Hybrid U-Net for segmentation of SEEG electrodes on post-operative CT scans

A common treatment option for pharmacoresistant epilepsy is to surgically remove epileptogenic zone. Stereoelectroencephalography (SEEG)_is a minimally invasive surgical procedure used to identify such zones. Precisely determining positions of all of implanted SEEG electrodes is crucial to design a resection plan. Metallic electrode contacts produce strong artefacts in CT scans which makes localisation process difficult and imprecise. We propose an automatic approach for accurate localisation of SEEG electrode contacts using a combination of a 2D and 3D U-Net architecture. The proposed hybrid network makes the best out of both models and makes more accurate predictions, resulting in a decrease of false-positive and false-negative segmentations. The network was trained on 36 data sets and evaluated on four different metrics. The Hybrid model outperformed both the 2D and the 3D U-Net model. To complete the electrode segmentation process, segmented contacts are linked into electrodes using Gaussian mixture models.