Robustness of lead reconstruction for deep brain stimulation modelling and probabilistic mapping

Deep brain stimulation (DBS) has become an efficacious therapy for multiple indications. With the advent of directional leads, increasing stimulation options complexify manual programming. Therefore, automated programming algorithms based on probablisitic mapping are being tested for parameter prediction. Such approaches require computational lead reconstruction routines that are already broadly used. However, the robustness of lead reconstruction across distinct image sets of a same patient remains unclear. To assess lead reconstruction systematically, we identified retrospectively 34 DBS patients with Parkinson's Disease (PD) or Essential Tremor, who received two distinct postoperative CT-scans. Each CT-scan was processed independently using the Lead-DBS toolbox. Between both image sets, we compared lead tip coordinates and volumes of tissue activation (VTA) for each hemisphere. Group-level probabilistic maps of clinical improvement were compared between sets for PD patients. Mean lead tip translation between CTs was 0.79mm (range: 0.21-2.35mm). Pneumocephalus did not significantly affect reconstruction robustness. Lead translation was comparable in the patient native space and after normalization to the template brain. Individual-level VTA comparison revealed a mean Dice coefficient of 0.73 (range: 0.33-0.94), which decreased with lower amplitudes of stimulation. Group-level N-images and clinical improvement maps were robust (Dice coefficient respectively 0.88 and 0.90). Computational normalization and pneumocephalus correction were satisfying in our cohort. However, individual-level VTA variability was observed, potentially caused by slightly inaccurate CT-to-MRI co-registration or by brain shift sources other than pneumocephalus. These variabilities vanish at the group level, suggesting that current lead reconstruction routines are sufficient for probabilistic sweet spot identification.