Image segmentation is one of the key analysis tools in biomedical imaging applications. Although level set segmentation algorithms have been explored thoroughly in the past, these approaches are non-scalable due to their inherent data dependencies. Given that the resolution of medical imaging hardware has continued to increase each year and CPU performance has not kept pace, we need to explore parallel solutions for processing today’s medical images.
Our prior work presented an efficient level set segmentation algorithm designed for parallel architectures for segmenting 2-D images. The algorithm segments an input image based on an initial curve that includes the following components: the member of the shape, member of the outside border of the shape, member of the inside border of the shape, not a member of the shape.
We extend our prior 2-D level set segmentation algorithms, providing a solution for 3-D images. We modify the algorithm by examining adjacent voxels at each step, instead of adjacent pixels. The initial curve is a user-provided sphere that is defined parametrically to reduce copy overhead to the compute device. The efficiency of the 2-D algorithm is preserved in the conversion, enabling the resulting algorithm to perform on the order of 10 times faster than existing GPU-accelerated 3-D level set segmentation implementations.
The implementation presented in this work supports real-time segmentation of 7T MRI images by leveraging the compute power of an NVIDIA Kepler K20 GPU. Our image segmentation algorithm supports identification of tumors, tissue volume measurements, and surgery planning at video rates.