A carregar...
Publicação
Physics-Guided Deep Learning for Sparse Data-Driven Brain Shift Registration
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 30.61 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Accurate neuronavigation in brain tumor surgery relies on the alignment between pre-operative imaging and the patient’s intra-operative anatomy. This alignment is severely compromised by brain shift, a complex and non-linear deformation caused by several factors, including gravity, cerebrospinal fluid loss, and tissue resection. Existing image registration approaches often require access to dense and high-quality intra-operative data or computationally expensive biomechanical simulations, limiting their practicality in real-time surgical workflows. Registration methods relying on sparse keypoints offer a promising alternative but typically rely on simple geometric interpolators that ignore the biomechanical behavior of brain tissue to estimate dense displacement fields, resulting in physically implausible deformations. This thesis proposes a physics-guided deep learning framework that acts as an advanced interpolator of sparse displacement information for brain shift correction. The method integrates data-driven learning with biomechanical modeling by training a residual 3D U-Net-based network to refine initial displacement fields derived from sparse matched keypoints. Supervision is provided by a large-scale dataset of deformations generated through patient-specific biomechanical simulations, ensuring that the learned deformations are biomechanically consistent. The proposed network is designed to be modular, efficient, and compatible with existing keypoint-based registration pipelines that require minimal intra-operative data. Experimental results evaluated the simulated intra-operative brain deformations and the refinements estimated by the proposed network, demonstrating that the proposed approach significantly outperforms standard interpolation techniques, such as linear and thin-plate spline methods, in terms of displacement accuracy and physical plausibility. The network refines these initial interpolations by incorporating learned physical priors, enabling the generation of biomechanically plausible deformations within an image registration pipeline. These improvements are achieved with negligible computational overhead, supporting real-time applicability. Overall, the work detailed in this thesis advances the state of the art in keypoint-based data-driven brain registration by bridging the gap between biomechanical realism and deep learning efficiency.
Descrição
Tese de mestrado, Ciência de Dados, 2026, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
biomechanical modeling brain shift medical image registration physics-informed deep learning sparse keypoint interpolation