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

TL;DR

Brain shift during neurosurgery disrupts alignment between preoperative plans and evolving intraoperative anatomy. The authors synthesize 41 studies (2020–2025) across DL-based registration, direct displacement-field regression, synthesis-driven multimodal alignment, resection-aware architectures, and physics-informed/hybrid models, with careful attention to datasets and benchmarking. Key findings indicate DL methods dominate recent work and enable rapid inference, yet challenges in out-of-distribution robustness, multimodal registration, interpretability, and clinical translation persist. The review outlines concrete gaps and opportunities aimed at delivering robust, generalizable deformation compensation capable of supporting safe, effective image-guided neurosurgery.

Abstract

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.

Figures (7)

• Figure 1: Neurosurgical workflow from preoperative assessment through postoperative care, including intraoperative phases. Green stars indicate the same location across imaging modalities and time points. (A) Image Acquisition: Following clinical evaluation and discussion of treatment options, preoperative imaging is acquired. (B) Surgical Planning: Multiparametric magnetic resonance imaging (MRI) is the primary imaging modality used to delineate tumor boundaries and assess the proximity of eloquent and critical brain areas. Additional modalities, such as computed tomography, functional MRI, or diffusion MRI, may be incorporated when needed to guide individualized surgical plans and resection goals. (C) Neuronavigation and Microscopy: In the operating room, patient-to-image registration aligns the patient's anatomy with the preoperative imaging, establishing a common coordinate system for neuronavigation. The incision and craniotomy are planned and executed to provide surgical access to the lesion, while neuronavigation and intraoperative microscopy provide continuous image-guided visualization throughout the resection. (D) Intraoperative Ultrasound: Ultrasound is employed at critical points: predural to confirm tumor access, postdural to delineate margins and assess brain shift, and at intervals during resection to monitor for residual tumor. (E) Intraoperative MRI: Intraoperative MRI assesses the extent of resection and brain shift. If residual tumor is present, additional targeted resection is guided by updated MRI data alongside ultrasound and microscopy, before closure. (F) Postoperative Recovery: Patients undergo immediate postoperative clinical care and imaging to evaluate the final extent of resection and screen for postoperative complications. (G) Pathology, Adjuvant Therapies and Serial Imaging: Pathology informs adjuvant therapy decisions, and serial MRI with ongoing clinical follow-up monitors for recurrence and optimizes long-term management.
• Figure 2: Comparison between classic instance optimization and deep learning-based medical image registration frameworks. (A) Classic methods optimize the transformation parameters independently for each image pair by minimizing a similarity-based objective function through iterative optimization (Section \ref{['sec:classic_instance_optimization']}). (B) Learning-based approaches train a neural network on a dataset of image pairs to predict the deformation field directly in a single forward pass, minimizing both similarity ($L_{sim}$) and regularization ($L_{reg}$) losses during training (Section \ref{['sec:learning_based_registration']}). Blue arrows show the inputs to the objective function that guides backpropagation.
• Figure 3: Diagram illustrating the literature search process. (Top) A total of $712$ records were first identified through database searches in PubMed, IEEE Xplore, Scopus, and Web of Science. Duplicates were removed before screening. Titles and abstracts were reviewed to exclude irrelevant or ineligible studies, followed by full-text assessment based on predefined inclusion criteria. (Bottom) The final selection included $41$ studies categorized by their methodological approach, including classic instance optimization, biomechanical modeling, and deep learning-based methods.
• Figure 4: Summary of key trends across studies meeting inclusion criteria. (A) Number of publications per year from January 2020 to *April 2025, comparing all surveyed methods with those exclusively based on deep learning (DL) for image registration. A steady growth trend is observed, with DL approaches dominating recent contributions. Additional papers from the Learn2Reg ReMIND2Reg challenge (2024 and 2025) remain unpublished at this time and were therefore not included in the present review. (B) Methodological distribution among studies performing image registration. Deep learning-based registration represents the majority, followed by biomechanical modeling and hybrid strategies combining learning with instance optimization (IO) or physics-informed priors. (C) Registered modality pairs and datasets used in DL-based registration studies. Most methods performed MRI-MRI registration, followed by MRI-US. The BraTS-Reg dataset is the most widely adopted, while several works also employed combined RESECT and BITE data or private clinical datasets. For a more detailed description of the different datasets, please refer to Table \ref{['tab:datasets_summary']}.
• Figure 5: Geographic distribution of the reviewed publications. Darker shading indicates higher publication counts. Research output is dominated by the United States, followed by Germany and China, with additional contributions from Canada and other countries.