Deep learning-based automated segmentation of resection cavities on postsurgical epilepsy MRI

Area of Science:

• Neurosurgery
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of surgical resection sites is vital for clinical evaluation and neuroimaging research.
• Postoperative imaging analysis, particularly for epilepsy surgery, requires precise delineation of resection extent and remnant tissue.

Purpose of the Study:

• To develop and validate an automated algorithm for segmenting surgical resection cavities in postoperative MRI of epilepsy patients.
• To create a user-friendly graphical user interface (GUI) for estimating remnant brain volumes, including hippocampal remnant tissue.

Main Methods:

• A majority vote ensemble of three U-Net convolutional neural networks was employed for segmentation.
• The algorithm was trained and validated using retrospective T1-weighted MRI data from 62 temporal lobe epilepsy patients.
• Performance was evaluated using Dice-Sørensen coefficient (DSC), Hausdorff distance, and volume correlation with manual segmentations.

Main Results:

• The algorithm achieved median DSCs of 0.84 (cross-validation) and 0.74 (held-out test set), approaching inter-rater reliability.
• High correlation (r=0.94, r=0.87) was observed between automated and manual volume estimates.
• Strong agreement (r=0.90) was found for postoperative hippocampal remnant volumes, with a low false positive rate.

Conclusions:

• The developed automated segmentation algorithm provides accurate and reliable quantification of postoperative brain volumes after epilepsy surgery.
• The GUI-based pipeline facilitates clinical assessments, surgical outcome prediction, and neuroimaging research applications.
• This tool enables precise measurement of remnant brain tissue, improving the analysis of surgical interventions in epilepsy.