Abstract: This research presents a comprehensive system for automated localization of SEEG (Stereoelectroencephalography) electrode contacts in post-operative CT scans. The system employs a novel multi-stage ensemble approach combining deep learning-based brain extraction, adaptive image enhancement, global voting consensus, and machine learning-based confidence estimation to achieve sub-millimeter localization accuracy (mean error: 0.33 mm). The automated pipeline reduces processing time from over 4 hours to approximately 30 minutes while maintaining clinical accuracy standards of 98.8% within the 2mm threshold.
📄 Read Full Thesis (PDF)

Key Results

98.8% Accuracy (≤2mm)
0.33mm Mean Error
30min Processing Time
100% Detection Rate

System Overview

The SEEG electrode localization system implements a six-stage processing pipeline: brain extraction using MONAI-based 3D U-Net, multi-modal image enhancement with 7 parallel approaches, adaptive thresholding via machine learning, global voting ensemble across 38 mask variants, confidence-based authentication using LightGBM, and trajectory reconstruction through hybrid DBSCAN-Louvain clustering.

Automated trajectory reconstruction and visualization

Repository Ecosystem

This research is organized across multiple repositories, each focusing on specific components of the SEEG localization system:

Brain Segmentation Model

Deep learning-based 3D brain extraction using MONAI and PyTorch. Implements 3D U-Net architecture for automated brain tissue segmentation from post-operative CT scans.

PyTorch MONAI 3D U-Net

ML Threshold Models

Machine learning models for automated threshold prediction in CT-based electrode localization. Random Forest regressor achieving R²=0.990 accuracy.

Random Forest scikit-learn Optuna

Core Algorithms

Research repository containing pipeline implementation, ensemble voting algorithms, feature extraction, confidence scoring, and trajectory reconstruction methods.

LightGBM NetworkX DBSCAN

Slicer Extension

Production-ready 3D Slicer extension for clinical deployment. User-friendly interface integrating all pipeline stages for end-to-end electrode localization.

3D Slicer Python VTK

Methodology

The research employs a Leave-One-Patient-Out Cross-Validation (LOPOCV) strategy on an 8-patient development cohort from Hospital del Mar, Barcelona. The system combines classical image processing techniques with modern machine learning approaches to achieve robust performance across varying image qualities and electrode configurations.

Technical Documentation

Detailed technical slides covering specific aspects of the methodology:

Brain Extraction
MONAI-based 3D U-Net architecture and training methodology State-of-the-Art Comparison
Benchmarking against existing methods (GARDEL, iELVis) Trajectory Reconstruction
Hybrid clustering and consensus algorithms

Clinical Impact

This system has been validated and deployed at Hospital del Mar's Epilepsy Unit, demonstrating real-world clinical applicability. The automated pipeline addresses a critical bottleneck in epilepsy surgery planning, where manual electrode localization requires extensive collaboration between neurosurgeons and neurophysiologists. By reducing processing time by over 95% while maintaining clinical accuracy, the system enables faster surgical planning and improved patient outcomes.

Citation

@mastersthesis{avalos2025seeg,
  title={Medical Software Module in 3D Slicer for Automatic Segmentation
         and Trajectory Reconstruction of SEEG Electrodes Using AI
         and Data Science},
  author={Ávalos Morillas, Rocío},
  year={2025},
  school={Universitat Politècnica de Catalunya},
  type={Bachelor's Thesis},
  note={Biomedical Engineering}
}