Mwana AI
An ultrasound classifier for breast cancer screening, built to run offline on-device.
Mwana AI is an ultrasound image classification model for breast cancer screening, designed for offline deployment in low-resource Tanzanian clinical settings. The model is exported to ONNX for on-device inference and paired with a Flutter mobile front end targeting Android, so screening does not depend on connectivity or a server round-trip. Developed at the Emerging Technologies for Healthcare Lab, MUHAS.
python export.py --format onnx --target android
model ▸ ultrasound classifier
export ▸ torch → onnx [ok]
runtime ▸ onnxruntime · on-device
network ▸ not required
// engineering notes
Decisions behind it
01
Offline-first by requirement
The deployment target is a clinic that cannot assume connectivity. Inference runs on the device, which shapes every decision upstream — architecture, size, and export path included.
02
PyTorch to ONNX Runtime
Training in PyTorch and HuggingFace Transformers, then exporting to ONNX, keeps research iteration fast while the shipped artefact is a portable runtime that runs on Android hardware.
03
Flutter front end for the field
A mobile front end targeting Android puts the model where the ultrasound is, rather than behind a workstation somewhere else in the building.
04
Clinical context, engineering discipline
Built at the ETH Lab, MUHAS, against real screening workflow constraints rather than a benchmark leaderboard.