Description
(2) Material and Methods
A fast 3D SAFT implementation using multiple GPUs is presented. It is shown that for volumes of high resolution a speed of sound correction is necessary to overcome the defocusing by assuming constant speed of sound. Speed of sound and attenuation correction was incorporated in the GPU-based SAFT reconstruction, but leads also to a significant reduction of computational performance. An approximation to SAFT, i.e. a time of flight interpolation based GPU implementation (TOFI-SAFT), accelerates our previous GPU implementation of speed of sound corrected SAFT to allow reconstruction of speed of sound and attenuation corrected SAFT images as fast as non-corrected SAFT.
(4) Discussion and Conclusion
The GPU based 3D SAFT implementations show, especially with TOFI-SAFT on high performance GPUs, that high resolution 3D reflection tomography can be done in a clinically relevant reconstruction time in the order of minutes, even including speed of sound and attenuation correction.
(3) Results
TOFI-SAFT achieves a maximum performance of 104 GVA/s on a GPU server with eight GTX 590, which is a speed up of factor 3 compared to the corrected SAFT and approaches the maximum performance of 106 GVA/s of the uncorrected SAFT reconstruction. Tested on one new generation GPU, GTX Titan, TOFI-SAFT can be even faster than uncorrected SAFT: a GPU server of eight GTX Titans would result in a maximum performance of 210 GVA/s for uncorrected SAFT, 62 GVA/s for corrected SAFT and 442 GVA/s for TOFI-SAFT which is a speed up of more than factor 7. Comparing speed of sound corrected SAFT and TOFI-SAFT for clinical data resulted in only small degradation of the image quality, see figure.
