The goal of the simulation is to predict the size of the RFA lesion based on the patient 3D model, RFA needle position and a simulated RFA protocol.
Our simulation is based on the Penne's bioheat model, whereas finite element method is used for numerical discretization. The major advantage of our simulation software is the speed up and the accuracy.
The fast RFA simulation software uses the advanced heterogeneous GPU computing. The CPU part receives the data such as mesh, needle position and patient-specific parameters. The CPU reads the mesh, finds the neighbours and generates a CSR (compressed Sparse Row) matrix. The GPU receives the required data from the CPU and computes the following:
- Heat source or Point source values (each thread → each node)
- Shape function and Assembly matrix (each thread → each element)
- Solution of the linear system (CUBLAS and CUSPARSE library)
- Prediction of the cell death values (each thread → each node)
Our optimized code avoids the explicit numerical integration while computing the stiffness matrix and it assembles the stiffness matrix to the global matrix with the help of an efficient neighbour mapping algorithm.
Once the GPU computes and sends the data back to the CPU, the CPU computes the thermocouple values and loops again through GPU-CPU cycle until a certain criteria to stop the simulation becomes true. Also, the GPU saves the lesion values in a vtp file at a regular interval.