N-Anya 0.1: poultry droppings screening

N-Anya (Igbo for "eye") is the photo model inside Nche, a poultry health app built in Nigeria. A farmer photographs fresh droppings on any phone; N-Anya screens the photo for signs often seen with coccidiosis, salmonellosis and Newcastle disease, or a healthy flock, and the app tells the farmer in plain words what to do next and when to call their vet.

It is a screening tool, not a diagnosis. N-Anya gives a first look that tells a farmer what's worth checking. A veterinarian confirms the result and decides the treatment, always. In Nigeria, suspected bird flu must be reported to the veterinary authorities; N-Anya does not screen for it.

Version	droppings-0.1.0 (17 October 2026)
Task	Image classification, 4 classes: healthy, coccidiosis, salmonellosis, Newcastle disease
Architecture	EfficientNet-B0, ImageNet-pretrained, fully fine-tuned
Input	One colour photo, short side resized to 256 px, centre 224 by 224 px, ImageNet normalisation
Output	Four probabilities; below 0.6 confidence the app says "not sure, show your vet"
Runs as	ONNX, on Nche's own server; the app sends the photo, never the model to the phone
Weights	Not published. N-Anya is available through the Nche app: app.nche.app
Full results, explained for farmers	nche.app/ai

Results

Tested on 1,201 photos whose disease a laboratory confirmed by PCR (Machuve et al. 2021, the PCR-annotated set), none of which the model saw during training. Training used the separate farm-labelled photos from the same project.

Right on lab-confirmed photos	966 of 1,201 (80%); between 79% and 82% across three training runs
On farm photos like its training data	98.9% (validation)
Newcastle disease caught	178 of 186 as the top answer (96%); 166 of 186 named with confidence
Answered "not sure, show your vet"	25% of photos

Per class (the test set, top answer)

Lab found	Photos	Caught (recall)	Right when it says so (precision)
Healthy	312	56%	79%
Coccidiosis	357	78%	95%
Salmonellosis	346	97%	71%
Newcastle disease	186	96%	83%

Confusion matrix

Rows: what the lab found. Columns: N-Anya's top answer.

	Healthy	Coccidiosis	Salmonellosis	Newcastle
Healthy	174	15	107	16
Coccidiosis	30	277	30	20
Salmonellosis	8	1	337	0
Newcastle	7	0	1	178

Against published models

The same task (train on farm photos, test on the same 1,201 de-duplicated PCR photos), from the leakage audit by Pratama et al. 2026 (mean of three seeds each):

Model	Accuracy	Newcastle recall
DenseNet121	62%	41%
MobileNetV2	65%	38%
YOLO11n-cls	69%	27%
N-Anya 0.1 (EfficientNet-B0)	80%	96%

The main differences: the whole network is fine-tuned rather than a frozen backbone, the farm photos are de-duplicated with the audit's keep-lists and near-duplicate groups so no near-copy sits in both training and test, and training uses strong colour, blur and crop changes so the model learns the droppings rather than the photo.

Where it struggles

Published on purpose, so farmers, vets and partners know what to expect.

• Healthy droppings called salmonellosis: 107 of 312 healthy photos (34%). A healthy flock can be told "worth checking for salmonellosis". It happened in all three training runs (97 to 107 photos), so it's a pattern, not bad luck.
• Some coccidiosis called healthy: 30 of 357 (8%). The dangerous direction, 30 to 44 photos across the runs. It's why the app always says to call your vet if birds look sick, whatever the photo shows.
• "Not sure" on about a quarter of photos: safe, because those go straight to the vet, but frequent. Clearer photos and Nigerian training data should bring it down.
• Not tested on Nigerian farms yet: every test photo so far comes from Tanzania. Nigerian feed, breeds, floors and phones may look different. Proving it here is what the pilot with founding vets is for.

Intended use

• For: poultry farmers screening fresh droppings with a phone, to know early whether something looks wrong and when to call a vet; vets receiving the farmer's report, so they arrive knowing what to check.
• Not for: diagnosing a bird, choosing a treatment or a medicine, deciding whether to cull, certifying a flock as disease-free, or screening for avian influenza or any disease outside the four classes. Not for other species.
• Always with a vet. The app's advice says when to call a vet, and every report can be opened, confirmed and signed by one.

Training data

	Source	Licence	Used for
Farm-labelled droppings photos, Tanzania	Machuve et al. (2021), Machine Learning Dataset for Poultry Diseases Diagnostics, doi:10.5281/zenodo.4628934	CC BY 4.0	training and validation
PCR-annotated droppings photos	Machuve et al. (2021), PCR-annotated version, doi:10.5281/zenodo.5801834	CC BY 4.0	test only
Keep-lists and near-duplicate groups	Pratama et al. (2026), lab-to-farm leakage audit, doi:10.5281/zenodo.23374400	CC BY 4.0	de-duplication

The farm photos are split by near-duplicate group, so related photos never cross from training into validation. No photos from Nche's users were used to train this version.

Before it leaves the pilot

N-Anya stays in a pilot with founding vets until it meets all three, on every training run:

Rule	Needed	Runs 1 / 2 / 3
Sick flocks told to check	80% or more	81.2% / 79.6% / 78.3%
Healthy flocks wrongly alarmed	25% or less	20.8% / 23.7% / 16.7%
Newcastle cases caught	70% or more	89.2% / 93.5% / 88.7%

Two of the three runs fall just short on the first rule, so it is not cleared yet.

Privacy

Photos farmers send are used to give them a result and a report. Only if they allow it are they used, with what vets later find, to improve screening; they can switch that off in the app at any time. See the privacy notice.

Coming next

N-Ntị (Igbo for "ear"): screening 10 seconds of sound from the poultry house for breathing problems and distress. In training; no results to publish yet.

Citation

@misc{nche2026nanya,
  title  = {N-Anya 0.1: poultry droppings screening},
  author = {{Nche}},
  year   = {2026},
  url    = {https://nche.app/ai},
  note   = {Model droppings-0.1.0, EfficientNet-B0, tested on 1,201 PCR-confirmed photos}
}

Please also cite the datasets above (Machuve et al. 2021; Pratama et al. 2026).

Contact

Nche, built in Nigeria by our founder, who runs a commercial poultry farm in Nigeria. nche.app/founder · hello@nche.app · research and lab partnerships welcome.