Medical imaging departments in hospitals still scarcely use the multiview imaging procedure, known as diffraction tomography. During the HPCN'99 conference track on medical applications, Dr. Theofanis Maniatis from the Technical University of Athens, proved that this method, allowing to reconstruct cross-sections of an object from scattered field measurements, can be successfully implemented in a high performance computing (HPC) environment. The HPC platform enables to solve the time-consuming inverse scattering problem which consists of finding the object's shape and refractive index. The algorithm parallelization makes it possible to speed up both the integrals' calculation and the cost function partial derivatives with respect to the unknowns in the non-linear optimization method. Fast execution of the inverse scattering algorithm turns diffraction tomography into a feasible and clinically realistic medical imaging application.
Within diffraction tomography, the biological object is illuminated from many incident directions by a diffracting energy source, like ultrasound or microwaves. The interaction between the incident field and the scattering object has to be described in terms of the wave equation. Currently the job is done with non-perturbative and non-linear optimization techniques, because this approach has no theoretical restriction regarding the scattering object's size or refractive index. Since the numerical solution of the inverse scattering problem is computationally intensive, the research team of Dr. Maniatis has ported the algorithm on an HPC platform. The idea is to interconnect a range of different clinical sites with an HPC centre to offer these hospitals a chance to introduce real life diffraction tomography sessions.
Hospitals normally do not dispose of the HPC facilities, necessary to perform intense and iterative calculations. As a result, the need to establish powerful communication lines between the HPC site and the participating diagnostic centres is justified, in order to enable the required data exchange and remote execution of the time-consuming parts of the algorithm. The multiple use of the HPC facilities reduces the financial cost of the implementation. In fact, a whole range of diffraction tomography scanners interact with a single inverse scattering solver at the HPC centre. The required visual data is collected by each diffraction tomography device and included in a message package to be sent to the inverse scattering solver with sockets over the TCP/IP protocol. At the HPC site, the algorithm is run after which the reconstructed images are sent back to the different hospitals.
The inverse scattering solver is implemented as a series of processes running on a parallel computing system. Message Passing Interface (MPI) is used for the distributed programming systems to monitor the inter-process channels of communication. If available, it is also possible to apply the shared memory of the system. The clinical sites, located at the hospitals, only need relatively simple workstations to interconnect with the HPC centre. In the near future, the team of Dr. Maniatis plans to port the current implementation to another parallel programming platform. Therefore, a new implementation needs to be designed which is based on the use of MPI as the communication protocol of the distributed processing memory. This type of solution has the advantage that it can be used on various platforms. More research details are provided on the Web site of the National Technical University of Athens.