BUSegNet: A Speckle-Aware Multi-Scale CNN with Noise Modeling and Gated Calibration for Breast Ultrasound Diagnosis

Breast cancer diagnosis remains a critical global health challenge, where timely and accurate detection can save lives. Ultrasound is a widely used, noninvasive imaging tool, yet its interpretation is prone to error, and standard deep learning models often falter due to speckle noise, data scarcity, and class imbalance. We introduce BUSegNet, a custom CNN built to handle the challenges of breast ultrasound. It uses a Spiral Multiscale Block to capture noise resilient features, a Gated Classifier Head to improve prediction stability, and a speckle noise augmentation strategy that mimics real world variability. On the BUSI dataset, BUSegNet achieved an overall F1 score of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 6. 7 \%}$</tex>, with strong per class performance: <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 5. 5 \%}$</tex>…