SonoCNN — BreastMNIST breast-ultrasound classifier

A compact CNN (287,330 parameters) for benign-vs-malignant classification of 28×28 breast-ultrasound images, released with something most model uploads don't have: a script that re-derives every headline metric from the shipped weights and data, so the numbers below are checkable, not just claimed.

• Test accuracy 0.8718 and ROC AUC 0.8709 on the held-out BreastMNIST test split (156 images), vs a 0.7308 majority-class baseline — independently reproduced from the artifacts in this repo on 2026-10-28 (see Verify it yourself).
• Weights in three formats: model.safetensors, the original PyTorch checkpoint (assets/sonocnn.pt), and ONNX (sonocnn.onnx, parity vs PyTorch measured at 9.5e-07 max logit difference).
• Trained from scratch by Habib Ur Rehman for the SonoSight explainable-ultrasound project (OpenCV AI Competition 2026 / Bolt Builders Week).

Research and benchmark use only. This is a 28×28 benchmark model trained on 546 images. It is not a medical device, it has never been evaluated on clinical data at native resolution, and it must not be used for diagnosis or screening.

Verified results — BreastMNIST test split (n = 156)

Metric	Value	Source
Accuracy	0.8718 (136/156)	reproduced 2026-10-28
ROC AUC	0.8709	reproduced 2026-10-28
Balanced accuracy	0.8145	reproduced 2026-10-28
Malignant recall (sensitivity)	0.9386 (107/114)	reproduced 2026-10-28
Benign recall (specificity)	0.6905 (29/42)	reproduced 2026-10-28
Malignant precision	0.8917 (107/120)	reproduced 2026-10-28
Benign precision	0.8056 (29/36)	reproduced 2026-10-28
Confusion matrix	[[29, 13], [7, 107]] (rows: true benign, true malignant)	reproduced 2026-10-28
Majority-class baseline	0.7308	reproduced 2026-10-28
Best validation accuracy	0.8718 (68/78)	recorded at training time

Read the weak row honestly: the model is much better at catching malignant cases (93.9%) than at clearing benign ones (69.1%) — 13 of 42 benign test images are flagged malignant. On a test set this small, the 95% confidence interval on accuracy is roughly ±5 points (≈0.81–0.92). The full reproduced record is in assets/evaluation_results.json; the training-time record is in assets/sonocnn_metrics.json; they agree exactly on every headline number.

Quickstart

pip install "git+https://github.com/habib-analyst/sonocnn.git"

from sonocnn import load_model, predict

model = load_model("model.safetensors")   # or assets/sonocnn.pt
result = predict(model, "assets/demo_samples/benign.png")
# {'label': 'benign', 'confidence': 0.993,
#  'probabilities': {'benign': 0.993, 'malignant': 0.007}}

Preprocessing contract (identical to training): grayscale → resize to 28×28 → /255 → normalize with the BreastMNIST train-split mean/std (0.32758 / 0.20573). Labels: 0 = benign (normal/benign), 1 = malignant.

ONNX: sonocnn.onnx takes the same normalized (N, 1, 28, 28) float tensor and returns logits.

Verify it yourself

python examples/verify_metrics.py

This re-evaluates the shipped checkpoint on the bundled test split (assets/breastmnist.npz) and compares against the training-time metrics file. On 2026-10-28 it printed ALL HEADLINE METRICS REPRODUCED. Reproduction protocol matters: an earlier audit of this checkpoint briefly "failed" at 0.73 accuracy because the evaluator forgot the /255 step — the script exists so nobody has to take the card's word for anything.

python examples/predict_demo.py   # bundled samples + optional ViewConsensus panel
python -m pytest -q               # 8 tests: formats agree, metrics files agree, parity

The demo also composes SonoCNN with the sibling ViewConsensus library: on the first malignant test image, the single-view prediction is malignant at 70.2%, but the 8-view consensus panel reports 76.9% ±18.9% and returns an unsure — defer-to-human verdict. A confident single number is not the whole story; the spread is information.

Model details

• Architecture: 3 conv blocks (32→64→128 channels, BatchNorm, ReLU, max-pool, dropout) + 2-layer MLP head. 287,330 trainable parameters; 1.2 MB in safetensors.
• Training: from scratch on the BreastMNIST train split (546 images), 40 epochs, AdamW (lr 3e-3, weight decay 1e-4, cosine schedule), batch 32, class-weighted cross-entropy for the 73/27 malignant/benign imbalance; checkpoint selected by best validation accuracy. Recorded training time: 65.7 s on CPU.
• Augmentation: horizontal flips, ±3 px translations (edge-replicated), small Gaussian noise. The translation augmentation was added after Grad-CAM showed an earlier model fixating on the dark bottom border of BreastMNIST images instead of lesion regions — accuracy was traded (~0.6 pt) for attention on the lesion.
• Explainability: the original build includes Grad-CAM overlays; confidences are softmax outputs and are known to be overconfident on easy cases.

What SonoCNN is not

• Not a clinical tool. Native-resolution ultrasound (e.g. the bundled 256×256 stills) is out of distribution for this model: the bundled malignant still lands at p(malignant) = 0.499 — a coin flip — after resizing to 28×28, even though its native 28×28 test image scores 0.702. The demo prints this rather than hiding it.
• Not state of the art. It is a small from-scratch CNN on a small benchmark, released because its numbers are verified and its failure modes are documented, not because it tops a leaderboard.
• Not trained on private or clinical data. Only the public BreastMNIST benchmark (below).

Data and attribution

BreastMNIST is part of MedMNIST v2 (Yang et al., MedMNIST v2 — A large-scale lightweight benchmark for 2D and 3D biomedical image classification, Scientific Data 2023), derived from the BUSI breast-ultrasound dataset (Al-Dhabyani et al., 2020). The dataset is licensed CC BY 4.0; assets/breastmnist.npz is redistributed here for metric reproduction, with attribution. The MIT license of this repository covers the code and the trained weights, not the dataset.

Files

File	Contents
model.safetensors	Release weights (safetensors)
sonocnn.onnx / sonocnn.onnx.data	ONNX export (logits)
config.json	Architecture + preprocessing config
src/sonocnn/	model.py (architecture), infer.py (loading + prediction API)
examples/predict_demo.py	Inference demo incl. optional ViewConsensus panel
examples/verify_metrics.py	Full metric reproduction from the artifacts
assets/sonocnn.pt	Original training checkpoint (state dict + metrics)
assets/sonocnn_metrics.json	Training-time metrics record
assets/evaluation_results.json	Independently reproduced metrics (2026-10-28)
assets/breastmnist.npz	BreastMNIST splits (CC BY 4.0, see attribution)
assets/demo_samples/	Two 256×256 ultrasound stills used by the demo

License

MIT for code and weights (see LICENSE). Dataset under CC BY 4.0 as above.