Carotid plaque detection model training method based on CTA image

Abstract

The invention discloses a carotid plaque detection model training method based on a CTA image, particularly relates to the technical field of heart and cerebral vessels, and realizes automatic, efficient and accurate detection of carotid plaques by collecting CTA image data of a large number of patients and innovatively adopting a network architecture of ResUNet in combination with a pyramid scene analysis network PSPNet for model construction. The model can effectively and automatically identify plaque forms and plaque positions, and solves the problems that a traditional plaque detection method depends on doctor experience and is high in subjectivity and long in manual analysis time consumption. The model can be used as an auxiliary tool for clinicians, the diagnosis efficiency and accuracy are improved, and the model has important clinical application value.

Description

Technical Field

[0001] The present invention relates to the field of cardiovascular and cerebrovascular technology, and in particular to a method for training a carotid artery plaque detection model based on CTA images. Background Art

[0002] Stroke remains the leading cause of death and disability worldwide. Clinically, carotid atherosclerosis is closely associated with stroke and cerebral infarction. The formation of carotid plaques is closely related to atherosclerosis and is one of the important causes of ischemic stroke. Carotid atherosclerosis is mainly manifested as the formation of atherosclerotic plaques in the carotid arteries. The presence of plaques can lead to stenosis or occlusion of the carotid arteries in the corresponding segments, leading to the appearance of ischemic clinical symptoms. In addition, due to the unstable composition, vulnerable plaques are prone to rupture and displacement, which can lead to clinical symptoms and increase the burden on patients. Therefore, early and accurate identification and evaluation of carotid plaques are of great significance for the prevention of cerebrovascular events.

[0003] Atherosclerotic plaques in the carotid arteries can be diagnosed by a variety of imaging methods, such as ultrasound, CTA, MRI, DSA, etc., which can achieve qualitative and quantitative analysis of plaque composition, detect and monitor the progression of plaques at an early stage, and provide medical staff with the opportunity to implement preventive measures to reduce the risk of stroke and related complications.

[0004] CTA is less operator-dependent than ultrasound and has faster imaging speed than MRI, and has been widely used to evaluate carotid plaques. Traditionally, the detection of carotid plaques relies on manual inspection of images by experienced radiologists, which requires a long time to conduct a detailed evaluation of all vascular structures. In addition, CTA image analysis also requires expertise in cerebrovascular imaging, which relies on the subjective diagnosis of radiologists. Therefore, current methods for detecting carotid plaques have problems such as strong subjectivity and long time consumption.

[0005] In recent years, artificial intelligence algorithms have been increasingly used to improve the efficiency and accuracy of image analysis, and have shown high performance and rapid and accurate assessment of large amounts of data while reducing human errors. Two studies, "A Radiomics Approach to Assess High Risk Carotid Plaques: A Non-invasive Imaging Biomarker, Retrospective Study" and "Artificial Intelligence in Cardiovascular Imaging: JACC State-of-the-Art Review", pointed out that the performance of artificial intelligence models was better than the visual assessment model of the reader, indicating that artificial intelligence algorithms can be used as clinical tools to provide objective insights for disease diagnosis. Deep learning is an artificial intelligence method that uses artificial neural networks to generate automatic predictions directly from image data. It can achieve tasks such as image detection, classification, and reconstruction, and has been successfully applied in the medical field to assist in the clinical diagnosis and prognosis of diseases. Deep learning methods are based on various artificial neural networks to obtain information about relevant and effective features from image data. Compared with machine learning, it completely changes the way plaques are detected and segmented.

[0006] However, the current carotid plaque detection model is constructed based on the CTA image architecture and uses the ResUNet combined with PSPNet network architecture to help solve the problems of the current carotid plaque detection method relying on the experience and judgment of radiologists, which is highly subjective and time-consuming. Summary of the invention

[0007] To this end, the present invention provides a carotid artery plaque detection model training method based on CTA images to solve the problems raised in the background technology.

[0008] In order to achieve the above object, the present invention provides the following technical solution: a method for training a carotid artery plaque detection model based on CTA images, comprising the following steps:

[0009] Step 1: Data preparation and screening: Retrospectively collect CTA image data of patients with carotid atherosclerotic plaques and record the baseline data of the patients; screen the data according to the inclusion and exclusion criteria, exclude unclear or incomplete CTA image data, and divide the patient data into training set, validation set and test set;

[0010] Step 2: Image preprocessing and plaque grouping: divide and depict plaque contours, and group plaque locations and morphologies; normalize the original CTA image data according to window level and width; extract smaller patches from large-volume CTA images to alleviate memory limitations and promote training efficiency; the characteristics are:

[0011] Step 3: Model architecture design: Use the ResUNet and pyramid scene parsing network PSPNet fusion model for patch segmentation;

[0012] The fusion model combines the output of the encoder of ResUNet with the output of the PPM module of PSPNet, and ensures the effective fusion of local and global information through element-by-element addition or cascading to obtain a more comprehensive carotid plaque feature representation; starting from the last layer of the residual block, pooling operations are applied with different scales of 1x1, 3x3 and 7x7 to obtain multi-scale carotid plaque feature representations. The results of the pooling operation are upsampled to the dimension of the original feature map and spliced ​​to the original feature map to obtain the fused carotid plaque feature map; the fused feature map is sent to the decoder of ResUNet, which gradually restores the resolution and fuses low-level and high-level features through jump links to obtain a feature map with fine details and strong semantic information, and then the pyramid pool is further combined with the final segmentation result through the convolution layer. Finally, the segmentation head outputs the patch prediction segmentation mask, which represents the category affiliation of each pixel in the input image;

[0013] Step 4: Model training and configuration: Configure the model on the Pytorch framework, set the initial learning rate, optimizer Adam and its parameters (such as beta coefficient, weight decay); define the loss function, using a combination of cross entropy CE loss and Dice loss to achieve balanced optimization;

[0014] Step 5: Step-by-step verification of the model: The fusion model is gradually verified using comparative studies and model-human real clinical scenario studies to evaluate its clinical application value;

[0015] Step 6: Model evaluation: The evaluation of the fusion model includes model performance testing, group analysis of model results, and step-by-step verification of the model obtained in step 5.

[0016] Preferably, the inclusion criteria in step one are: patients diagnosed with carotid plaques by CTA; the exclusion criteria in step one are: 1) patients with a history of interventional or surgical treatment, such as carotid stenting or carotid endarterectomy; 2) patients with carotid hemangioma or carotid vascular malformation; 3) patients whose CTA images for subsequent image processing are unclear or incomplete.

[0017] Preferably, in step 2, the plaque position grouping is specifically as follows: the lumen of the largest cross-sectional layer of the carotid artery plaque on the axial image is divided into four 90° regions, namely the anterior wall, the outer wall, the posterior wall and the inner wall; if the plaque spans two slices, the region where the thickest part of the plaque is located is selected; in step 2, the plaque morphology grouping is divided into two forms, smooth and non-smooth, according to the intersection surface of the plaque and the residual vascular lumen, and a non-smooth plaque refers to a plaque with an uneven surface and uneven lumen distribution.

[0018] Preferably, in step 4, the initial learning rate is set to 0.0005, the beta coefficient of Adam during the optimization period is set to (0.9, 0.99), the weight decay is 0.001, and the weight ratio of the cross entropy CE loss and the Dice loss in the loss function is 1:1.

[0019] Preferably, in step six, the model performance test uses recall (lesion level sensitivity), sensitivity (patient level) and precision indicators to evaluate the model performance on the test data set.

[0020] The present invention has the following advantages:

[0021] The present invention collects a large number of patients' CTA image data and innovatively uses the network architecture of ResUNet combined with the pyramid scene parsing network PSPNet to build a model, thereby realizing automatic, efficient and accurate detection of carotid artery plaques. The model can effectively and automatically identify plaque morphology and plaque location, solving the problem that traditional plaque detection methods rely on doctor experience, are highly subjective, and require long manual analysis time. The model of the present invention can be used as an auxiliary tool for clinicians to improve diagnostic efficiency and accuracy, and has important clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A model training flowchart of a method for training a carotid artery plaque detection model based on CTA images by fusing ResUNet and PSPNet provided in an embodiment of the present invention.

[0023] Figure 2 The main structural diagram of the model provided for the embodiment of the present invention.

[0024] Figure 3 A diagram showing some modeling results provided in an embodiment of the present invention.

[0025] Figure 4 This is a diagram of the diagnostic results of a comparative study provided in an embodiment of the present invention.

[0026] Figure 5 Diagnostic results of the model-human body study provided in the embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] In order to achieve the purpose of the present invention, this embodiment provides a method for training a carotid artery plaque detection model based on CTA images by integrating ResUNet and PSPNet. The training process is as follows: Figure 1 As shown, the implementation steps include the following steps:

[0029] Step 1: Data preparation and screening:

[0030] Patients with carotid atherosclerotic plaques who consecutively underwent head and neck CTA examination in a tertiary hospital from October 2020 to October 2022 were included.

[0031] The inclusion criteria were: 1) patients diagnosed with carotid artery plaque by CTA.

[0032] Exclusion criteria were: 1) patients with a history of interventional or surgical treatment, such as carotid artery stenting or carotid endarterectomy; 2) patients with carotid artery hemangioma or carotid vascular malformation; 3) patients with unclear or incomplete CTA images for subsequent image processing.

[0033] Patients were randomly divided into training, validation, and test sets in a ratio of 7:1.5:1.5.

[0034] A total of 647 patients with 1982 plaques were included in this example. The patients were randomly divided into a training set (N=475), a validation set (N=86) and a test set (N=36) according to the proportion. The average age was 65.08±9.73 years, 64.77±9.41 years and 66.94±9.85 years, respectively. The patients included 221 (34.16%) females and 426 (65.84%) males, of which 158 (33.26%) females and 317 (66.74%) males were in the training set, 28 (32.56%) females and 58 (67.44%) males were in the validation set, and 35 (40.70%) females and 51 (59.30%) males were in the test set (Table 1). The patients’ clinical baseline information, such as symptoms, medical history (including hypertension, diabetes), and life history (such as smoking and drinking), is also recorded in Table 1 .

[0035] Table 1 Clinical characteristics

[0036]

[0037]

[0038] Instruments and reagents for CTA image acquisition: A dual-source CT scanner was used, and the scanning parameters were set as follows: current: 125 mA, voltage: 100 kV, collimation: 16 × 0.6 mm, slice thickness: 0.75 mm. In addition, an intravenous indwelling needle, a double-barreled syringe, iodixanol contrast agent (320 mg I / ml) and 40 ml of normal saline were also used.

[0039] Scanning protocol for CTA image acquisition: During the examination, the patient was placed in the supine position with the head tilted back and held his breath at the end of exhalation for scanning. The scanning range was from the aortic arch to the top of the head, and the scanning direction was from the feet to the head. After the routine scan was completed, contrast agent was injected for enhanced scanning, and the scanning time was (8.5±1.5)s. Imaging of all patients was performed by an experienced imaging technician, who advised the patient not to swallow before the scan and to fix the head during the scan.

[0040] Step 2: Image preprocessing and patch grouping:

[0041] Each image was reviewed and labeled by two radiologists who had been engaged in head and neck imaging diagnosis for more than five years, and when the two reviewers disagreed on the results, a discussion was held to reach a consensus.

[0042] Plaque contour: Select the plaque at the carotid bifurcation and set the boundaries of the carotid plaque and the region of interest from distal to proximal on the axial image using ITK-SNAP software. Determination of the extent of plaque delineation: The entire extent of the plaque at the common carotid bifurcation. If the plaque is extensive, the lowest contour line is drawn to the aortic arch and the highest contour line is drawn to the skull base. The vessels on the occluded side of the carotid artery are not depicted, but the plaque on the lateral vessels needs to be depicted.

[0043] Plaque location grouping: Select the largest cross-sectional layer of the carotid plaque on the axial image and divide the cross-sectional lumen into four 90° regions: anterior wall, outer wall, posterior wall, and inner wall. If the plaque spans 2 slices, select the region where the thickest part of the plaque is located.

[0044] Plaque morphology grouping: According to the intersection of the plaque and the residual vascular lumen, there are two forms: smooth and rough. Rough plaque refers to a plaque with uneven surface and uneven lumen distribution.

[0045] The original CTA image is normalized to 0-1 according to the window level 400 and the window width 1200. Secondly, for large volume CTA images, smaller patches are extracted to facilitate training and alleviate memory limitations. A patch size of 256*256*128 is used to simplify the input data of this example, overlapping half of the pre-processed CTA image.

[0046] Step 3: Model architecture design:

[0047] This embodiment uses a model architecture that combines ResUNet with a pyramid scene parsing network PSPNet for patch segmentation, and 3D-CNN and ResNet-50 are the main architectures of the network.

[0048] The ResUNet architecture consists of an encoder, skip connections, and a decoder. The encoder is built using the residual blocks of ResNet, which gradually reduces the spatial resolution through a series of residual blocks (ResNet-50) while increasing the depth of the feature map. Skip connections are placed between the encoder and decoder to help retain more detail information when restoring resolution during decoding. The decoder, inspired by the UNet structure, gradually restores the resolution through skip connections and fuses low-level and high-level features to retain detailed information of the image.

[0049] PSPNet can capture global context information at different scales through the pyramid pooling module PPM. This module can fuse features of receptive fields of different sizes to improve the network's ability to understand the overall background of the image.

[0050] This embodiment combines the output of the encoder of ResUNet with the output of the PPM module of PSPNet, ensuring the effective fusion of local and global information through element-by-element addition or cascading. Starting from the last layer of the residual block, pooling operations are applied at different scales (1x1, 3x3, and 7x7). The results of these pooling operations are upsampled to the dimensions of the original feature map and spliced ​​to the original feature map. The fused feature map is fed into the decoder of ResUNet, where the pyramid pooling is further combined with the final segmentation result through the convolutional layer, and finally the patch prediction segmentation mask is output by the segmentation head.

[0051] This structure enables the network to not only obtain high-resolution features in local areas using ResUNet, but also obtain contextual information globally through PSPNet, thereby achieving better performance in patch segmentation tasks.

[0052] Step 4: Model training and configuration:

[0053] The model of this embodiment is trained on the PyTorch framework, using a combination of ResUNet and PSP modules. Combined with the warm-up strategy, the initial learning rate is set to 0.0005. The beta coefficient of the Adam optimizer is set to (0.9, 0.99), and the weight decay is 0.001. Figure 2 The loss function shown is a combination of the Cross Entropy (CE) loss and the Dice loss, assigned equal weights (1:1) to achieve a balanced optimization approach.

[0054] Step 5: Model verification step by step:

[0055] Comparative study: This example uses a test set to compare the detection performance of a separate AI model and six radiologists. In addition, no clinical information or other comparison images were provided for each CTA examination. Among the six radiologists, three junior radiologists had less than five years of working experience (D1, D2, D3), and the other three junior radiologists had more than five years of working experience (D4, D5, D6). Radiologists rely on their clinical expertise to diagnose plaques. The radiologists' test results and times were recorded and used for comparison with the model, such as Figure 2 shown.

[0056] Model-human real clinical scenario study: Six additional radiologists interpreted each CTA case with the help of the AI ​​model during the reading process. In addition, no clinical information or other comparison images were provided for each CTA examination. Three of them were junior staff with less than 5 years of work experience, namely D7, D8, and D9, and the other three were senior staff with more than 5 years of work experience (D10, D11, and D12). The final test results and reading time were recorded (automatically recorded). This study was used to simulate the gains of the model by radiologists in real clinical scenarios, such as Figure 2 shown.

[0057] Step 6: Model evaluation:

[0058] Model performance test: CTA images of 647 patients were included in the dataset, of which 475 were used for training, 86 for validation, and 86 for testing. The DL model based on CTA images has good plaque diagnosis accuracy (validation set: accuracy = 80.49%, sensitivity = 90.70%, recall rate = 84.62%; test set: accuracy = 78.37%, sensitivity = 91.86%, recall rate = 84.58%) (Table 2). In addition, some of the modeling results are shown in Figure 3(TP stands for True Positives, which refers to the number of samples that are actually positive but correctly predicted as positive; FN stands for False Negatives, which refers to the number of samples that are actually positive but incorrectly predicted as negative; FP stands for False Positives, which refers to the number of samples that are actually negative but incorrectly predicted as positive.)

[0059] Table 2 Diagnosis results of deep learning model

[0060]

[0061]

[0062] Group analysis of model results in the test set: Subgroup analysis was performed on the plaques in the test set to evaluate the accuracy of the model based on plaque location and plaque morphology. According to the location of the plaques, they were divided into four groups, including the anterior wall, outer wall, posterior wall, and medial wall. By comparing the plaque location identified by the model with the manually identified plaque location, it was found that the recall rates of the plaque locations in the anterior, posterior, inner, and outer groups were 83.72%, 76.32%, 89.25%, and 83.02%, respectively. According to the plaque morphology, the plaques were divided into smooth and non-smooth groups. The results showed that the recall rates of the plaque morphology in the smooth and non-smooth groups were 86.03% and 79.17%, respectively. (Table 3).

[0063] Table 3 Subgroup analysis of model results in the test set

[0064] TP FN quantity Recall value Anterior wall 36 7 43 0.8372 outer wall 29 9 38 0.7632 Back wall 83 10 93 0.8925 Inner wall 44 9 53 0.8302 smooth 154 25 179 0.8603 Not smooth 10 38 48 0.7917

[0065] Model step-by-step verification results:

[0066] Comparative study: For plaque diagnosis, the results showed that the recall rate of plaque diagnosis by the three junior radiologists was low, and the model recall value was higher than that of the three radiologists (p<0.001); for the three senior radiologists, the recall values ​​of plaque diagnosis were all high, and only the plaque recall value of the D5 radiologist was higher than that of the model, which showed that the model had higher performance in plaque diagnosis and could help doctors improve the accuracy of plaque diagnosis. The model diagnosis time of plaque was 6 seconds, which was significantly shorter than the diagnosis time of the six radiologists (97.50±15.24 seconds) (p<0.001)( Figure 4 ).

[0067] Model-human real clinical scenario study: For plaque diagnosis, the results showed that 6 radiologists (D7-D12) improved diagnostic recall. The radiologists' diagnosis (D7 D10 and D12) showed significantly higher accuracy than the model with the assistance of the model. Junior radiologists (D7) took longer to achieve higher diagnostic performance. Senior radiologists spent less time and performed better. The model diagnosis time for plaques was 6 seconds, which was still significantly shorter than the diagnosis time of 6 radiologists (74.67±36.12 seconds) (p<0.001)( Figure 5 ).

[0068] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for training a carotid artery plaque detection model based on CTA images, comprising the following steps: Step 1: Data preparation and screening: Retrospectively collect CTA image data of patients with carotid atherosclerotic plaques and record the baseline data of the patients; screen the data according to the inclusion and exclusion criteria, exclude unclear or incomplete CTA image data, and divide the patient data into training set, validation set and test set; Step 2: Image preprocessing and plaque grouping: divide and depict plaque contours, and group plaque locations and morphologies; normalize the original CTA image data according to window level and width; extract small-sized patches from large-volume CTA images to facilitate training and alleviate memory limitations; the characteristics are: Step 3: Model architecture design: Use the ResUNet and pyramid scene parsing network PSPNet fusion model for patch segmentation; The fusion model combines the output of the encoder of ResUNet with the output of the PPM module of PSPNet, and ensures the effective fusion of local and global information through element-by-element addition or cascading to obtain a more comprehensive carotid plaque feature representation; starting from the last layer of the residual block, pooling operations are applied with different scales of 1x1, 3x3 and 7x7 to obtain multi-scale carotid plaque feature representations. The results of the pooling operation are upsampled to the dimension of the original feature map and spliced ​​to the original feature map to obtain the fused carotid plaque feature map; the fused feature map is sent to the decoder of ResUNet, which gradually restores the resolution and fuses low-level and high-level features through jump links to obtain a feature map with fine details and strong semantic information, and then the pyramid pool is further combined with the final segmentation result through the convolution layer. Finally, the segmentation head outputs the patch prediction segmentation mask, which represents the category affiliation of each pixel in the input image; Step 4: Model training and configuration: Configure the model on the Pytorch framework, set the initial learning rate, optimizer Adam and its parameters; define the loss function, using a combination of cross entropy CE loss and Dice loss to achieve balanced optimization; Step 5: Step-by-step verification of the model: The fusion model is gradually verified using comparative studies and model-human real clinical scenario studies to evaluate its clinical application value; Step 6: Model evaluation: The evaluation of the fusion model includes model performance testing, group analysis of model results, and step-by-step verification of the model obtained in step 5.

2. The method for training a carotid artery plaque detection model based on CTA images according to claim 1, characterized in that: The inclusion criteria in step one are: patients diagnosed with carotid plaque by CTA; the exclusion criteria in step one are: 1) patients with a history of interventional or surgical treatment; 2) patients with carotid hemangioma or carotid vascular malformation; 3) patients with unclear or incomplete CTA images for subsequent image processing.

3. The method for training a carotid artery plaque detection model based on CTA images according to claim 1, characterized in that: In the step 2, the plaque position grouping is specifically as follows: the lumen of the largest cross-sectional layer of the carotid artery plaque on the axial image is divided into four 90° regions, namely the anterior wall, the outer wall, the posterior wall and the inner wall; if the plaque spans two slices, the region where the thickest part of the plaque is located is selected; in the step 2, the plaque morphology grouping is divided into two forms, smooth and non-smooth, according to the intersection surface of the plaque and the residual vascular lumen. Non-smooth plaque refers to a plaque with an uneven surface and uneven lumen distribution.

4. The method for training a carotid artery plaque detection model based on CTA images according to claim 1, characterized in that: In step 4, the initial learning rate is set to 0.0005, the beta coefficient of Adam during the optimization period is set to (0.9, 0.99), the weight decay is 0.001, and the weight ratio of the cross entropy CE loss and the Dice loss in the loss function is 1:

1.

5. The method for training a carotid artery plaque detection model based on CTA images according to claim 1, characterized in that: In step six, the model performance test uses recall, sensitivity, and precision indicators to evaluate the model performance on the test data set.

Images