A Noise-Aware Robustness Evaluation Framework for Breast Cancer Classification in Ultrasound Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a leading cause of mortality worldwide, necessitating accurate detection methods.
• Ultrasound is a widely used imaging modality for breast cancer screening due to its accessibility and safety.
• Deep learning (DL) models show promise for breast ultrasound analysis, but their reliability is limited by image noise and acquisition variations.

Purpose of the Study:

• To develop and evaluate a noise-aware training framework for deep learning-based breast ultrasound classifiers.
• To assess the robustness of DL models against clinically relevant image degradations, including Gaussian, Poisson, and speckle noise.
• To improve the accuracy and malignant recall of breast ultrasound analysis in the presence of noise.

Main Methods:

• A noise-aware training framework was proposed, incorporating controlled noise modeling and robustness assessment.
• The framework was applied to the publicly available Breast Ultrasound Images (BUSI) dataset.
• Evaluated a custom Convolutional Neural Network (CNN) and a pretrained Inception V3 model using noise-matched training strategies.

Main Results:

• Gaussian noise significantly impacted the custom CNN's accuracy, while Poisson noise affected the Inception V3 model.
• Speckle noise led to reduced malignant recall, particularly under high noise conditions.
• Noise-matched training enhanced accuracy by up to 56.5% at a 5 dB signal-to-noise ratio (SNR), demonstrating improved model performance.

Conclusions:

• Noise-aware evaluation and training are crucial for enhancing the reliability of deep learning models in breast ultrasound analysis.
• The proposed framework effectively mitigates the negative impact of image noise on diagnostic performance.
• Implementing noise-aware strategies is essential for safer and more accurate breast cancer detection using ultrasound.