A Systematic Review on Data-Driven Brain Deformation Modeling for Image-Guided Neurosurgery

Accurate compensation of brain deformation is a critical challenge for reliable image-guided neurosurgery, as surgical manipulation and tumor resection induce tissue motion that misaligns preoperative planning images with intraoperative anatomy and longitudinal studies. In this systematic review, we synthesize recent AI-driven approaches developed between January 2020 and April 2025 for modeling and correcting brain deformation. A comprehensive literature search was conducted in PubMed, IEEE Xplore, Scopus, and Web of Science, with predefined inclusion and exclusion criteria focused on computational methods applied to brain deformation compensation for neurosurgical imaging, resulting in 41 studies meeting these criteria. We provide a unified analysis of methodological strategies, including deep learning-based image registration, direct deformation field regression, synthesis-driven multimodal alignment, resection-aware architectures addressing missing correspondences, and hybrid models that integrate biomechanical priors. We also examine dataset utilization, reported evaluation metrics, validation protocols, and how uncertainty and generalization have been assessed across studies. While AI-based deformation models demonstrate promising performance and computational efficiency, current approaches exhibit limitations in out-of-distribution robustness, standardized benchmarking, interpretability, and readiness for clinical deployment. Our review highlights these gaps and outlines opportunities for future research aimed at achieving more robust, generalizable, and clinically translatable deformation compensation solutions for neurosurgical guidance. By organizing recent advances and critically evaluating evaluation practices, this work provides a comprehensive foundation for researchers and clinicians engaged in developing and applying AI-based brain deformation methods.

💡 Research Summary

This systematic review examines AI‑driven brain deformation modeling techniques published between January 2020 and April 2025, focusing on their application to image‑guided neurosurgery. A comprehensive search of PubMed, IEEE Xplore, Scopus, and Web of Science identified 41 studies that met predefined inclusion criteria emphasizing computational methods for intra‑operative brain shift compensation. The authors categorize the approaches into five major families.

1. Deep learning‑based image registration – encoder‑decoder networks such as VoxelMorph, TransMorph, and DiffReg learn dense deformation fields directly from pre‑operative and intra‑operative image pairs, optimizing similarity losses (e.g., normalized mutual information, structural similarity) without handcrafted features. These models achieve near‑real‑time inference on GPUs.

2. Direct deformation field regression – 3‑D U‑Net, ResNet‑Transformer hybrids, and other convolutional architectures take paired volumes as input and output voxel‑wise displacement vectors, bypassing iterative optimization.

3. Synthesis‑driven multimodal alignment – Generative adversarial networks (CycleGAN, Pix2Pix) or variational auto‑encoders translate low‑quality intra‑operative ultrasound (iUS) into MRI‑like representations, mitigating modality gaps and improving downstream registration accuracy.

4. Resection‑aware architectures – Networks incorporate binary resection masks or use partial‑observation learning to handle missing correspondences caused by tissue removal, allowing dynamic updating of deformation fields as the tumor is excised.

5. Hybrid physics‑informed frameworks – Biomechanical finite‑element simulations or elasticity regularization terms are embedded as priors or loss functions, providing physical plausibility especially when training data are scarce.

Public datasets dominate the experimental landscape: MICCAI Brain Shift Challenge, RESECT, BRATS, and several iUS‑MRI co‑registration collections, typically comprising 30–50 patients (≈1,200–2,500 volumes). However, data imbalance (tumor size, location) and variable annotation quality are repeatedly cited as sources of performance variance.

Evaluation metrics include mean squared error, Dice similarity coefficient, target registration error (TRE), and Hausdorff distance. Most studies employ k‑fold or leave‑one‑out cross‑validation, yet independent external validation is rare, limiting confidence in out‑of‑distribution robustness. Uncertainty quantification is addressed through Monte‑Carlo dropout, Bayesian neural networks, or ensemble predictions, with visualizations intended to guide surgeons in high‑risk regions.

The review highlights several systemic limitations. AI models, while achieving higher accuracy and faster runtimes than classic iterative registration, often lack robustness to unseen anatomical variations, suffer from non‑standardized preprocessing pipelines, and remain opaque to clinicians. Real‑time clinical deployment is further hampered by integration challenges (data transfer, synchronization), regulatory hurdles, and the need for lightweight inference on operating‑room hardware.

Future research directions proposed include: (i) building large, multi‑institutional, harmonized datasets with standardized benchmarks; (ii) developing hybrid self‑supervised or physics‑informed models that combine data efficiency with physical consistency; (iii) advancing rigorous uncertainty estimation and risk‑aware decision support; (iv) engineering edge‑AI solutions for low‑latency intra‑operative use; and (v) establishing regulatory and ethical validation frameworks. Addressing these gaps could translate AI‑based deformation compensation from research prototypes into reliable, real‑time tools that improve tumor resection completeness and patient safety in neurosurgical practice.