Extracting keyframes of breast ultrasound video using deep reinforcement learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Ultrasound (US) is crucial for breast cancer screening, particularly for dense breasts.
• Current computer-aided diagnosis tools often overlook the importance of keyframe selection from dynamic US videos.
• Lesion variability in shape, size, and location complicates accurate diagnosis and keyframe identification.

Purpose of the Study:

• To develop a reinforcement learning framework for automatic keyframe extraction from breast US videos.
• To address challenges in lesion recognition and class imbalance during screening.
• To improve the representativeness of selected frames for enhanced diagnostic accuracy.

Main Methods:

• Proposed a reinforcement learning-based framework for automatic keyframe extraction from breast US videos of variable length.
• Integrated a detection-based nodule filtering module.
• Developed a novel reward mechanism incorporating anatomical and diagnostic lesion features.
• Implemented a loss function to mitigate class imbalance issues.

Main Results:

• The proposed framework successfully extracts representative keyframe sequences from breast US videos.
• The innovations, including the reward mechanism and loss function, significantly benefit the keyframe selection process.
• The system demonstrates effectiveness across various screening conditions.

Conclusions:

• The developed reinforcement learning framework offers an effective solution for automatic keyframe selection in breast cancer screening.
• This approach enhances the diagnostic potential of US by focusing on critical visual information.
• The method shows promise for improving the efficiency and accuracy of AI-assisted breast cancer diagnosis.