A convolutional neural network Cascade for plantar pressure images registration

Area of Science:

• Biomedical Engineering
• Computer Vision
• Medical Imaging

Background:

• High-resolution plantar pressure images (PPIs) are crucial for clinical gait analysis.
• Image registration is essential for functional analysis of PPIs, but traditional methods are computationally expensive.
• Existing iterative optimization techniques for PPI registration often lack speed and efficiency.

Purpose of the Study:

• To develop a rapid and accurate PPI registration technique.
• To investigate a single PPI registration method that maintains adequate correspondence using various metrics.
• To overcome the computational limitations of conventional PPI registration methods.

Main Methods:

• A cascaded convolutional neural network (CNN) was proposed for PPI registration.
• The CNN model was trained to regress registration parameters directly from image differences.
• Registration performance was evaluated using Mean Squared Error (MSE), Exclusive OR (XOR), and Mutual Information (MI) metrics, compared against Principal Axes (PA) and Center of Pressure (COP) methods.

Main Results:

• The CNN-based method achieved registration accuracy comparable to MSE and XOR methods.
• Registration speed comparison: MSE (30.6s), XOR (24.2s), PA (0.02s), COP (25.6s), and the proposed CNN model (0.05s).
• The CNN model demonstrated significantly faster registration times than iterative methods like MSE and XOR.

Conclusions:

• The proposed CNN-based registration approach offers a practical solution for high-accuracy PPI registration.
• This method significantly reduces computational time, enabling near-real-time applications.
• The pre-trained CNN model is suitable for developing efficient clinical gait analysis tools.