CN112085830A - Optical coherent angiography imaging method based on machine learning - Google Patents

Abstract

The invention discloses an optical coherent angiography imaging method based on machine learning. According to the invention, an original data set required by network model training is generated by utilizing OCT three-dimensional structural images of a sample acquired by OCTA equipment, a whole group of OCT structural images with poor registration effect are removed, an OCTA algorithm is adopted for radiography and imaging to generate a training data set, a machine learning network model is established and trained, so that OCTA radiography is carried out through the machine learning network model; the method can play a great role in the field of OCTA, can generate an angiogram with higher signal-to-noise ratio and better vascular connectivity, and inhibits the common speckle effect in OCT images to a great extent; the label image is automatically generated by an algorithm, so that the applicability of the method is expanded without being influenced by system errors caused by different systems; the damage can be reduced by using smaller detection power for imaging, or the data volume required by imaging is reduced during imaging, and scanning can be completed more quickly.

Description

A Machine Learning-Based Optical Coherence Angiography Imaging Method

technical field

The invention relates to an optical coherence angiography imaging technology, in particular to an optical coherence angiography imaging method based on machine learning.

Background technique

Optical Coherent Tomography (OCT) is a high-resolution, non-contact and fast three-dimensional imaging technology. It utilizes the coherence principle of scattered light in biological tissues, and its signal contrast is derived from the difference in light scattering ability of different biological tissues. OCT technology combines semiconductor and ultrafast laser technology, and uses core components such as broadband light source, Michelson interferometer and photodetector to obtain the backscattered signal of biological tissue, and finally can obtain real-time micron-scale biological tissue through digital signal processing by computer. tomographic image. Therefore, OCT technology has long become one of the important means of anatomical imaging diagnosis. It not only plays an important role in ophthalmology clinical examination, but also plays an important role in fields such as dermatology, gastroenterology, cardiology and neurology. effect.

With the development of science and technology, OCT technology has experienced many major breakthroughs and developments in software and hardware in the past nearly 30 years, with faster imaging speed and higher system sensitivity. Especially after 2002, with the maturity of frequency domain OCT technology, OCT technology has received attention and applications in various fields.

In 1991, Huang et al. of the Massachusetts Institute of Technology built the first OCT prototype with a longitudinal resolution of 15 μm, and published the first isolated human retina OCT scan image and the corresponding tissue slice in the journal Science. , which verifies the feasibility of the OCT system. In 2002, Wojtkowski et al. obtained the world's first live human retinal image based on frequency domain OCT technology. Johannes and Leitgeb successively compared the parameters of frequency domain OCT compared with time domain OCT theoretically and experimentally, and proved that Frequency domain OCT has higher sensitivity and faster imaging speed. Since then, frequency domain OCT has gradually replaced time domain OCT, and has received extensive attention and applications.

Optical coherent angiography (Optical Coherent Tomography Angiography, OCTA) is a new type of non-invasive vascular imaging technology that appeared in recent years. During the specific imaging, the signal light scans the sample through the galvanometer system. The scanning area is generally rectangular and divided into the fast axis direction and the slow axis direction. During scanning, the signal light continuously scans repeatedly in the fast axis direction for many times (usually 4 times). ) to record the OCT signals at the same position at different times, and then remove tissue information through algorithm processing, extract blood flow signals, and generate angiography images. It cleverly uses the flowing red blood cells as a contrast agent, that is, when the red blood cells are continuously flowing in the blood vessels, the OCT signal in the blood vessels is constantly changing, so as to distinguish it from the stable signal of the static tissue.

At present, the imaging algorithms of OCTA are mainly divided into three types of imaging algorithms based on phase changes, amplitude changes, and combined phase and amplitude changes according to the source of blood vessel information. Its essence is to compare the OCT signals at the same location at different times by means of analytical calculation. However, these methods often only utilize a part of the information in the OCT signal, which leads to problems such as low signal-to-noise ratio and serious speckle in angiography images. At present, the main method to solve this problem is to increase the number of scans at the same position and enhance the intensity of the blood vessel signal. This method will lead to too long scanning time, and the vibration of the sample will produce artifacts, such as the shaking of the patient's eyes and breathe. In addition, prolonged laser irradiation can also cause damage to biological tissue.

SUMMARY OF THE INVENTION

In view of the above problems in the prior art, the present invention proposes an optical coherence angiography imaging method based on machine learning.

The optical coherence angiography imaging method based on machine learning of the present invention comprises the following steps:

1) Generate the original dataset:

The OCT three-dimensional structure image of the sample collected by OCTA equipment is used to generate the original data set required for network model training. The original data set includes j×k groups of OCT structure image sequences, each group of OCT structure image sequences includes i two-dimensional horizontal The OCT structure image of the cross-section (B-Scan), where k is the number of samples, j is the number of slow-axis scanning positions of each sample, i is the scanning times of the same slow-axis scanning position of the same sample, i is a natural number > 4, j is a natural number > 50, and k > 5;

2) Data filtering:

The rigid registration algorithm is used to register the i B-Scan surface OCT structure images in the same group of OCT structure images. After registration, the correlation algorithm is used to calculate the registration accuracy, and the whole group of OCT structures with poor registration effect is eliminated. image, retain n groups of filtered OCT structure images, n is a natural number, and

3) Generate a training dataset:

Using the n groups of screened OCT structural images obtained in step 2), the OCTA algorithm is used for contrast imaging, and each group of OCT structural images will obtain an OCTA contrast image of the B-Scan surface, which is called a label image; The image corresponds to the i B-Scan surface OCT structure images of a set of OCT structure images, and m B-Scan surface OCT structure images are taken out, which are called input data and are paired with label images. The input data and label images form the network model training center. The required training data set, where m=2, 3 or 4;

4) Build a machine learning network model:

Build a machine learning network model and set the hyperparameters of the machine learning network model; and divide the training data set into n ₁ sets of training sets and n ₂ sets of test sets, the training sets and test sets are independent of each other, n ₁ and n ₂ are respectively natural numbers, and

5) Train the machine learning network model:

Using the machine learning network model established in step 4), the input data in the training data set is used as the input of the machine learning network model, wherein n ₁ groups of training sets are used for training the machine learning network model, and n ₂ groups of test sets are used for Test the performance of the machine learning network model; during the training process, the training set will be divided into multiple batches and repeatedly input into the machine learning network model for multiple rounds of training, and the difference between the output image of the machine learning network model and the label image will be judged or calculated as The training error is used to train the machine learning network model. After each batch of training, use the test set to test the performance of the machine learning network model. When the performance test indicators of the machine learning network model are stabilized, it is considered that machine learning After the training of the network model is completed, save the trained machine learning network model;

6) Machine learning network model for OCTA imaging:

Using the trained machine learning network model, the OCT structure image of the sample collected by the OCTA device is used as the input, and the output image is the OCTA angiography image.

Wherein, in step 1), when the OCTA device collects the sample, it scans a slow-axis scanning position of a scanning position of a sample once to obtain an OCT structure image of the B-Scan surface, and scans the same slow-axis scanning position. i times, each sample has j slow-axis scanning positions, and there are k samples in total, so as to obtain the OCT structure images of i×j×k B-Scan surfaces, and the OCT structures of i×j×k B-Scan surfaces are obtained. The images are divided into j×k groups of OCT structural images, each group of OCT structural images includes i OCT structural images of the B-Scan surface, that is, each slow-axis scanning position corresponds to a group of OCT structural images, thereby obtaining the original data set.

In step 2), for the registered OCT structure images of the j × k groups, the registration accuracy is compared, and the first n groups of OCT structure images with higher registration accuracy are retained, and the rest are considered to have poor registration effects. These sets of OCT structural images were excluded.

In step 4), the machine learning network model adopts deep convolutional neural network CNN, generative adversarial network GAN or recurrent neural network RNN; hyperparameters include the number of network layers, convolution kernel, learning rate, parameter initialization, number of training rounds and batch size.

In step 5), the mean square error, structural similarity or peak signal-to-noise ratio between the output image and the label image is used as the training error and performance test indicator, using stochastic gradient descent (SGD), adaptive moment One of the estimation optimization algorithm (Adaptive Moment Estimation, Adam) and the momentum algorithm (Momentum) minimizes the training error to train the machine learning network model.

In step 6), the OCT structure image of the sample is scanned multiple times at the same slow-axis scanning position, and the number of scans is ≤4.

Machine learning is an inevitable product of the development of artificial intelligence research to a certain stage, and has become the core research topic of artificial intelligence. Its purpose is to allow computers to acquire knowledge or skills by imitating human learning behaviors and can continuously learn new knowledge to improve performance. Machine learning draws on physiology, psychology, cognition and other disciplines. Through the understanding of the self-learning mechanism of human beings, a computational model or cognitive model similar to human learning is established, thus forming various learning theories and learning methods. And a learning system with specific applications is established for specific tasks.

At present, the commonly used algorithms in machine learning include artificial neural network, support vector machine, naive Bayes, random forest, sparse dictionary, reinforcement learning, representation learning and similarity metric learning. With the development of computer hardware, deep learning has gradually developed, which is a comprehensive evolution of artificial neural networks. Deep learning expands the depth and width of artificial neural networks, and can infinitely approximate more complex nonlinear models, so as to learn the objective laws and internal connections hidden in the data. In a general sense, deep learning algorithms include deep belief networks, deep neural networks and convolutional neural networks, where the structures of deep belief networks and deep neural networks are very similar. The deep learning networks that are currently used in image processing are convolutional neural networks and generative adversarial networks. At present, machine learning has been widely used in the field of medical image reconstruction, enhancement and segmentation, but it has not been used in OCTA image reconstruction.

Advantages of the present invention:

The invention can play a huge role in the field of OCTA, and its powerful data mining ability helps OCTA equipment to generate angiography images with higher signal-to-noise ratio and better blood vessel connectivity, and to a large extent suppresses common occurrences in OCT images. Speckle effect; it is worth mentioning that the label image in the present invention is automatically generated by the algorithm, which is different from common machine learning applications, which need to obtain label data through expert labeling, which expands the applicability of this method without being affected by different The system brings the influence of its own system error. In addition, in the same OCTA equipment, in order to obtain the same level of OCTA imaging, the present invention can use a smaller detection power for imaging, reduce the damage of laser to biological tissue (such as ophthalmology), or reduce the imaging requirements during imaging. The amount of data, that is, the number of scans at the same position is reduced, the scan can be completed faster, and the artifacts (such as patient shaking and breathing during fundus imaging) caused by too long scanning time and sample tremors are reduced.

Description of drawings

FIG. 1 is a flowchart of an optical coherence angiography imaging method based on machine learning of the present invention;

FIG. 2 is an OCTA image obtained by an embodiment of the optical coherence angiography imaging method based on machine learning according to the present invention.

Detailed ways

Below in conjunction with the accompanying drawings, the present invention will be further described through specific embodiments.

The optical coherence angiography imaging method based on machine learning in this embodiment, as shown in FIG. 1 , includes the following steps:

1) Generate the original dataset:

The OCT three-dimensional structure image of the retina collected by OCTA equipment is used to generate the original data set required for network model training. The same slow-axis scanning position of the same sample is scanned 50 times, and the slow-axis scanning position of each sample is 100. For 30 human eyes, the original data set includes 100×30 groups of OCT structure image sequences, and each group of OCT structure image sequences includes 50 B-Scan surface OCT structure images;

2) Data filtering:

The rigid registration algorithm is used to register 50 B-Scan surface OCT structure images in the same group of OCT structure images. After registration, the correlation algorithm is used to calculate the registration accuracy, and the whole group of OCT structures with poor registration effect is eliminated. Images, retain 70%, that is, 100 × 30 × 0.7 groups of OCT structural images after screening;

3) Generate a training dataset:

Using the 100 × 30 × 0.7 groups of screened OCT structural images obtained in step 2), the OCTA algorithm is used for contrast imaging, and each group of OCT structural images will obtain an OCTA contrast image of the B-Scan surface, which is called a label image; From the 50 B-Scan surface OCT structure images of a set of OCT structure images corresponding to each label image, 4 B-Scan surface OCT structure images are taken out, which are called input data and paired with the label image. The input data is composed of the label image. The training dataset required for network model training;

4) Build a machine learning network model:

The machine learning network model adopts the deep convolutional neural network DnCNN, which is composed of 20 convolutional layers, of which the first layer uses 64 3×3×4 convolution kernels to convolve the input 4 OCT structure images to generate 64 images. feature map, and use the ReLU function as the activation function; layers 2 to 19 use 64 3×3×64 convolution kernels to convolve the feature maps of the previous layer, and connect the ReLU function after batch normalization as Output; the 20th layer uses a 64 3×3×4 convolution kernel to convolve the feature map of the previous layer, and the output image is the output image of the network; in the network parameters, the learning rate is set to 0.001, and the network parameters Initialization uses Kaiming initialization method, the number of repeated training rounds is 50, the batch size is 16, and the ratio of training set to test set is 7:3;

5) Train the machine learning network model:

Using the machine learning network model established in step 4), the input data in the training data set is used as the input of the machine learning network model, of which 100 × 30 × 0.7 × 0.7 sets of training sets are used to train the machine learning network model, 100 × 30 × 0.7 ×0.3 sets of test sets are used to test the performance of the machine learning network model; during the training process, the training set will be input into the machine learning network model in batches and repeated for 50 rounds of training, and the difference between the output image of the machine learning network model and the label image will be calculated. As the training error, the adaptive moment estimation optimization algorithm (Adam) is used to minimize the training error to train the machine learning network model; after each batch of training, use the test set to test the performance of the machine learning network model , when the performance test index of the machine learning network model (the mean square error between the output image and the label image) tends to be stable, it is considered that the training of the machine learning network model is completed, and the trained machine learning network model is saved;

6) Machine learning network model for OCTA imaging:

Using the trained machine learning network model, the OCT structure image of the sample collected by the OCTA device is used as input, the number of scans at the same slow-axis scanning position is 4, and the output image is the OCTA angiography image, as shown in Figure 2.

Finally, it should be noted that the purpose of publishing the embodiments is to help further understanding of the present invention, but those skilled in the art can understand that various replacements and modifications can be made without departing from the spirit and scope of the present invention and the appended claims. It is possible. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection of the present invention shall be subject to the scope defined by the claims.

Claims (6)

1. An optical coherent angiography imaging method based on machine learning, characterized in that the optical coherent angiography imaging method comprises the following steps:

1) generating an original data set:

acquiring an OCT three-dimensional structural image of a sample by using OCTA equipment, and generating an original data set required by network model training, wherein the original data set comprises j multiplied by k groups of OCT structural image sequences, each group of OCT structural image sequences comprises i OCT structural images with two-dimensional cross sections (B-Scan), k is the number of the sample, j is the number of scanning positions of a slow axis of each sample, i is the scanning times of the same scanning position of the slow axis of the same sample, i is a natural number larger than 4, j is a natural number larger than 50, and k is a natural number larger than 5;

2) and (3) screening data:

registering i B-Scan face OCT structural images in the same group of OCT structural images by adopting a rigid registration algorithm, calculating registration accuracy by utilizing a correlation algorithm after registration, removing the whole group of OCT structural images with poor registration effect, and reserving n groups of screened OCT structural images, wherein n is a natural number, and

3) generating a training data set:

carrying out contrast imaging by using the n groups of screened OCT structural images obtained in the step 2) by adopting an OCTA algorithm, wherein each group of OCT structural images obtains a B-Scan surface OCTA contrast image called a label image; taking m B-Scan surface OCT structural images from i B-Scan surface OCT structural images of a group of OCT structural images corresponding to each label image, wherein m is 2,3 or 4, the m is called input data, the input data is matched with the label images, and the input data and the label images form a training data set required by network model training;

4) establishing a machine learning network model:

constructing a machine learning network model, and setting hyper-parameters of the machine learning network model; and dividing the training data set into n₁Group training set and n₂Group test set, training set and test set independent of each other, n₁And n₂Are respectively natural numbers, and n₁+n₂＝n，

5) Training a machine learning network model:

using the machine learning network model established in the step 4), taking the input data in the training data set as the input of the machine learning network model, wherein n is used₁The group training set is used for training the machine learning network model and takes n₂The group test set is used for verifying the performance of the machine learning network model; in the training process, the training set is divided into a plurality of batches and repeatedly input into the machine learning network model for training for a plurality of times, the difference between the output image and the label image of the machine learning network model is judged or calculated as a training error to train the machine learning network model, after the training of each batch is finished, the test set is used for carrying out performance test on the machine learning network model, when the performance test indexes of the standby machine learning network model tend to be stable, the training of the machine learning network model is considered to be finished, and the trained machine learning network model is stored;

6) performing OCTA imaging by using a machine learning network model:

and (3) by utilizing the trained machine learning network model, taking the OCT structural image of the sample acquired by the OCTA equipment as input, wherein the output image is the OCTA contrast image.

2. The method as claimed in claim 1, wherein in step 1), when the OCTA device collects the samples, the OCT structure images of one B-Scan plane are obtained by scanning one slow axis scanning position of one sample once, the OCT structure images of the same slow axis scanning position are scanned i times, each sample has j slow axis scanning positions, there are k samples, so as to obtain i × j × k OCT structure images of the B-Scan plane, and the i × j × k OCT structure images of the B-Scan plane are divided into j × k groups of OCT structure images, each group of OCT structure images includes i OCT structure images of the B-Scan plane, that is, each slow axis scanning position corresponds to one group of OCT structure images, so as to obtain the original data set.

3. The optical coherence angiography imaging method according to claim 1, wherein in step 2), for the j × k sets of OCT structure images after registration, the registration accuracy is compared, the first n sets of OCT structure images with higher registration accuracy remain, the remaining sets are considered to have poor registration effect, and these sets of OCT structure images are rejected.

4. The optical coherence angiography imaging method according to claim 1, wherein in step 4), the machine learning network model employs a deep convolutional neural network CNN, a generative countermeasure network GAN, or a recurrent neural network RNN.

5. The method of claim 1, wherein in step 4), the hyper-parameters include the number of network layers, convolution kernel, learning rate, parameter initialization, number of training rounds, and batch size.

6. The method of claim 1, wherein in step 5), a mean square error, a structural similarity or a peak signal-to-noise ratio between the output image and the label image is used as a training error, and one of a stochastic gradient descent, an adaptive moment estimation optimization algorithm and a momentum algorithm is used to minimize the training error to train the machine learning network model.

Images