Intraoperative Ultrasound-Based Displacement Mapping Through Deep Learning.

Area of Science:

• Neurosurgery
• Medical Imaging
• Artificial Intelligence

Background:

• Brain shift during surgery compromises navigation accuracy.
• Current methods rely on intraoperative MRI, limiting real-time application.
• Intraoperative ultrasound (iUS) offers a low-cost, real-time alternative.

Purpose of the Study:

• To develop a deep learning framework for predicting voxel-wise brain deformation from iUS scans.
• To enable localized brain shift compensation without preoperative MRI.
• To improve intraoperative adaptability in neurosurgery.

Main Methods:

• Trained two 3D neural network architectures and an ensemble on 13 patients' iUS data.
• Utilized paired pre-resection and postresection 3D iUS scans with landmark annotations.
• Evaluated performance using regression metrics and leave-one-patient-out cross-validation.

Main Results:

• The ensemble model balanced root median squared error and directional accuracy.
• Gradient-weighted Class Activation Mapping identified key deformation regions.
• Most patients showed sub-2mm median absolute error, with challenges noted for atypical anatomy.

Conclusions:

• Demonstrated feasibility of estimating brain shift directly from iUS-to-iUS scans using deep learning.
• The approach provides dense, real-time deformation fields for enhanced surgical navigation.
• Future work requires larger, diverse datasets and multitask learning for improved accuracy.