Deep Learning Real-Time Adaptive Physics-based Non-Rigid Registration for Accurate Geometry Representation of Brain in Modeling Deformation During Glioma Resection
Description/Abstract/Artist Statement
The Physics-based Non-Rigid Registration (PBNRR) framework allows for accurate real-time medical image registration and geometry representation of the brain in modeling deformation during glioma resection. Existing adaptive PBNRR (APBNRR) shows promise in being able to be utilized in time-constrained image-guided neurosurgery operations, but the issue of determining patient-specific input parameters to allow for optimal registration remains an open problem. We present a deep feedforward neural network that can predict sets of possible optimal or suboptimal input parameters that lead to a low Hausdorff distance of the registered image from the preoperative image. The neural network is trained on output produced by over 2.6 million retrospective APBNRR executions consisting of an almost exhaustive parameter study using cloud computing on 13 patient cases spanning from partial to excessive tumor resection. By utilizing the neural network, we can greatly reduce the parameter space that needs to be evaluated with APBNRR in order to achieve optimal results, and initial experiments have been very promising.
Faculty Advisor/Mentor
Dr. Nikos Chrisochoides
Presentation Type
Oral Presentation
Disciplines
Computer Sciences | Numerical Analysis and Scientific Computing | Theory and Algorithms
Session Title
Chemistry
Location
Learning Commons @ Perry Library Conference Room 1310
Start Date
2-2-2019 11:30 AM
End Date
2-2-2019 12:30 PM
Deep Learning Real-Time Adaptive Physics-based Non-Rigid Registration for Accurate Geometry Representation of Brain in Modeling Deformation During Glioma Resection
Learning Commons @ Perry Library Conference Room 1310
The Physics-based Non-Rigid Registration (PBNRR) framework allows for accurate real-time medical image registration and geometry representation of the brain in modeling deformation during glioma resection. Existing adaptive PBNRR (APBNRR) shows promise in being able to be utilized in time-constrained image-guided neurosurgery operations, but the issue of determining patient-specific input parameters to allow for optimal registration remains an open problem. We present a deep feedforward neural network that can predict sets of possible optimal or suboptimal input parameters that lead to a low Hausdorff distance of the registered image from the preoperative image. The neural network is trained on output produced by over 2.6 million retrospective APBNRR executions consisting of an almost exhaustive parameter study using cloud computing on 13 patient cases spanning from partial to excessive tumor resection. By utilizing the neural network, we can greatly reduce the parameter space that needs to be evaluated with APBNRR in order to achieve optimal results, and initial experiments have been very promising.