From pre- to intra-operative MRI: predicting brain shift in temporal lobe resection for epilepsy surgery

Brain shift reduces the accuracy of neuronavigation based on preoperative magnetic resonance imaging (MRI). Intraoperative MRI can depict this deformation but is costly, disruptive, and not widely available.

We propose NeuralShift, a U-Net-based framework that predicts a dense brain displacement field from preoperative MRI and resection laterality for patients undergoing temporal lobe resection. Of 98 paired preoperative and intraoperative MRI cases, eight were reserved as a fixed validation set for checkpoint selection. The remaining 90 were divided into nine disjoint folds; nine independently initialised models were trained with 80 cases and evaluated on a previously unseen 10-case test fold. Performance was assessed using registration-referenced Target Registration Error (TRE) at ipsilateral and midline landmarks and overlap between predicted and intraoperative brain masks.

The predicted masks achieved an unweighted mean fold-wise Dice score of 0.97, with a mean within-fold patient-wise standard deviation of 0.015 (fold means, 0.95-0.98), compared with 0.93 and 0.014, respectively, before deformation. Mean post-prediction TRE ranged from 1.12 to 3.05 mm across the evaluated landmarks and resection sides, compared with 1.46 to 4.76 mm before deformation.

NeuralShift demonstrates the feasibility of predicting a cohort-level prior for brain deformation using only information available before surgery. The registration-derived supervision, single-centre homogeneous cohort, and absence of external or independent physical validation preclude claims of clinical equivalence to biomechanical methods; prospective multi-centre validation is required. Code will be made publicly available after acceptance at https://github.com/SurgicalDataScienceKCL/NeuralShift .