection results
curl -X POST "http://localhost:8000/inspect/fundus-artifacts/" \
     -H "accept: application/json" \
     -F "file=@fundus_image.jpg"

# Annotated image
curl -X POST "http://localhost:8000/visualize/fundus-artifacts/" \
     -H "accept: image/jpeg" \
     -F "file=@fundus_image.jpg" \
     --output detection_result.jpg

Docker

docker build -t fundusnap-lesiondet .
docker run -p 8000:8000 fundusnap-lesiondet
curl http://localhost:8000/

Direct

pip install -r requirements.txt
uvicorn main:app --host 0.0.0.0 --port 8000

Interactive docs are at http://localhost:8000/docs. The model is loaded once at import time; if models/fundus_artifacts.pt is missing (e.g. Git LFS was not fetched), the service still starts but both detection endpoints return HTTP 500 — check GET / for an empty loaded_models.

License

Released under CC BY-NC 4.0 — attribution required, non-commercial use only.

Note for redistributors: these weights are fine-tuned from Ultralytics YOLO11, and the checkpoint itself records license: AGPL-3.0 (https://ultralytics.com/license). Ultralytics treats derivative weights as covered by AGPL-3.0 unless you hold an Ultralytics Enterprise licence, so review Ultralytics' licensing terms before redistributing or building on this model. The training data carries its own separate terms, which are not published here.

Citation

@software{fundusnap_lesiondet_yolo11m_20m,
  title  = {fundusnap-v1-lesiondet-yolo11m-20m: YOLO11m retinal lesion detector},
  author = {Fundusnap},
  url    = {https://huggingface.co/fundusnap/fundusnap-v1-lesiondet-yolo11m-20m},
  license = {CC-BY-NC-4.0}
}