Make It Less Complex: Autoencoder for Speckle Noise Removal-Application to Breast and Lung Ultrasound

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Ultrasound (US) imaging is crucial for diagnosing conditions like COVID-19 and breast cancer.
• Speckle Noise (SN) in US images degrades image quality and reduces lesion visibility.
• Existing SN removal methods, including complex deep learning models, often focus on simulated noise and lack real-world applicability.

Purpose of the Study:

• To develop and evaluate a simpler Convolutional Neural Network Autoencoder (CNN-AE) for removing naturally occurring SN from breast and lung US images.
• To compare the performance of the proposed CNN-AE against traditional Median and Lee filters.
• To assess the impact of SN removal on the classification of malignant versus benign breast lesions using a CNN model.

Main Methods:

• A CNN-AE with fewer than 30,000 parameters was trained using original US images as targets and noisy images (with simulated SN at four levels) as input.
• Performance was evaluated using Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR) against original images.
• A CNN classifier was trained using original, denoised (CNN-AE and filters), and noisy US images to differentiate breast lesions.

Main Results:

• The CNN-AE significantly outperformed Median and Lee filters in removing simulated SN across all noise levels, as measured by SSIM and PSNR.
• While original US images yielded the highest overall accuracy and Matthews Correlation Coefficient (MCC) for lesion classification, CNN-AE denoised images showed superior sensitivity and negative predicted values, especially at higher noise levels.
• The simpler DL model demonstrated fewer misclassifications of malignant breast lesions compared to using original images or Median filter denoised images.

Conclusions:

• A less complex, clinically focused deep learning approach (CNN-AE) effectively removes speckle noise from real-world ultrasound images.
• This method improves the diagnostic performance for breast lesion classification, particularly in enhancing sensitivity and negative predictive value.
• The study highlights the importance of developing simpler, clinically applicable deep learning models for medical image noise reduction.