KR20190066922A - Processing System and method for optical coherence tomography - Google Patents

Abstract

The present invention relates to a high-speed optical coherence tomography (OCT) system which provides a CPU-GPU parallel processing concept in an A-scan data unit so as to improve the processing speed of the OCT system.

Description

[0001] The present invention relates to an optical coherence tomography system,

The present invention relates to a high-speed optical coherence tomography (OCT) system, and more particularly, to a method for improving the processing speed of an OCT system through CPU-GPU parallel processing.

The optical coherence tomography (OCT) system is an optical imaging device that has been developing since its inception in the early 1990s. It has a high resolution of micrometer (㎛) sectional image and a three-dimensional volume image.

The OCT system basically acquires signal data in the depth direction of the sample through an optical beam, and this is called an A-scan (axial scan) data.

To obtain a two-dimensional tomographic image, the OCT system sequentially moves the optical beam in one axial direction through a Galvano mirror or MEMs and sequentially acquires the A-scan data.

In order to acquire a 3D volume image, the OCT system moves the 2D tomographic image in one axial direction (axial direction orthogonal to the axis for acquiring the 2D tomographic image) and successively acquires the 2D tomographic image, , And generally acquires a three-dimensional volume image by moving the optical beam in a raster pattern (eg, TV scanning).

1 shows a two-dimensional tomographic image and a three-dimensional volume image of a sample (eye cornea).

The OCT system has recently been greatly developed due to the improvement of the performance (speed) of the data acquisition device, and it is expected that the high-speed / high-resolution OCT system can be realized because the volume data of the sample can be acquired in real time.

However, data processing, 2D cross-sectional image, and 3D volume image generation are still being performed in the post-processing stage because they are relatively slow compared to the data acquisition (acquisition) speed, and thus the range of application fields is limited.

In order to improve the data processing and the speed of 2-D / 3-D image generation, methods using FPGA (field programmable gate array), multi-core CPU and DSP (digital signal processing) In recent years, the development of graphics cards has been actively researched and developed using GPU (graphics processing unit).

The GPU-based OCT system basically acquires the data from the CPU, acquires the 2-dimensional tomographic image through the data processing in the GPU, and acquires the single tomographic image is the sum of the data acquisition time and the data processing time.

GPU-based methods show improved performance in terms of processing speed compared to existing methods, but require additional processing time not related to data processing, such as GPU-to-CPU data transfer rates, GPU-to-GPU data bandwidth). Therefore, a technique for improving the data transfer speed between the GPU and the CPU is required in the OCT system using the GPU.

Related Prior Art Patent No. 1409234 (Optical coherence tomography apparatus capable of parallel processing of interference signals) is for improving data processing speed in FD-OCT including SS-OCT and SS-OCT, The present invention is characterized in that the obtained data is grouped into a plurality of data sets and the plurality of data sets are processed in parallel. In order to improve the efficiency of using the GPU, .

The present invention provides an OCT system and method using a GPU capable of acquiring a 2D sectional image and a 3D volume image.

Also, OCT system and method which can reduce the processing speed compared with existing methods through CPU-GPU parallel processing and acquire 3-D volume image in real time in acquisition of 2D sectional image I want to.

The present invention relates to an optical coherence tomography (OCT) system for generating a two-dimensional tomographic image or a three-dimensional volume image on the basis of A-Scan data by a central processing unit (CPU) and a graphics processing unit , The CPU is a multi-core CPU and includes a plurality of buffers into which A-scan data is switched and input; A pinned memory coupled to each of the buffers; An input processor for performing a parallel switching process in which A-scan data for a two-dimensional tomographic image is sequentially input to the plurality of buffers and then moved to the fixed memory; And a data transfer unit for transferring the A-Scan data moved to the fixed memory to the GPU memory, wherein the GPU includes an image generation unit for generating image data for A-Scan data stored in the GPU memory, A signal processing unit for performing an operation; And an image transmission unit for transmitting the image to the CPU when the data processing for the two-dimensional tomographic image is completed by the signal processing unit. And a plurality of A-scan data are processed in parallel by the operations of the input processor, the data transmitter, the signal processor, and the image transmitter.

According to an aspect of the present invention, the GPU includes a rendering processing unit for accessing a set of two-dimensional images stored in the GPU memory generated by the signal processing unit to visualize a three-dimensional volume image.

According to another aspect of the present invention, there is provided an optical coherence tomography (OCT) system that generates a two-dimensional tomographic image or a three-dimensional volume image based on A-scan data by a central processing unit (CPU) and a graphics processing unit (1) acquiring A-Scan data from a first buffer by a multi-core CPU; 2) A multi-core CPU acquires the next A-Scan data in the second buffer and moves the A-Scan data of the first buffer to a first pinned memory connected thereto ; 3) A multi-core CPU acquires the next A-Scan data in the first buffer and concurrently acquires A-Scan data of the second buffer in a second fixed memory connected thereto Transferring the A-Scan data of the first fixed memory to the GPU memory; 4) A multi-core CPU acquires the next A-Scan data in the second buffer and concurrently acquires A-Scan data of the first buffer in a first fixed memory connected thereto And transmits the A-Scan data of the second fixed memory to the GPU memory, and the signal processing unit of the GPU executes an image generation operation on the A-Scan data previously stored in the GPU memory; 5) A multi-core CPU acquires the next A-Scan data in the first buffer and concurrently acquires A-Scan data of the second buffer in a second fixed memory connected thereto And transmitting the A-Scan data of the first fixed memory to the GPU memory, wherein the signal processing unit of the GPU executes an image generation operation for the A-scan data previously stored in the GPU memory; And 6) repeating the steps 4 and 5 to transmit the image to the CPU when the data processing of the 2-dimensional tomographic image by the GPU is completed.

According to an aspect of the present invention, the method further includes a rendering step of visualizing a three-dimensional volume image by accessing a set of two-dimensional images stored in the GPU memory after data processing is completed by the step 6).

According to the present invention, it is possible to dramatically increase the number of 2-dimensional tomographic images (FPSs) that can be acquired per second as compared with the conventional technology in which only a CPU or a CPU and a GPU are serially processed , Which enables real-time OCT system implementation.

Figure 1 shows a two-dimensional tomographic image and a three-dimensional volume image of a sample (eye cornea).
2 is a block diagram of a system configuration according to the present invention.
3 is a conceptual diagram of a CPU-GPU parallel processing according to the present invention.
4 is a data processing conceptual diagram of a system according to the present invention.
5 is a flowchart of a processing method according to the present invention.
6 is a conceptual diagram of a GPU processing method according to the present invention.

The present invention relates to a method for simultaneously processing data collection and processing through CPU-GPU parallel processing of A-Scan data units to reduce the acquisition time of a tomographic image, and the CPU performs A-scan data collection corresponding to a 2-dimensional tomographic image Once completed, it is transferred to the GPU memory and data acquisition for the next tomographic image is started without waiting for the GPU processing to complete. As a result, the processing time (data collection time) in the CPU or the processing time (data processing time) The performance of the system is determined.

The present invention provides a CPU-GPU parallel processing method for improving OCT system processing speed (two-dimensional sectional image generation speed), wherein parallel processing means that data acquisition and image generation are processed simultaneously, It is possible because it can be regarded as each unique processor.

Hereinafter, technical features of the present invention will be described in detail with reference to the drawings.

2 is a block diagram of an optical coherence tomography (OCT) system according to an embodiment of the present invention. Referring to FIG. 2, a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit) Or an optical coherence tomography (OCT) system for generating a three-dimensional volume image, the CPU 1 includes a buffer 10, a fixed memory 20, an input processing unit 30, and a data transfer unit 40 The GPU 2 includes a GPU memory 50, a signal processing unit 60, an image transmitting unit 70, and a rendering processing unit 80.

In the present invention, a multi-core CPU is used for parallel processing, and the buffer 10 is divided into a plurality of buffers 101 and 102 so that the A-scan data is switched (cross-over) do.

Preferably, the first buffer 101 and the second buffer 102 can be formed, and the number of buffers can be increased according to the required processing capacity.

A pinned memory 20 is applied to reduce data transfer time and is connected to each of the buffers 101 and 102 in the same manner as the first fixed memory 201 and the second fixed memory 202.

The input processing unit 30 is a plurality of cores for performing a parallel switching process in which A-scan data for a two-dimensional tomographic image is sequentially inputted to the plurality of buffers 101 and 102 and then moved to the fixed memories 201 and 202 .

The data transfer unit 40 transfers the A-scan data sequentially and sequentially transferred to the fixed memories 201 and 202 to the GPU memory 50. The data transfer unit 40 transfers the A- It is possible.

The signal processing unit 60 performs an image creation operation on the A-scan data transmitted by the data transfer unit 40 and stored in the GPU memory 50, and stores the image data in the GPU memory 50.

The image transfer unit 70 determines whether the data processing operation of the A-scan data for the two-dimensional tomographic image is completed, and transmits the two-dimensional tomographic image to the CPU.

As described above, a plurality of A-scan data are processed in parallel by the operations of the input processing unit 30, the data transfer unit 40, the signal processing unit 60, and the image transfer unit 70.

Figure 3 shows a flow chart for CPU-GPU parallel processing.

While the CPU acquires signal data (interference data) for a single two-dimensional tomographic image, the GPU processes the acquired data to generate a two-dimensional cross-sectional image, which results in a two-dimensional It is possible to increase the number of cross-sectional image (frame) acquisitions (frames per second (FPS)).

The performance of the entire system is determined by the data acquisition and processing time of the present invention, and the performance of the entire system is determined. For example, in FIG. 2, since the data acquisition time is slower than the processing time, .

For example, if data is acquired using a 200 K light source (200,000 A-Scan data acquisitions per second) and a two-dimensional image is generated with 500 A-Scan data, the present invention provides 400 FPS Two-dimensional image generation, 200,000 / 500 = 400) system performance.

4 shows a flow chart from data acquisition to two-dimensional image generation.

The CPU sequentially acquires the A-scan data (interference signal data) from the data acquisition device (digitizer) and transmits it to the GPU memory when the data acquisition for the two-dimensional image is completed.

In the present invention, a pinned memory (fixed memory) is used to reduce the CPU-GPU data transfer time. To this end, a process of moving data acquired from the digitizer stored in the buffer to the pinned memory is added.

To minimize the time required to move from the buffer to the pinned memory, parallel processing is performed using a multi-core CPU. For parallel processing, two buffers are allocated as shown in FIG. 3, and switching between Buffer 1 and Buffer 2 is performed. And collects signal data from the digitizer.

Collected data moves to pinned memory while collecting data from other buffers. For example, after acquiring a signal from Buffer 1, acquire the signal from Buffer 2, and while Buffer 2 acquires the signal, move the signal data of Buffer 1 to Pinned Memory 1.

The GPU generates a two-dimensional image by performing a signal processing algorithm on the A-scan unit from the data received from the pinned memory. Since the GPU internally supports parallel processing, it is possible to process multiple A-scan data simultaneously, Thereby reducing the execution time in image generation.

The signal processing algorithm for one A-scan data consists of Rescaling, Fast Fourier transform (FFT), and Log-scaling compression.

The GPU actually takes a significant amount of time to transfer data between the CPU and the GPU.

In order to solve this problem, the present invention concurrency of data transmission and processing within the GPU.

Data transmission and processing in the GPU is largely composed of three steps: data transfer from the CPU to the GPU, data processing in the GPU, and data transfer from the GPU to the CPU. In the parallel processing, data obtained from the CPU is divided, do.

When a three-dimensional image is required, the GPU further accesses a set of two-dimensional images stored in the GPU memory created by the signal processing unit using the rendering processing unit 80 to visualize a three-dimensional volume image. For this purpose, a Volume Rendering algorithm .

That is, the sequentially acquired two-dimensional image sets (for example, three-dimensional volume data obtained by the raster pattern) generated through the signal processing algorithm are stored in the GPU memory, and the rendering processing unit 80 supplies the GPU memory And visualizes the 3D volume image.

The volume rendering algorithm is based on Ray casting and texture slicing.

5 is a flowchart of an optical coherence tomography (OCT) processing method according to an embodiment of the present invention. Referring to FIG. 5, a first buffer 101, a second buffer 102, the first fixed memory 201, and the second fixed memory 202. [

FIG. 6 is a conceptual view of a parallel processing performed in the GPU. FIG. 6 shows an example of performing parallel processing by dividing data acquired by a CPU by 1/4. The first data (first 1/4 data) (A) and transmits the second data (second 1/4 data) to the GPU while processing the transmitted data (b). Data that has been processed is transferred (c) to the CPU again.

6, the total processing time (d) in the GPU is similar to the sum of the data transfer time (e) from the CPU to the GPU and the data transfer time (f) from the GPU to the CPU, In conclusion, the total processing time in the GPU is determined by the data transfer speed between the CPU and GPU rather than the data processing (image generation) time.

Therefore, the processing process is summarized with reference to FIG. 5 as follows.

In the first buffer data obtaining step s10, which is the first step, the multi-core CPU obtains the A-scan data in the first buffer 101. [

In step 2, the multi-core CPU acquires the next A-scan data in the second buffer (s201) and simultaneously transmits the A-Scan data of the first buffer to the first fixed memory (S202).

In step 3, the multi-core CPU acquires the next A-scan data in the first buffer (s301), and concurrently acquires the A-scan data of the second buffer in the second fixed memory (S302), and transmits the A-Scan data of the first fixed memory to the GPU memory (s303).

In step 4, the multi-core CPU acquires the next A-scan data in the second buffer (s401), and concurrently acquires the A-scan data of the first buffer in the first fixed memory Scan data of the second fixed memory to the GPU memory (s403), and the signal processing unit of the GPU transfers the A-Scan data of the second fixed memory to the pinned memory An image creation job is executed (s404).

In step 5, the multi-core CPU acquires the next A-scan data in the first buffer (s501), and concurrently acquires the A-scan data of the second buffer in the second fixed memory (S502), the A-Scan data of the first fixed memory is transferred to the GPU memory (s503), and the signal processing unit of the GPU transfers the A-Scan data of the A-Scan data stored in the GPU memory to the pinned memory An image creation job is executed (s504).

In step S6, it is determined whether or not the data processing of the two-dimensional tomographic image by the GPU is completed while the steps 4 and 5 are repeatedly performed. If completed, the corresponding image is transmitted to the CPU in operation s60.

If a three-dimensional image is additionally required, the GPU accesses the set of two-dimensional images stored in the GPU memory after the data processing is completed by the step 6 using the rendering processor 80, (S60) for visualizing the rendering process (step s60).

10: buffer 20: fixed memory
30: input processor 40:
50: GPU memory 60: signal processor
70: image transmitting unit 80: rendering processing unit

Claims (4)

An optical coherence tomography (OCT) system for generating a two-dimensional tomographic image or a three-dimensional volume image based on an A-scan data by a central processing unit (CPU) and a graphics processing unit (GPU)
The CPU is a multi-core CPU,
A plurality of buffers into which A-Scan data is switched and input;
A pinned memory coupled to each of the buffers;
An input processor for performing a parallel switching process in which A-scan data for a two-dimensional tomographic image is sequentially input to the plurality of buffers and then moved to the fixed memory; And
And a data transfer unit for transferring the A-Scan data moved to the fixed memory to the GPU memory,
The GPU includes:
A signal processor for performing an image generation operation on the A-scan data transmitted by the data transmitter and stored in the GPU memory; And
An image transmitting unit for transmitting the image to the CPU when the data processing of the two-dimensional tomographic image is completed by the signal processing unit; Including,
Wherein the plurality of A-scan data are processed in parallel by the operations of the input processor, the data transmitter, the signal processor, and the image transmitter. The method according to claim 1,
The GPU includes:
And a rendering processor for accessing a set of two-dimensional images stored in the GPU memory generated by the signal processor to visualize a three-dimensional volume image. 1. An optical coherence tomography (OCT) processing method for generating a two-dimensional tomographic image or a three-dimensional volume image based on an A-scan data by a central processing unit (CPU) and a graphics processing unit (GPU)
1) A multi-core CPU acquires A-Scan data with a first buffer;
2) A multi-core CPU acquires the next A-Scan data in the second buffer and moves the A-Scan data of the first buffer to a first pinned memory connected thereto ;
3) A multi-core CPU acquires the next A-Scan data in the first buffer and concurrently acquires A-Scan data of the second buffer in a second fixed memory connected thereto Transferring the A-Scan data of the first fixed memory to the GPU memory;
4) A multi-core CPU acquires the next A-Scan data in the second buffer and concurrently acquires A-Scan data of the first buffer in a first fixed memory connected thereto And transmits the A-Scan data of the second fixed memory to the GPU memory, and the signal processing unit of the GPU executes an image generation operation on the A-Scan data previously stored in the GPU memory;
5) A multi-core CPU acquires the next A-Scan data in the first buffer and concurrently acquires A-Scan data of the second buffer in a second fixed memory connected thereto And transmitting the A-Scan data of the first fixed memory to the GPU memory, wherein the signal processing unit of the GPU executes an image generation operation for the A-scan data previously stored in the GPU memory; And
6) Repeating the steps 4) and 5) to transmit the image to the CPU when the data processing of the 2-dimensional tomographic image by the GPU is completed. The method of claim 3,
And a rendering step of accessing a set of two-dimensional images stored in the GPU memory by visualizing the three-dimensional volume image after the data processing is completed by the step 6) after the step 6) A tomographic image processing method.

Images