Method and device for classifying vascular objects using classification trees evaluated on a graphics processing unit
Methods and devices for optimizing classification tree evaluation for online and real-time characterization of tissue by using GPUs are disclosed that give a significant improvement in run time compared to data decomposition. The methods and devices use a parallel algorithm designed for single instruction, multiple data (SIMD) architectures commonly found in GPUs. Various optimizations and their effects are also disclosed.
1. A method of identifying components within a vascular object from a signal, the method comprising:
transmitting a signal within a selected region of the vascular object and collecting a backscatter signal reflected from the vascular object;
reconstructing an image of the selected region from the backscatter signal;
preparing a histology image of the selected region of the vascular object, the histology image distinguishing different components of the vascular object;
selecting a region of interest from the histology image, the region of interest representing one component from the different components of the vascular object and being at a coordinate location on the histology image;
mapping the coordinate location to a location on the image corresponding to the coordinate location;
translating the location on the image to a signal portion in the backscatter signal corresponding to the location, the signal portion representing a signal signature of the one component of the vascular object; and
correlating the signal signature to the one component using a classification tree scheme using a graphic processing unit (GPU), wherein using the classification tree scheme includes performing data decomposition of a sample of the signal portion by evaluating in parallel all nodes in a tree associated with the sample of the signal portion and thereafter synchronously reducing a path through the tree, the path being defined by the evaluation of all nodes in the tree, wherein evaluating all nodes in the tree includes implementing branchless tree traversal, wherein reducing the path through the tree includes providing a static path for each of a subset of all nodes in the tree, and wherein reducing the path through the tree includes reducing the path a plurality of times per computational loop, wherein the correlating further comprises generating a processor-node map that correlates a subdivision of the GPU with an index of a node of the tree associated with the sample of the signal portion.
2. The method of claim 1 wherein the signal is an intravascular ultrasound signal.
3. The method of claim 1 wherein the signal is an optical coherence tomography signal.
4. The method of claim 2 , wherein the intravascular ultrasound signal is transmitted from at least one of: a single transducer element and an array of ultrasound transducers.
5. The method of claim 1 , wherein the GPU comprises a single instruction, multiple data (SIMD) architecture.
6. The method of claim 1 , wherein the GPU comprises a plurality of streaming multiprocessors.
7. A device for identifying components within a vascular object from a signal comprising:
a transducer for transmitting a signal within a selected region of the vascular object and collecting a backscatter signal reflected from the vascular object; and
a computing system including a graphics processing unit (GPU), the computing system configured to:
reconstruct an image of the selected region from the backscatter signal;
prepare a histology image of the selected region of the vascular object, the histology image distinguishing different components of the vascular object;
select a region of interest from the histology image, the region of interest representing one component from the different components of the vascular object and being at a coordinate location on the histology image;
map the coordinate location to a location on the image corresponding to the coordinate location;
translate the location on the image to a signal portion in the backscatter signal corresponding to the location, the signal portion representing a signal signature of the one component of the vascular object; and
correlate the signal signature to the one component using a classification tree scheme using the GPU, wherein the computing system is further configured to use the GPU to perform data decomposition of a sample of the signal portion by evaluating in parallel all nodes in a tree associated with the sample of the signal portion with branchless tree traversal and thereafter synchronously reducing a path through the tree, the path being defined by the evaluation of all nodes in the tree.
8. The device of claim 7 , wherein the transducer comprises at least one of: a single transducer element and an array of ultrasound transducers.
9. The device of claim 7 , wherein the GPU comprises a single instruction, multiple data (SIMD) architecture.
10. The device of claim 7 , wherein the computing system is configured to use the classification tree scheme to implement a static path through the tree associated with the sample of the signal portion for each of a subset of all nodes in the tree.
11. The device of claim 7 , wherein the computing system is configured to use the classification tree scheme to implement a processor-node map that correlates a subdivision of the GPU with an index of a node of the tree associated with the sample of the signal portion.
12. The device of claim 7 , wherein the computing system is configured to use the classification tree scheme to reduce a path through the tree multiple times per computational loop.
13. The device of claim 7 , wherein the GPU comprises a plurality of streaming multiprocessors.
14. A method of identifying components within a vascular object from a signal, the method comprising:
transmitting a signal within a selected region of the vascular object and collecting a backscatter signal reflected from the vascular object;
reconstructing an image of the selected region from the backscatter signal;
preparing a histology image of the selected region of the vascular object, the histology image distinguishing different components of the vascular object;
selecting a region of interest from the histology image, the region of interest representing one component from the different components of the vascular object and being at a coordinate location on the histology image;
mapping the coordinate location to a location on the image corresponding to the coordinate location;
translating the location on the image to a signal portion in the backscatter signal corresponding to the location, the signal portion representing a signal signature of the one component of the vascular object; and
correlating the signal signature to the one component using a classification tree scheme using a graphic processing unit (GPU), wherein using the classification tree scheme includes evaluating in parallel all nodes in a tree associated with a sample of the signal portion and thereafter synchronously reducing a path through the tree, the path being defined by the evaluation of all nodes in the tree.
15. The method of claim 14 , wherein using the classification tree scheme includes implementing branchless tree traversal for the tree associated with the sample of the signal portion.
16. The method of claim 14 , wherein using the classification tree scheme includes providing a static path through the tree associated with the sample of the signal portion for each of a subset of all nodes in the tree.
17. The method of claim 14 , wherein using the classification tree scheme includes generating a processor-node map that correlates a subdivision of the GPU with an index of a node of the tree associated with the sample of the signal portion.
18. The method of claim 14 , wherein using the classification tree scheme includes reducing a path through the tree multiple times per computational loop.