Predicting patient-specific brain shift by generating intraoperative brain MRI during neurosurgical interventions based on preoperative brain MRI.

Could a data-driven model be used to predict a patient-specific updated MRI image representative of changes in the brain caused by brain shift either via image synthesis or a deformation field?

Objectives:
Y1: Dataset generation and brain deformation model design (direct image and vector field)
Y2: Extend prediction model with physics AI (Physics Informed Neural Network - PINNs)
Y3: Real-time brain shift prediction

Methodology:
Y1. Build synthetic dataset: 1) 3D printing plastic model, whereby deformations will be applied, and recorded using medical imaging (at rest and deformed states), 2) by generating a post-op MRI from pre-op MRI and post-op CT of Deep Brain Stimulation (DBS) cases (KCL), and 3) by generating synthetic post-op MRI conditioned by pre-op MRI via latent diffusion models (UCL).

Y1-2. Design and implement data-driven deep learning (DL) methods to predict brain shift: 1) post-op MRI directly from pre-op MRI via latent diffusion models, 2) post-op MRI resulting from learning the deformation vector field via PCA encoder/decoder and guided supervision, and 3) extending (1) and (2) via model optimization when applying PINNs which aims to provide prior knowledge by learning governing physical equations.

Y3. Real-time brain shift prediction. By capturing intraoperative information with equipment such as O-arm (CT), Kinevo 900 microscope, intraoperative ultrasound and/or laser scanner that will serve as inputs to the prediction models that will be tailored for this task.