NeuralShift predicts brain shift in temporal lobe resection from preoperative MRI alone, reaching Dice 0.97 and landmark TRE as low as 1.12 mm

A preoperative-to-intraoperative registration pipeline that accounts for the resection cavity and generates deformation fields quantitatively capturing brain shift. Standard registration tools struggle to normalise intraoperative MRI directly to a common stereotaxic space because of the surgical cavity and intensity heterogeneity; the pipeline first reorients iMRI using a subject-specific orthogonal AC–PC–IH coordinate system defined by three manually annotated landmarks, rigidly registers it to the corresponding pMRI in native space, and propagates the pMRI-to-MNI transformations to iMRI so both modalities share one MNI space. Qualitative MNI-space overlays for two representative cases show improved spatial correspondence while retaining the post-resection cavity; the pipeline also serves as the basis for constructing the supervision signals.

NeuralShift, a U-Net model that jointly predicts a dense displacement field, the intraoperative brain mask, and its signed distance function (SDF). The authors note that only one prior study, Shimamoto et al.'s W-Net, proposed a deep learning model for brain tissue deformation, and that its compensation targeted only the pre-resection state; this work extends prediction to deformation after temporal lobe resection and adds mask and SDF supervision to complement voxel-wise displacement regression with global shape constraints. Training uses a weighted multi-task objective: Cartesian MSE and spherical-coordinate losses (elevation, azimuth, magnitude) for displacement, Dice and edge losses for the mask, and MSE for the SDF; because voxel-wise physical ground truth is unavailable, the displacement field from NiftyReg F3D non-rigid registration is used as surrogate supervision, which the authors explicitly describe as a consistent surrogate rather than exact biomechanical ground truth.

Quantitative validation on 98 paired preoperative and intraoperative T1-weighted MRI scans with 9-fold cross-validation shows the predicted displacement field consistently reduces landmark TRE and improves mask overlap. Evaluation covers both global shape (Dice) and local geometric accuracy (TRE), reported on resection-side and midline landmarks, matching the clinical characteristic that larger deformations occur near the surgical corridor in temporal lobe resection. Dice between predicted and intraoperative masks is 0.97±0.01, higher than the 0.92±0.01 between preoperative and intraoperative masks; for left resections TRE drops from 4.58±1.47 mm to 2.96±1.08 mm (P1), 4.31±1.39 to 2.78±1.01 (P2), 4.76±1.53 to 3.05±1.09 (P4), and 4.49±1.46 to 2.87±1.02 (P6), and for right resections from 4.41±1.42 to 2.89±1.03 (P1), 4.18±1.35 to 2.71±0.97 (P2), 4.63±1.49 to 2.98±1.04 (P4), and 4.32±1.41 to 2.79±0.96 (P6), with midline landmarks P3 and P5 reaching a minimum TRE of 1.12 mm.

The work targets epilepsy patients undergoing left or right temporal lobe resection, takes only preoperative T1-weighted MRI plus a hemisphere indicator encoding resection laterality as input, and outputs a dense displacement field mapped to intraoperative space, the intraoperative brain mask, and its SDF, suited to a preoperative planning scenario. The authors split brain shift prediction into two tasks and investigate only the first here—learning shift from preoperative data alone; the second, incorporating intraoperative observations such as intraoperative ultrasound or electrophysiological measurements, is left for future work. The authors state the contributions will be publicly available after acceptance at the specified repository and plan broader validation on larger cohorts and task-specific surgical assessments.

The supervision target comes from the displacement field obtained by NiftyReg F3D registration, which the authors explicitly call a consistent surrogate rather than exact biomechanical ground truth, so TRE measures agreement with the registration result rather than deviation from real brain tissue. Open questions the authors list include: the limited dataset size may affect generalisability across surgical cases, the cohort may not fully capture variability across clinical scenarios, using preoperative MRI alone may miss complementary information from intraoperative ultrasound or electrophysiology, and TRE and Dice quantify geometric alignment and mask overlap but do not fully capture clinical usability, so prospective validation such as surgeon-in-the-loop evaluation or surgical outcome analysis is still needed. In addition, the timing of intraoperative MRI acquisition varies across procedures, and this is a retrospective study.