3D Ultrasonic Brain Imaging with Deep Learning Based on Fully Convolutional Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Ultrasound imaging offers advantages over MRI and CT but faces challenges in adult transcranial brain imaging due to skull acoustic impedance.
• Conventional ultrasound struggles with skull penetration, limiting its application in brain imaging.

Purpose of the Study:

• To develop and validate a 3D AI algorithm for precise transcranial brain ultrasound reconstruction.
• To overcome the limitations of conventional ultrasound in adult brain imaging.

Main Methods:

• A 3D AI algorithm, Brain Imaging Full Convolution Network (BIFCN), was developed, integrating waveform modeling and deep learning.
• The BIFCN network architecture includes an input layer, four convolution layers, and one pooling layer for training.
• The algorithm was tested using simulation experiments and laboratory reconstructions with pure water as the initial model.

Main Results:

• Simulation experiments showed a high Pearson correlation coefficient between reconstructed and true brain images.
• Laboratory results demonstrated impressive 3D reconstruction accuracy, even without prior information.
• The 3D BIFCN network trains in 8 hours and reconstructs 10 samples in 12.67 seconds, indicating high efficiency.

Conclusions:

• The 3D BIFCN algorithm provides an accurate and efficient method for 3D brain imaging using ultrasound.
• This AI-driven approach enables fast and precise brain tissue imaging, mapping wavefield data to 3D models.
• The frequency shift phenomenon observed in blood may offer new quantitative insights for whole-brain blood imaging using BIFCN.