LU500798B1 - Full-Automatic Segmentation Method for Coronary Artery Calcium Lesions Based on Non-Contrast Chest CT - Google Patents

Abstract

The present invention discloses a full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT, including the following steps: pre-processing an original non-contrast chest CT image to obtain pre-processed data; inputting the pre-processed data into a heart segmentation model, and segmenting a heart region in the original non-contrast chest CT image to obtain a heart segmentation image corresponding to the original non-contrast chest CT image and including the heart region; screening suspected coronary artery calcium lesions in the heart region; extracting a plurality of candidate sample blocks from the screened suspected coronary artery calcium lesions; and obtaining coronary artery calcium lesion segmentation results of the candidate sample blocks by using a calcium lesion segmentation model based on the plurality of candidate sample blocks and corresponding coordinates thereof in a coordinate system corresponding to the original non-contrast chest CT image.

Description

Description Full-Automatic Segmentation Method for Coronary Artery Calcium Lesions Based on Non-Contrast Chest CT Technical Field The present invention relates to the technical field of medical image segmentation, and more particularly, to a full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT.

Background Art Cardiovascular diseases currently have the highest mortality rate and the prevalence and mortality of cardiovascular diseases continue to rise.

Elderly people with cardiovascular diseases have a weakened immunity and are more likely to contract other diseases and turn into critically ill patients.

Therefore, the early diagnosis of coronary heart diseases is particularly necessary.

Sufficient medical evidences have been raised to support that coronary artery calcium is an independent risk factor for adverse cardiovascular events in asymptomatic patients.

Quantitative analysis on coronary artery calcium (CAC) can predict the possibility of cardiovascular diseases.

Nowadays, cardiac CT with ECG-gating is often used clinically to assess coronary artery calcium.

The specific post-processing requires manually selecting coronary artery calcium lesions on cardiac CT image using dedicated software, and then the calcium score is given based on the selected lesions, which is so called coronary artery calcium scoring (CACS). This is a semi-automatic process.

The dedicated calcium-scoring CT scan is usually incorporated in coronary CT angiography (CTA) examination, the price of which is several times that of a non- contrast chest CT, and the radiation dose is also much higher.

In addition, there is a certain risk in coronary CTA and an ECG gating device is required as well as strict breathing cooperation by patients.

Coronary CTA is not available at every level of hospital.

The related prior art for automatic calculation of CACS based on cardiac CT still has the following shortcomings: a) The existing solution adopts relatively traditional artificially-designed image features during the process of feature extraction from a suspected calcium region: histogram of oriented gradient (HOG) features, local binary pattern (LBP) features, Haar features and texture features. b) The existing solution screens the suspected calcium region by means of an area threshold, and then uses a simple classifier to determine whether the suspected calcium region belongs to a calcified one.

This setting of the area threshold must be based on prior knowledge, which is often a rule summed up from regular data, and some special circumstances may not fit this rule.

Assuming that there is a calcium lesion outside the set area threshold, there is no chance of being classified in the existing solution at all, and missed detection will occur.

The non-contrast chest CT is relatively inexpensive and widely used, and it is also available in the ordinary physical examination institutions.

If the coronary artery calcium score can be automatically calculated based on non-contrast chest CT, the efficiency of clinical diagnosis will be greatly improved (avoiding additional cardiac CT scans, providing additional clinical information to complete one-stop cardiopulmonary screening, and no need for manual selection of a calcium region by a doctor), the cost of the patients and the radiation dose will be reduced.

However, compared with the above prior art, the calculation of the calcium score based on non- contrast chest CT is more complicated, especially in the chest CT with lung window setting of more noise.

In this case, the classification method (traditional random forest method, etc.) used in the above prior art for processing the suspected calcium region will be greatly challenged.

In view of the above shortcomings of the prior art, a full-automatic segmentation solution for coronary artery calcium lesions based on non-contrast chest CT is urgently needed in the art.

Summary of the Invention In view of this, the objective of the embodiments of the present invention is to propose a full-automatic segmentation method for coronary artery calcium lesions based on non- contrast chest CT, which can solve the problems in the above prior art that full- automatic segmentation of coronary artery calcium lesions based on non-contrast chest CT cannot be achieved and the like.

Based on the above objective, one aspect of the embodiments of the present invention provides a full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT, including the following steps: step 1, pre-processing an original non-contrast chest CT image to obtain pre-processed data: step 2, inputting the pre-processed data into a pre-trained heart segmentation model, and segmenting a heart region in the original non-contrast chest CT image to obtain a heart segmentation image corresponding to the original non-contrast chest CT image and including the heart region; step 3, screening suspected coronary artery calcium lesions in the heart region; step 4, extracting a plurality of candidate sample blocks from the screened suspected coronary artery calcium lesions; and step 5, obtaining a coronary artery calcium lesion segmentation result of each candidate sample block by using a pre-trained calcium lesion segmentation model based on the plurality of candidate sample blocks and corresponding coordinates thereof in a coordinate system corresponding to the original non-contrast chest CT image.

In some embodiments, step 1 further includes: step 11, scaling the original non-contrast chest CT image to a preset size; and step 12, normalizing the scaled original non-contrast chest CT image.

In some embodiments, step 3 further includes: step 31, restoring the size of the heart segmentation image to that of the original non- contrast chest CT image; step 32, determining the boundary of the heart region in the original non-contrast chest CT image based on the heart segmentation image of the restored size; and step 33, screening the suspected coronary artery calcium lesions within the boundary of the heart region.

In some embodiments, step 33 further includes: step 331, marking all suspected coronary artery calcium lesions within the boundary of the heart region according to CT values; and step 332, clustering all the marked suspected coronary artery calcium lesions, and screening the suspected coronary artery calcium lesions according to the clustering results and corresponding prior knowledge.

In some embodiments, before step 5, the method further includes: assembling the plurality of candidate sample blocks into a plurality of batches and then inputting into the calcium lesion segmentation model batch by batch.

In some embodiments, after step 5, the method further includes: obtaining final segmentation results by using a voting mechanism based on the coronary artery calcium lesion segmentation result of each candidate sample block.

In some embodiments, the training of the heart segmentation model includes the following steps: collecting non-contrast chest CT images used for training, and labeling the hearts to obtain corresponding heart Mask images; and training a first deep neural network by using the non-contrast chest CT images for training and the heart Mask images until a preset first convergence condition is satisfied, to obtain the heart segmentation model.

In some embodiments, the training of the calcium lesion segmentation model includes the following steps: labeling coronary artery calcium lesions on the non-contrast chest CT images for training to obtain corresponding coronary artery calcium lesion Mask images; extracting a plurality of Mask image sample blocks of a preset size from the coronary artery calcium lesion Mask images, and extracting a plurality of original image sample blocks of the preset size corresponding to the plurality of Mask image sample blocks from the non-contrast chest CT images for training, each Mask image sample block and the corresponding original image sample block constituting a sample pair; and training a second deep neural network by using the sample pairs until a preset second convergence condition is satisfied, to obtain the calcium lesion segmentation model.

In some embodiments, extracting a plurality of Mask image sample blocks of a preset size from the coronary artery calcium lesion Mask images includes the following steps: clustering the coronary artery calcium lesion Mask images; and extracting the Mask image sample blocks respectively by taking each center of clustering as a center.

In some embodiments, extracting a plurality of original image sample blocks of the preset size corresponding to the plurality of Mask image sample blocks from the non- contrast chest CT images for training includes the following steps: extracting a corresponding original image sample block from the non-contrast chest CT images for training according to extraction coordinates of each Mask image sample block.

Compared with the prior art, the present invention has the following technical effects: The embodiments of the present invention provide a two-stage automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT: in the first stage, a deep neural network trained with heart labeling data is first used to segment a heart region; and in the second stage, another deep neural network trained with coronary artery calcium lesion labeling data is used to segment coronary artery calcium lesions in the heart region obtained in the first stage.

The whole process does not require human involvement, which will greatly improve the efficiency of clinical diagnosis and provide more possibilities for the promotion of integrated cardiopulmonary screening.

The solution proposed by the present invention is suitable for the risk assessment of coronary heart diseases, helps to improve the accuracy of coronary heart disease screening, improves the individualized predictive value of adverse cardiovascular events, assists the clinical individualized comprehensive treatment plans, and reduces the risk of adverse cardiovascular events in patients.

The present invention uses the currently very successful deep neural network to automatically extract image features instead of traditional manual feature extraction.

This is also an obvious trend in current image processing tasks.

The neural network architecture adopted by the present invention is reasonably designed, and the automatic feature extraction process using the neural network replaces the tedious manual feature design and calculation process.

In the present invention, after the heart region is obtained in the first stage, suspected calcium regions are first clustered and then screened preliminarily using an area threshold.

After the preliminary screening, sample blocks are extracted based on the center of clustering (due to relatively large spatial sizes, these sample blocks are likely to overlap), and then these sample blocks are predicted by a trained deep segmentation network.

The calcium lesions previously excluded because of the area threshold may still be “recovered” during the inferring of the neural network, and the prediction results of these sample blocks with certain spatial overlap will finally determine whether each point in the space is a real calcium point through a voting mechanism.

In this way, more robust results can be obtained than using a simple classifier.

In summary, 1) the deep segmentation network in the second stage of the present invention can further cover the setting of the area threshold; and 2) the voting mechanism makes the segmentation results more robust.

Brief Description of the Drawings In order to describe the technical solutions in the embodiments of the present invention or in the prior art more clearly, the following briefly introduces the accompanying drawings required in the description of the embodiments or the prior art.

Apparently, the accompanying drawings in the following description show only some embodiments of the present invention, and a person of ordinary skill in the art may still derive other embodiments from these accompanying drawings without any creative effort.

Fig. 1 is a schematic flowchart of a full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT according to an embodiment of the present invention; Fig. 2 is a schematic flowchart of a pre-processing process of an original non-contrast chest CT image; Fig. 3 is a schematic flowchart of a process of screening suspected coronary artery calcium lesions in a heart region; Fig. 4 is a schematic flowchart of screening suspected coronary artery calcium lesions within a boundary of the heart region; Fig. 5 is a schematic flowchart of a process of training a heart segmentation model; Fig. 6 is a schematic diagram of labeling data in the heart segmentation model; Fig. 7 is a schematic flowchart of a process of training a calcium lesion segmentation model;

Fig. 8 is a schematic diagram of labeling data in the calcium lesion segmentation model; and Fig. 9 is a comparison diagram of effects achieved when the method of the present invention is used in a real segmentation task.

Detailed Description of the Invention In order to make the objectives, technical solutions and advantages of the present invention clearer, the following further describes the embodiments of the present invention in detail with specific embodiments and the accompanying drawings.

It should be noted that all expressions such as “first” and “second” in the embodiments of the present invention are used to distinguish a plurality of different entities or different parameters with the same name.

Hence, the expressions “first”, “second”, etc. are only for the convenience of description, and should not be construed as limiting the embodiments of the present invention, and subsequent embodiments will not describe this one by one.

Based on the above objective, the present invention proposes an embodiment of a full- automatic segmentation method for coronary artery calcium lesions based on non- contrast chest CT.

Fig. 1 shows a schematic flowchart of the method.

This method 1s mainly based on a segmentation model commonly used in deep learning, such as UNet.

However, for those skilled in the art, the segmentation network used in the present invention may also be other segmentation networks that can achieve the purpose of the present invention in addition to the basic UNet.

The segmentation solution of coronary artery calcium lesions in the present invention was divided into two main stages, because the coronary artery calcium lesions are very small in the entire non-contrast chest CT.

In a 512x512x100 three-dimensional CT image, the calcium lesions occupy less than one in ten thousand of pixels.

Therefore, in order to accomplish such a segmentation task, the present invention adopts a coarse-to-fine strategy.

This method segments a heart region through a heart segmentation model in the first stage, then implements a more delicate segmentation operation through a calcium lesion segmentation model in the second stage based on the obtained heart region, and finally obtains coronary artery calcium lesion segmentation results.

Specifically, as shown in Fig. 1, the full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT includes the following steps: Step S1, an original non-contrast chest CT image is pre-processed to obtain pre- processed data.

When a coronary artery calcium lesion segmentation operation is required, previous non-contrast chest CT information is processed to obtain data in the format of DICOM (Digital Imaging and Communications in Medicine), the data of each DICOM image including image data of a slice of a certain layer of the chest.

For the convenience of storage and management, multi-layer DICOM image data of a patient may be merged into a nii digital format file to save the original non-contrast chest CT image.

Then, the original non-contrast chest CT image in the nii file is pre-processed.

Step S2, the pre-processed data is inputted into a pre-trained heart segmentation model, and a heart region in the original non-contrast chest CT image is segmented to obtain a heart segmentation image corresponding to the original non-contrast chest CT image and including the heart region.

This step is the first stage of segmentation, that is, heart segmentation, which may also be understood as coarse segmentation.

To achieve the segmentation of the heart region, a suitable deep neural network (for example, UNet) is first selected to construct the heart segmentation model.

The training process of the heart segmentation model will be described in detail below.

Step S3, suspected coronary artery calcium lesions are screened in the heart region.

Before the second stage of segmentation, all suspected coronary artery calcium lesions in the heart region need to be preliminarily screened to filter out some suspected calcium lesions, thereby reducing the burden of the second-stage segmentation model.

Step S4, a plurality of candidate sample blocks are extracted from the screened suspected coronary artery calcium lesions.

From the screened suspected coronary artery calcium lesions, a corresponding candidate sample block is extracted by taking each center of clustering as a center.

The candidate sample blocks have a size of (16, 64, 64) (that is, the input size of the second- stage segmentation model). However, the size is not fixed (16, 64, 64). Those who are skilled in the art may set a suitable size based on GPU configuration and segmentation results.

Step S5, a coronary artery calcium lesion segmentation result of each candidate sample block is obtained by using a pre-trained calcium lesion segmentation model based on the plurality of candidate sample blocks and corresponding coordinates thereof in a coordinate system corresponding to the original non-contrast chest CT image.

This step is the second stage of segmentation, that is, coronary artery calcium lesion segmentation, which may also be understood as fine segmentation.

To achieve the segmentation of coronary artery calcium lesions, a suitable deep neural network (for example, UNet) is first selected to construct the calcium lesion segmentation model.

The training process of the calcium lesion segmentation model will be described in detail below.

Fig. 2 shows a schematic flowchart of the pre-processing process of the original non- contrast chest CT image.

As shown in Fig. 2, the image pre-processing process of step 1 described above further includes: step S11, the original non-contrast chest CT image is scaled to a preset size.

Since different parameters of different CT data acquisition devices result in heterogeneous CT images, the nii data needs to be transformed into a uniform size (which can be achieved by a Reshape function) by interpolation.

For example, the size may be (64, 256, 256) (that is, the input size of the first-stage segmentation model). However, the size is not fixed (64, 256, 256). Those who are skilled in the art may set a suitable size based on GPU configuration and segmentation results.

Step S12, the scaled original non-contrast chest CT image is normalized, that is, the data is transformed to a [0, 1] space and then inputted into the trained heart segmentation model.

Fig. 3 shows a schematic flowchart of the process of screening suspected coronary artery calcium lesions in the heart region.

As shown in Fig. 3, step S3 above further includes: Step S31, the size of the heart segmentation image is restored to that of the original non-contrast chest CT image.

In this step, the heart segmentation image inferred by the heart segmentation model needs to be restored to the original size by interpolation again (which can be achieved by the Reshape function again). Step S32, a boundary of the heart region is determined in the original non-contrast chest CT image according to the heart segmentation image of the restored size.

The boundary may be represented by a rectangular BoundingBox, and then coordinates of a corner of the BoundingBox relative to the coordinates of the entire CT are recorded.

Step S33, the suspected coronary artery calcium lesions are screened within the boundary of the heart region.

Specifically, as shown in Fig. 4, this step further includes: step S331, all suspected coronary artery calcium lesions are marked within the boundary of the heart region according to CT values, specifically, a threshold (for example, 130 HU (Hounsfield Unit)) is set, then for each pixel in the heart region, if its CT value is greater than 130 HU, the pixel is marked as the suspected calcium point, otherwise, the pixel is marked as the background; and step S332, all the marked suspected coronary artery calcium lesions are clustered, and the suspected coronary artery calcium lesions are screened according to the clustering results and corresponding prior knowledge, wherein each cluster in the clustering results has its cluster center and corresponding area, and some suspected calcium lesions can be preliminarily filtered according to the corresponding prior knowledge (for example, experience tells us what range the coronary artery calcium lesions should be in). Next, when step S4 1s performed, for the suspected coronary artery calcium lesions left after the preliminary screening, a plurality of candidate sample blocks may be extracted based on the centers of clustering, and their coordinates in the heart BoundingBox may be transformed to the original CT coordinates and recorded.

When there are many candidate sample blocks to be extracted, if a single sample block (Patch) is serially inputted to the calcium lesion segmentation model for inferring, the efficiency will be relatively low.

Therefore, in order to accelerate the inferring of the calcium lesion segmentation model, the candidate sample blocks may be assembled into a plurality of batches (each batch may include 128 samples or more, depending on GPU configuration), and then inputted into the calcium lesion segmentation model batch by batch for parallel inferring, so as to achieve the purpose of speeding up the inferring.

After the segmentation result of each candidate sample block is obtained by the inferring of the calcium segmentation model, the segmentation results are assembled one by one back to the original CT coordinates to obtain the coronary artery calcium lesion segmentation results based on the non-contrast chest CT image.

Different candidate sample blocks may overlap, so in order to make the results more robust, the final calcium segmentation results may be determined according to a voting mechanism, which may be majority voting here.

Fig. 5 shows a schematic flowchart of a process of training the heart segmentation model.

As shown in Fig. 5, the training of the heart segmentation model includes the following steps: step S51, non-contrast chest CT images for training are collected and the hearts are labeled to obtain corresponding heart Mask images; and step S52, a first deep neural network is trained by using the non-contrast chest CT images for training and the heart Mask images until a preset first convergence condition is satisfied, to obtain the heart segmentation model.

The following illustrates an example of training the heart segmentation model.

First, hundreds of non-contrast chest CT data are collected from hospitals, each of which includes varying degrees of coronary artery calcium.

Then, hundreds of heart Masks are labeled among the data to train a deep neural network heart segmentation model, that is, in each case of non-contrast chest CT, the pixels of the heart region are labeled 1, the aorta is labeled 2, and the background is labeled 0, as shown in Fig. 6. Each case of non-contrast chest CT data has a labeled corresponding heart Mask image.

Then the specified model can be trained.

The training indicates that a deep neural network is given, network parameters are first initialized, then the network performs forward operations on the given input data to obtain output results, and differences between the output results and the labeled data are calculated and propagated back to the network for updating the parameters to achieve the purpose of network training.

After converging the network parameters, the expected output result for a given input can have the same expression form as the labeled data.

The differences in the training process are generally calculated by a loss function.

The loss function used in the present invention is TverskyLoss, which is a generalized DiceLoss and can directly calculate the degree of overlap between the network output results and the labeled Masks.

The formula is as follows: T(a, 8) = — Dies Poig pre Poigoi + a Si, Poigu + à Se P1i9o: This formula expresses a two-category segmentation form, which can be expanded to a multi-category form during specific implementation.

For example, the heart segmentation model of the present invention aims at three-category segmentation.

It should be noted that, before the chest CT data with relatively large size difference is inputted into the network for training, all the data needs to be reshaped to a uniform size, such as (64, 256, 256). Of course, the data pre-processing stage also includes common random clipping, random mirroring and normalization operations. Then an Adam optimizer is used to train for 200 rounds. The initial learning rate of the Adam optimizer is set to 2e-4, the momentum parameter B1 is set to 0.5, and Bz is set to 0.99. BatchSize is set to 4.

Fig. 7 shows a schematic flowchart of a process of training the calcium lesion segmentation model. As shown in Fig. 7, the training of the calcium lesion segmentation model includes the following steps: step S71, coronary artery calcium lesions are labeled on the non-contrast chest CT images for training to obtain corresponding coronary artery calcium lesion Mask images; step S72, a plurality of Mask image sample blocks of a preset size are extracted from the coronary artery calcium lesion Mask images, and a plurality of original image sample blocks of the preset size corresponding to the plurality of Mask image sample blocks are extracted from the non- contrast chest CT images for training, each Mask image sample block and the corresponding original image sample block constituting a sample pair; and step S73, a second deep neural network is trained by using the sample pairs until a preset second convergence condition is satisfied, to obtain the calcium lesion segmentation model. Extracting a plurality of Mask image sample blocks of a preset size from the coronary artery calcium lesion Mask images specifically includes: the coronary artery calcium lesion Mask images are clustered; and the Mask image sample blocks are respectively extracted by taking each center of clustering as a center. Extracting a plurality of original image sample blocks of the preset size corresponding to the plurality of Mask image sample blocks from the non-contrast chest CT images for training specifically includes: a corresponding original image sample block is extracted according to extraction coordinates of each Mask image sample block from the non-contrast chest CT images for training.

Here, the following strategy is adopted for the collection of training data for the calcium lesion segmentation model. First, hundreds of coronary artery calcium lesion Masks are labeled for training another deep neural network model, as shown in Fig. 8. Then the coronary artery calcium lesion Masks are clustered, and sample blocks of a preset size of (16, 64, 64) are extracted by taking each center of clustering as a center. Meanwhile, the extracted reference coordinates are implemented to the original image. Then, N sample pairs of the preset size are constituted (an original image sample block and a corresponding Mask sample block constitute a sample pair). Finally, the calcium lesion segmentation model is trained based on these sample pairs.

The specific training process is similar to the heart segmentation model training process exemplified above, so details are not described herein again.

Preferably, after the neural network completes the inferring, the present invention also sets up some post-processing operations, the main purpose of which is to eliminate false positives.

The false positives may be divided into two categories, one is the false positive in the rib region (the HU value of the bone region is generally high), and the other is the false positive on the aorta (this is also the cause of labeling the aorta alone when the heart Mask is labeled in the above embodiment). In order to achieve the above purpose, when the present invention is implemented, the segmented heart Mask may be set to 0 for the non-1 value, and then the two parts of false positives can be removed by multiplying with the calcium segmentation results.

The present invention can achieve full-automatic segmentation of coronary artery calcium lesions based on non-contrast chest CT by using two stages of deep segmentation networks, and can obtain good segmentation results for highly noisy data.

The heart segmentation model in the first stage considers the segmentation of the aorta into a single category, which provides a basis for removing the false positives on the aorta from the calcium segmentation results in the second stage.

In addition, the method of clustering first and then extracting training samples for the segmentation of small targets such as calcium is also an innovation in the task of calcium segmentation.

Fig. 9 shows the effect achieved when the method of the present invention is used in a real segmentation task.

It can be seen from the comparison of the original image, Ground Truth, and segmentation results of coronary artery calcium lesions shown in Fig. 9, the present invention achieves the expected segmentation effect.

It should be noted that those who have ordinary skill in the art can understand that all or part of the processes in the operations of the foregoing embodiments, and can be implemented by instructing relevant hardware through a computer program.

The program may be stored in a computer-readable storage medium.

The program, when executed, may include the processes of the embodiments of the above operations.

The computer program can achieve the same or similar effects as the corresponding foregoing operation embodiments.

In addition, it should be understood that the computer-readable storage medium (for example, a memory) used to implement the operations of the present invention may be a volatile memory or a non-volatile memory, or may include a volatile memory and a non-volatile memory. As an example and not limitation, the non-volatile memory may include a read-only memory (ROM), a programmable ROM (PROM), an electrically programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), or a flash memory. The volatile memory may include a random access memory (RAM), which may serve as an external cache memory. As an example and not limitation, the RAM may be obtained in many forms, such as a synchronous RAM (DRAM), a dynamic RAM (DRAM), a synchronous DRAM (SDRAM), a double data rate SDRAM (DDR SDRAM), an enhanced SDRAM (ESDRAM), a synchronous link DRAM (SLDRAM) and a direct Rambus RAM (DRRAM). The storage devices in the disclosed aspects are intended to include, but are not limited to, these and other suitable types of memories.

Those who are skilled in the art will also understand that the various exemplary logic blocks, modules, circuits, and algorithm steps described in the disclosure herein can be implemented or executed using the following components designed to perform the functions described herein: a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination of these components. The general- purpose processor may be a microprocessor, but alternatively, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in combination with a DSP, and/or any other such configuration.

In one or more exemplary designs, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions can be stored as one or more instructions or codes on a computer-readable medium or transmitted by the computer-readable medium. The computer-readable medium includes computer storage media and communication media, and the communication media include any media that facilitate the transfer of a computer program from one location to another. The storage media may be any available media that can be accessed by a general-purpose or special-purpose computer. As an example and not limitation, the computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage devices, magnetic disk storage devices or other magnetic storage devices, or may be used for carrying or storing required program codes in the form of instructions or data structures, and can be accessed by a general- purpose or special-purpose computer or a general-purpose or special-purpose processor. In addition, any connection may be properly referred to as a computer-readable medium. For example, if a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technology such as infrared, radio and microwave 1s used to send software from a website, server, or other remote source, the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technology such as infrared, radio and microwave is included in the definition of the medium. As used herein, the magnetic disks and optical disks include compact disks (CDs), laser disks, optical disks, digital versatile disks (DVD), floppy disks, and Blu-ray disks. The magnetic disks usually reproduce data magnetically, while the optical disks optically reproduce data by means of laser. Combinations of the above contents should also be included in the scope of the computer-readable medium.

The foregoing are exemplary embodiments disclosed in the present invention. The sequence disclosed in the foregoing embodiments of the present invention 1s only for description, and does not represent the superiority or inferiority of the embodiments. However, it should be noted that the discussion of any of the above embodiments 1s only exemplary, and is not intended to imply that the scope of the disclosure (including the claims) in the embodiments of the present invention 1s limited to these examples, and many changes and modifications may be made without departing from the scope defined by the claims. The functions, steps and/or operations of the claims according to the disclosed embodiments described herein do not need to be performed in any specific order. In addition, although the elements disclosed in the embodiments of the present invention may be described or required in individual forms, they may also be understood as plural unless explicitly limited to a singular number.

Claims (10)

Claims

1. A full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT, characterized by comprising the following steps: step 1, pre-processing an original non-contrast chest CT image to obtain pre-processed data; step 2, inputting the pre-processed data into a pre-trained heart segmentation model, and segmenting a heart region in the original non-contrast chest CT image to obtain a heart segmentation image corresponding to the original non-contrast chest CT image and including the heart region; step 3, screening suspected coronary artery calcium lesions in the heart region; step 4, extracting a plurality of candidate sample blocks from the screened suspected coronary artery calcium lesions; and step 5, obtaining a coronary artery calcium lesion segmentation result of each candidate sample block by using a pre-trained calcium lesion segmentation model based on the plurality of candidate sample blocks and corresponding coordinates thereof in a coordinate system corresponding to the original non-contrast chest CT image.

2. The full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT according to claim 1, characterized in that step 1 further comprises: step 11, scaling the original non-contrast chest CT image to a preset size; and step 12, normalizing the scaled original non-contrast chest CT image.

3. The full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT according to claim 2, characterized in that step 3 further comprises: step 31, restoring the size of the heart segmentation image to that of the original non- contrast chest CT image; step 32, determining a boundary of the heart region in the original non-contrast chest CT image based on the heart segmentation image of the restored size; and step 33, screening the suspected coronary artery calcium lesions within the boundary of the heart region.

4. The full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT according to claim 3, characterized in that step 33 further comprises: step 331, marking all suspected coronary artery calcium lesions within the boundary of the heart region according to CT values; and step 332, clustering all the marked suspected coronary artery calcium lesions, and screening the suspected coronary artery calcium lesions according to the clustering results and corresponding prior knowledge.

5. The full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT according to claim 4, characterized in that before step 5, the method further comprises: assembling the plurality of candidate sample blocks into a plurality of batches and then inputting into the calcium lesion segmentation model batch by batch.

6. The full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT according to claim 1, characterized in that after step 5, the method further comprises: obtaining final segmentation results by using a voting mechanism based on the coronary artery calcium lesion segmentation result of each candidate sample block.

7. The full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT according to claim 1, characterized in that the training of the heart segmentation model comprises the following steps: collecting non-contrast chest CT images used for training, and labeling hearts to obtain corresponding heart Mask images; and training a first deep neural network by using the non-contrast chest CT images for training and the heart Mask images until a preset first convergence condition is satisfied, to obtain the heart segmentation model.

8. The full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT according to claim 7, characterized in that the training of the calcium lesion segmentation model comprises the following steps:

labeling coronary artery calcium lesions on the non-contrast chest CT images for training to obtain corresponding coronary artery calcium lesion Mask images; extracting a plurality of Mask image sample blocks of a preset size from the coronary artery calcium lesion Mask images, and extracting a plurality of original image sample blocks of the preset size corresponding to the plurality of Mask image sample blocks from the non-contrast chest CT images for training, each Mask image sample block and the corresponding original image sample block constituting a sample pair; and training a second deep neural network by using the sample pairs until a preset second convergence condition is satisfied, to obtain the calcium lesion segmentation model.

9. The full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT according to claim 8, characterized in that extracting a plurality of Mask image sample blocks of a preset size from the coronary artery calcium lesion Mask images comprises the following steps: clustering the coronary artery calcium lesion Mask images; and extracting the Mask image sample blocks respectively by taking each center of clustering as a center.

10. The full-automatic segmentation method for coronary artery calcium lesions based on non-contrast chest CT according to claim 8, characterized in that extracting a plurality of original image sample blocks of the preset size corresponding to the plurality of Mask image sample blocks from the non-contrast chest CT images for training comprises the following steps: extracting a corresponding original image sample block from the non-contrast chest CT images for training according to extraction coordinates of each Mask image sample block.