Noise-aware framework tests breast ultrasound AI robustness

In breast ultrasound classification, Gaussian noise reduced a custom CNN's accuracy by 36.1% at 5 dB SNR, and Poisson noise reduced Inception V3 by 35.2%. Noise-matched training boosted accuracy up to 56.5% at 5 dB, underscoring the need for noise-robust evaluation.

• The study used the publicly available BUSI dataset (780 ultrasound images, 600 patients) and evaluated a custom CNN and a pretrained Inception V3.
• Gaussian noise reduced custom CNN accuracy by 36.1% at 5 dB SNR; Poisson noise reduced Inception V3 by 35.2%; speckle noise primarily affected malignant recall.
• Noise-matched training improved accuracy by up to 56.5% at 5 dB, with smaller gains at lower noise levels.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.