Displacement detection with sub-pixel accuracy and high spatial resolution using deep learning.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Conventional ultrasound displacement detection methods struggle with high spatial resolution and sub-pixel accuracy.
• Existing techniques often assume neighborhood uniformity, limiting their ability to detect small tissue deformations.

Purpose of the Study:

• To develop and validate a deep learning-based approach for high spatial resolution, two-dimensional, and sub-pixel displacement detection using ultrasound.
• To overcome the limitations of conventional methods in accurately measuring tissue displacement.

Main Methods:

• A deep learning network, adapted from FlowNet2, was developed to process ultrasound images and output displacement distributions.
• A simulated ultrasound image dataset was created for training the network.
• Performance was evaluated using simulated data and clinical ultrasound images of liver tissue treated with high-intensity focused ultrasound (HIFU), compared against the Lucas-Kanade method.

Main Results:

• The deep learning method achieved high accuracy (above 0.5 µm and 0.2 µm) and precision (above 0.4 µm and 0.3 µm) for displacements within ±40 µm.
• Excellent spatial resolution was demonstrated (1.1 mm lateral, 0.8 mm axial).
• The method successfully visualized lateral displacement distributions, crucial for identifying lesion margins in HIFU treatments, and showed improvements in experimental data.

Conclusions:

• A deep learning methodology enables precise two-dimensional and sub-pixel displacement detection with high spatial resolution via ultrasound.
• This advanced technique allows for effective monitoring of localized tissue deformations, particularly those induced by HIFU exposure.