Generation of synthetic training data for SEEG electrodes segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurosurgery

Background:

• Stereoelectroencephalography (SEEG) is vital for localizing epileptogenic zones.
• Accurate identification of SEEG electrode contacts is critical for treatment efficacy.
• Post-operative CT scans present significant metal streak artifacts, hindering deep learning segmentation.

Purpose of the Study:

• To develop a method for generating synthetic data to train neural networks for SEEG contact localization.
• To improve the accuracy of deep learning segmentation algorithms in the presence of CT metal artifacts.

Main Methods:

• Generated synthetic SEEG electrodes based on manufacturer specifications.
• Simulated metal artifacts in CT images using radon transform, beam hardening, and filtered back projection.
• Trained a UNet neural network using combinations of real, augmented, and synthetic data.

Main Results:

• Training with synthetic data significantly improved contact detection and segmentation accuracy compared to using only real or augmented data.
• Models trained solely on real/augmented data frequently resulted in misclassified artifacts or missed contacts.
• The trained network achieved rapid segmentation of post-operative CT slices.

Conclusions:

• Synthetic data generation is an effective strategy to enhance neural network performance for SEEG contact segmentation.
• This approach overcomes limitations of real-world data scarcity and artifact interference in medical imaging.