Ultrasound Image Quality Evaluation using a Structural Similarity Based Autoencoder.

Area of Science:

• Medical Imaging
• Machine Learning
• Ultrasound Technology

Background:

• Ultrasound (US) imaging is crucial in clinical settings but demands extensive user training.
• Image quality is vital for accurate interpretation, yet errors and noise frequently compromise US image quality.
• Current quality assessment relies on experienced sonographers, posing a challenge for novices.

Purpose of the Study:

• To investigate the efficacy of autoencoders (AE) for automated ultrasound image quality assessment.
• To develop a machine learning model capable of distinguishing between high-quality and low-quality US images.
• To provide an objective measure of US image quality, assisting less experienced users.

Main Methods:

• An autoencoder (AE) model was trained using ultrasound images from 49 healthy subjects.
• Two loss functions, Structural Similarity Index Measure (SSIM) and Mean Squared Error (MSE), were employed for AE training.
• Reconstruction errors from the AE were used to train a Random Forest Classifier (RFC) for image quality classification.

Main Results:

• The SSIM-based AE achieved 71% accuracy for user-error-induced artifacts and 91% for noise.
• The MSE-based AE yielded 76% accuracy for artifacts and 83% for noise.
• The AE approach demonstrated potential in differentiating image quality based on reconstruction errors.

Conclusions:

• Autoencoders show promise for automated ultrasound image quality assessment.
• This technology can serve as an objective tool to aid researchers and clinicians in evaluating US image quality.
• The AE-based method could enhance diagnostic accuracy by ensuring reliable image quality.