Intracranial blood vessel wall magnetic resonance image quality control method based on deep learning

Abstract

The invention discloses an intracranial vascular wall magnetic resonance image quality control method based on deep learning. The method comprises the following steps: S1, data acquisition and screening; step S2, influence checking; s3, anonymizing the image in the DICOM format, and performing slice sampling and image block generation based on a blood vessel center line; s4, independently scoring the display quality of the blood vessel wall, the blood flow inhibition effect and the motion artifact influence by a doctor; s5, constructing a joint learning network model based on the ResNet-18 architecture, wherein the model comprises a shared backbone network and three independent full connection layers; s6, performing model training; and S7, inputting the preprocessed multi-channel image into the trained model, and outputting a quality scoring result of each dimension. The method has the advantages that low-quality images can be quickly recognized, intervention is prompted, a doctor is assisted in image quality control, and therefore the clinical application effect and diagnostic value of MR-VWI are improved.

Description

Technical Field

[0001] This invention relates to the field of medical imaging, and more particularly to a method for quality control of intracranial vascular wall magnetic resonance images based on deep learning. Background Technology

[0002] Intracranial atherosclerotic disease is one of the main pathogenic factors of ischemic stroke in my country. In recent years, with the development of magnetic resonance imaging technology, high-resolution magnetic resonance vascular wall imaging (MR-VWI) of intracranial arteries can perform morphological and compositional analysis of intracranial atherosclerotic plaques, providing important imaging evidence for assessing plaque stability, evaluating stroke risk, and guiding individualized treatment.

[0003] However, MR-VWI image quality is affected by a variety of factors, such as the hardware performance of the magnetic resonance imaging (MRI) equipment, the setting of scanning sequence parameters, the patient's cooperation, and artifact interference during the scanning process. Once image quality deteriorates, it directly affects the identification of vascular structures and lesion plaque characteristics, which may lead to misdiagnosis or missed diagnosis, thereby affecting the accuracy of clinical diagnosis and treatment decisions.

[0004] Currently, the assessment of MR-VWI image quality in clinical practice mainly relies on the subjective judgment of physicians, typically using manual scoring for image quality control. This method is not only time-consuming and labor-intensive, but also has poor repeatability, with scoring standards varying from person to person and being highly subjective. Furthermore, because manual assessment often lags behind image acquisition, it is difficult to provide timely feedback and correction of imaging quality problems, which may lead to duplicate scans or invalid images, increasing the burden on patients and medical costs, and may even cause medical disputes.

[0005] Therefore, it is necessary to provide a new deep learning-based method for quality control of intracranial vascular wall magnetic resonance images to solve the above-mentioned technical problems. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based method for quality control of intracranial vascular wall magnetic resonance images, in order to solve the problems mentioned in the background art.

[0007] To achieve the above-mentioned objectives, this invention provides a deep learning-based method for quality control of intracranial vascular wall magnetic resonance images, comprising the following steps:

[0008] Step S1: Obtain raw intracranial artery MR-VWI data from the medical imaging archive system, and screen valid case data based on preset exclusion criteria;

[0009] Step S2: Perform multi-sequence scanning using a 3.0T MRI device. The sequences include at least a 3D TOF (time-of-flight)-MRA sequence, a 3D black-blood T1WI SPACE sequence before and after contrast agent injection, and a 2D black-blood T2WI TSE (Turbo Spin Echo) sequence.

[0010] Step S3: Anonymize the DICOM (Digital Imaging and Communications in Medicine) format image, and perform slice sampling and image block generation based on the blood vessel centerline;

[0011] In step S4, two senior radiologists independently scored the quality of the vessel wall visualization, the effect of blood flow suppression, and the impact of motion artifacts.

[0012] Step S5: Construct a joint learning CNN (Convolutional Neural Network) model based on the ResNet-18 architecture. The model includes a shared backbone network and three independent fully connected layers, which correspond to the vessel wall score (CNN-1), blood flow inhibition score (CNN-2), and motion artifact score (CNN-3), respectively.

[0013] Step S6: Stratified random sampling is used to divide the training set, validation set, and test set. The network parameters are optimized by minimizing the mean square error between the predicted score and the doctor's score.

[0014] Step S7: Input the preprocessed multi-channel image into the trained model and output the quality score results for each dimension.

[0015] Furthermore, the exclusion criteria include:

[0016] Imaging data from patients with non-atherosclerotic lesions were excluded;

[0017] Exclude cases that have already undergone vascular interventional treatment;

[0018] Exclude invalid data that is incompletely acquired or has severe artifacts.

[0019] Furthermore, the image preprocessing in step S3 includes:

[0020] Centerline interpolation is performed on the target vessel segment;

[0021] Slices were taken at 2mm intervals along the centerline of the blood vessel.

[0022] Perform dynamic cropping on each slice to generate standardized 224×224 pixel blocks;

[0023] Five consecutive slices are combined into multi-channel input data.

[0024] Furthermore, the subjective quality assessment described in step S4 employs the following scoring mechanism:

[0025] The vessel wall scoring dimensions include vessel wall-cerebrospinal fluid contrast (1-4 points) and vessel wall structural clarity (1-4 points).

[0026] The blood flow inhibition score is divided into qualified (2 points) and unqualified (1 point) based on the homogeneity of the signal in the blood vessel lumen.

[0027] Motion artifact scoring dimensions are divided into no artifacts (3 points), mild artifacts (2 points), and severe artifacts (1 point) based on the degree of image ghosting.

[0028] Furthermore, in step S5, during the construction of the CNN model:

[0029] The backbone network adopts the convolutional layer structure of ResNet-18 and contains 8 residual modules;

[0030] Each fully connected layer is independently configured as follows: the vessel wall scoring network contains 512→128→4 nodes, the blood flow inhibition scoring network contains 512→64→2 nodes, and the motion artifact scoring network contains 512→64→3 nodes.

[0031] Furthermore, in step S6, the training set, validation set, and test set are randomly divided into layers in a 6:2:2 ratio; the optimizer uses the Adam algorithm, with an initial learning rate of 0.001 and a batch size of 16; the training period is 100 epochs, and the validation set performance is evaluated every 5 epochs.

[0032] Furthermore, the quality scoring results for each dimension in step S7 include:

[0033] The seven major vascular segments in the 3D whole-brain vascular wall imaging (SPACE sequence) were independently scored;

[0034] Local enhancement assessment of the target vessel segment in 2D vessel wall imaging (TSE sequence);

[0035] Output the classification probability distribution for each rating dimension and the final rating suggestion.

[0036] Compared with existing technologies, this system and method have the following advantages:

[0037] 1. This invention introduces a multi-task learning strategy into the network structure, integrates the correlation between blood vessel wall, blood flow inhibition and motion artifact scores, and dynamically balances the influence of the three on the backbone network training through a weight adaptive adjustment mechanism, thereby improving the overall scoring stability and generalization ability of the model. This type of joint multi-task scoring structure is the first of its kind to be proposed in intracranial VWI image quality control tasks.

[0038] 2. This invention introduces a spatial context-aware module into the model to enhance the modeling ability of spatial relationships between image slices and improve the model's accuracy in recognizing low-contrast blood vessel segments and motion artifact regions; the design of continuous slices simulates the visual judgment logic of radiologists on continuous blood vessel layers, which is different from the single-slice evaluation method; this invention is the first to introduce this structure as a multi-channel spatial context input, which has a significant effect on improving the consistency of scoring.

[0039] 3. This invention employs an interpolation and dynamic cropping strategy based on the vascular centerline direction, which can provide a consistent image input standardization mechanism for intracranial vascular systems with different individuals, angles, and complex structural curvatures. Compared with traditional static sampling methods, the accuracy in the quality control model is improved by more than 10%.

[0040] 4. The model constructed in this invention can replace senior physicians in performing automated subjective image quality scoring of intracranial artery MR-VWI images, solving the problems of time-consuming and poor repeatability of traditional manual quality control. It improves the efficiency and accuracy of clinical MR-VWI image quality control, helping clinicians to assess intracranial atherosclerotic plaques more quickly and accurately, and providing a reliable basis for clinical decision-making. Three CNN frameworks are proposed to evaluate the vessel wall, blood flow inhibition, and motion artifacts of intracranial artery MR-VWI respectively. Combined with clinical diagnostic requirements, the model highlights the key points of vessel wall quality assessment, avoids the previous general and single scoring methods, and enhances the pertinence and interpretability of image quality control, especially suitable for the intracranial vascular system with complex anatomical structure. Attached Figure Description

[0041] Figure 1 This is a flowchart of a deep learning-based method for quality control of intracranial vascular wall magnetic resonance images. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] like Figure 1The diagram shown is a flowchart of the method of the present invention. An embodiment of the present invention provides a method for quality control of intracranial vascular wall magnetic resonance images based on deep learning, comprising the following steps:

[0044] Step S1, Data Acquisition and Screening. Raw intracranial arterial MR-VWI data are obtained from the medical imaging archive system, and valid case data are screened based on preset exclusion criteria. The exclusion criteria include:

[0045] Imaging data from patients with non-atherosclerotic lesions were excluded;

[0046] Exclude cases that have already undergone vascular interventional treatment;

[0047] Exclude invalid data that is incompletely acquired or has severe artifacts.

[0048] Step S2, Imaging Examination. A 3.0T MRI scanner was used to perform multi-sequence scanning, including at least 3D TOF-MRA sequence, 3D black-blood T1WI SPACE sequence before and after contrast agent injection, and 2D black-blood T2WI TSE sequence;

[0049] Step S3, Image Preprocessing. The DICOM format image is anonymized, and slices are generated based on the vessel centerline. The image preprocessing specifically includes:

[0050] Centerline interpolation is performed on the target vessel segment;

[0051] Slices were taken at 2mm intervals along the centerline of the blood vessel.

[0052] Perform dynamic cropping on each slice to generate standardized 224×224 pixel blocks;

[0053] Five consecutive slices are combined into multi-channel input data.

[0054] Step S4, Subjective Quality Assessment. Two senior radiologists independently scored the quality of vessel wall visualization, blood flow suppression effect, and motion artifact impact; the subjective quality assessment adopted the following scoring mechanism:

[0055] The vessel wall scoring dimensions include vessel wall-cerebrospinal fluid contrast (1-4 points) and vessel wall structural clarity (1-4 points).

[0056] The blood flow inhibition score is divided into qualified (2 points) and unqualified (1 point) based on the homogeneity of the signal in the blood vessel lumen.

[0057] Motion artifact scoring dimensions are divided into no artifacts (3 points), mild artifacts (2 points), and severe artifacts (1 point) based on the degree of image ghosting.

[0058] Step S5, CNN model construction. A joint learning network based on the ResNet-18 architecture is established, which includes a shared backbone network and three independent fully connected layers, corresponding to vessel wall scoring (CNN-1), blood flow inhibition scoring (CNN-2), and motion artifact scoring (CNN-3), respectively; the backbone network adopts the convolutional layer structure of ResNet-18 and contains 8 residual modules;

[0059] Each fully connected layer is independently configured as follows: the vessel wall scoring network contains 512→128→4 nodes, the blood flow inhibition scoring network contains 512→64→2 nodes, and the motion artifact scoring network contains 512→64→3 nodes.

[0060] Step S6, Model Training. Stratified random sampling is used to divide the dataset into training, validation, and test sets. Network parameters are optimized by minimizing the mean squared error between predicted scores and doctor scores. The model training uses the following parameter configuration:

[0061] The training set, validation set, and test set are randomly divided into stratified groups in a 6:2:2 ratio.

[0062] The optimizer uses the Adam algorithm, with an initial learning rate of 0.001 and a batch size of 16.

[0063] The training period is 100 epochs, and the validation set performance is evaluated every 5 epochs.

[0064] Step S7, Quality Prediction. The preprocessed multi-channel image is input into the trained CNN model, which outputs quality scores for each dimension. Quality prediction includes:

[0065] The seven major vascular segments in the 3D whole-brain vascular wall imaging (SPACE sequence) were independently scored;

[0066] Local enhancement assessment of the target vessel segment in 2D vessel wall imaging (TSE sequence);

[0067] Output the classification probability distribution for each rating dimension and the final rating suggestion.

[0068] Example:

[0069] We selected 642 cases of intracranial arterial MR-VWI examinations performed between August 2020 and July 2021 from the PACS (Picture Archiving and Communication System) of the Department of Imaging Medicine, First Affiliated Hospital of Naval Medical University. After excluding cases with non-atherosclerotic lesions, cases that had undergone interventional treatment, and duplicate entries, 497 valid examination data were finally included.

[0070] All cases were examined using the same 3.0T MRI scanner (Siemens, Skyra, Erlangen, Germany) and a 20-channel head and neck coil. 3D TOF-MRA sequences, whole-brain diffusion-weighted (DWI) sequences were acquired sequentially, followed by canal wall imaging, including 3D T1WI SPACE plain and enhanced sequences, and 2D T2WI TSE sequences, each with specific scanning parameters.

[0071] The acquired images were exported in DICOM format and anonymized. The images were then displayed and scored using Horos V3.3.5 software on a MacBook Pro (13-inch, Mid 2014).

[0072] Two senior radiologists independently completed subjective image quality scoring for all 497 MR-VWI examinations, assessing vessel wall quality (1-4 points), blood flow inhibition (1-3 points), and motion artifacts (1-3 points). Six months later, 100 cases were randomly selected and re-scored by one of the senior radiologists. For vessel wall and blood flow inhibition assessment, 3D whole-brain vessel wall imaging was used to score the major intracranial vessels in seven segments, while 2D vessel wall imaging was used to assess the target vessel segment. Motion artifacts were scored across the entire sequence of images.

[0073] Based on the ResNet-18 convolutional neural network architecture, three subjective rating CNN prediction models were developed: vessel wall (CNN-1), blood flow inhibition (CNN-2), and motion artifacts (CNN-3). All models share a common backbone network structure, with independent fully connected layers at the ends for different tasks. Image sampling and interpolation of the centerline between the start and end slices of the arterial segment, as labeled by radiologists, yielded five 224*224 pixel image blocks, which were then used as multi-channel inputs to the CNN for image quality assessment. The network output underwent linear transformation and a stage to obtain the subjective ratings for each task. Training was conducted to minimize the mean squared error between the predicted and actual subjective ratings.

[0074] The dataset of 434 patients was randomly divided into a training set (299 cases, 60%), a validation set (99 cases, 20%), and a test set (99 cases, 20%). A CNN model was built using PyTorch 1.10 on an Ubuntu 18.04 LTS system, and training and testing were performed on a workstation equipped with an Intel Core i7-6700K CPU 4.0 GHz processor, 32GB DDR3 RAM, a Toshiba 1TB HDD, and an NVIDIA 3090 GPU with 24GB of memory. Statistical analysis was performed using R 4.1.1. Cohen's Kappa test was used to analyze the consistency of subjective scores for vessel wall and motion artifacts, and accuracy was used to evaluate the results of binary classification of blood flow inhibition.

[0075] This invention introduces a multi-task learning strategy into the network structure, integrating the correlations between vessel wall, blood flow inhibition, and motion artifact scores. A weight adaptive adjustment mechanism dynamically balances the impact of these three factors on the backbone network training, improving the overall scoring stability and generalization ability of the model. This type of joint multi-task scoring structure is the first of its kind proposed for intracranial VWI image quality control tasks. Furthermore, this invention introduces a spatial context-aware module to enhance the modeling ability of spatial relationships between image segments, improving the model's accuracy in identifying low-contrast vessel segments and motion artifact regions. The design of continuous segments simulates the visual judgment logic of radiologists on continuous vessel layers, differing from... Compared to single-slice evaluation methods, this invention introduces this structure as a multi-channel spatial context input for the first time, significantly improving scoring consistency. The invention employs an interpolation and dynamic cropping strategy based on the vessel centerline direction, providing a consistent image input standardization mechanism for intracranial vascular systems with different individuals, angles, and complex structural curvatures. Compared to traditional static sampling methods, the accuracy in the quality control model is improved by more than 10%. The constructed model maintains a scoring accuracy above 0.82 even in images with severe motion artifacts or complex bifurcation structures (such as skull base branches), demonstrating superior robustness and universality compared to traditional quality control methods. Compared with related technologies, the deep learning-based intracranial vascular wall magnetic resonance image quality control method provided by this invention has the following beneficial effects:

[0076] This invention provides a deep learning-based method for quality control of intracranial vascular wall magnetic resonance images. As shown in Table 1, the CNN model showed good consistency with senior physicians in subjective image quality assessment of intracranial arteries using MR-VWI. Regarding vascular wall quality scoring, in most vessels of the 3D T1WI SPACE plain scan (before contrast agent injection) sequence, the AI ​​model showed good or better inter-group consistency with senior physician A, with Cohen's Kappa values ​​ranging from 0.62 to 0.81. The highest consistency was observed in the right internal carotid artery (ICA) of the enhanced sequence, with a Kappa value of 0.81, reaching a level of "almost perfect consistency." Compared to the inter-group consistency between senior physicians A and B (Kappa values ​​mostly ranging from 0.41 to 0.65), the AI ​​model's scoring consistency in most vascular segments was significantly higher than that of the manual scoring. In the blood flow inhibition scoring, the AI ​​model achieved an accuracy of 0.73 to 0.96 in all vascular segments across the three sequences, except for the right middle cerebral artery (MCA) in the enhanced sequence (AI...). The AI ​​scoring accuracy for other areas was higher than that for senior physicians (0.73 for general practitioners and 0.76 for doctors). In terms of motion artifact scoring, the AI ​​scores for 3D T1WI SPACE plain scan and enhanced sequence showed good consistency with those for senior physicians, and were both higher than the inter-group consistency for senior physicians (Cohen's Kappa values ​​of 0.74 and 0.72, respectively).

[0077] Table 1. Consistency between AI and subjective ratings from senior physicians

[0078]

[0079] Note: ICA is the internal carotid artery, MCA is the middle cerebral artery, VA is the vertebral artery, and BA is the basilar artery.

[0080] This indicates that the model can accurately simulate the assessment scale of senior physicians and effectively reduce assessment differences caused by human factors. The model constructed in this invention can replace senior physicians in performing automated subjective image quality scoring of intracranial artery MR-VWI images, solving the problems of time-consuming and poor repeatability of traditional manual quality control, improving the efficiency and accuracy of clinical MR-VWI image quality control, and helping clinicians to assess intracranial atherosclerotic plaques more quickly and accurately, providing a reliable basis for clinical decision-making. Three CNN frameworks are proposed to evaluate the vessel wall, blood flow inhibition, and motion artifacts of intracranial artery MR-VWI respectively. Combined with clinical diagnostic requirements, the model highlights the key points of vessel wall quality assessment, avoids the previous general and single scoring methods, and enhances the pertinence and interpretability of image quality control, especially suitable for the anatomically complex intracranial vascular system.

[0081] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for quality control of intracranial vascular wall magnetic resonance images, characterized in that, Includes the following steps: Step S1: Obtain raw intracranial artery MR-VWI data from the medical imaging archive system, and screen valid case data based on preset exclusion criteria; Step S2: Perform multi-sequence scanning using a 3.0T MRI device. The sequences include at least a 3D TOF-MRA sequence, a 3D black-blood T1WI SPACE sequence before and after contrast agent injection, and a 2D black-blood T2WI TSE sequence. Step S3: Anonymize the DICOM format image, and perform slice sampling and image block generation based on the blood vessel centerline; In step S4, two senior radiologists independently scored the quality of the vessel wall visualization, the effect of blood flow suppression, and the impact of motion artifacts. Step S5: Construct a CNN model based on the ResNet-18 architecture. The model includes a shared backbone network and three independent fully connected layers, which correspond to the vessel wall score, blood flow inhibition score, and motion artifact score, respectively. Step S6: Stratified random sampling is used to divide the training set, validation set, and test set. The network parameters are optimized by minimizing the mean square error between the predicted score and the doctor's score. Step S7: Input the preprocessed multi-channel image into the trained model and output the quality score results for each dimension.

2. The method for quality control of intracranial vascular wall magnetic resonance images based on deep learning according to claim 1, characterized in that, The exclusion criteria include: Imaging data from patients with non-atherosclerotic lesions were excluded; Exclude cases that have already undergone vascular interventional treatment; Exclude invalid data that is incompletely acquired or has severe artifacts.

3. The method for quality control of intracranial vascular wall magnetic resonance images based on deep learning according to claim 1, characterized in that, The image preprocessing in step S3 includes: Centerline interpolation is performed on the target vessel segment; Slices were taken at 2mm intervals along the centerline of the blood vessel. Perform dynamic cropping on each slice to generate standardized 224×224 pixel blocks; Five consecutive slices are combined into multi-channel input data.

4. The method for quality control of intracranial vascular wall magnetic resonance images based on deep learning according to claim 1, characterized in that, The subjective quality assessment described in step S4 adopts the following scoring mechanism: The vessel wall scoring dimensions include vessel wall-cerebrospinal fluid contrast and vessel wall structural clarity; The blood flow inhibition score is categorized as qualified or unqualified based on the homogeneity of intravascular signal. Motion artifact scoring dimensions are categorized into no artifacts, mild artifacts, and severe artifacts based on the degree of image ghosting.

5. The method for quality control of intracranial vascular wall magnetic resonance images based on deep learning according to claim 1, characterized in that, In step S5, during the construction of the CNN model: The backbone network adopts the convolutional layer structure of ResNet-18 and contains 8 residual modules; Each fully connected layer is independently configured as follows: the vessel wall scoring network contains 512→128→4 nodes, the blood flow inhibition scoring network contains 512→64→2 nodes, and the motion artifact scoring network contains 512→64→3 nodes.

6. The method for quality control of intracranial vascular wall magnetic resonance images based on deep learning according to claim 1, characterized in that, In step S6, the training set, validation set, and test set are randomly divided into layers in a 6:2:2 ratio; the optimizer uses the Adam algorithm, with an initial learning rate of 0.001 and a batch size of 16; the training period is 100 epochs, and the validation set performance is evaluated every 5 epochs.

7. The method for quality control of intracranial vascular wall magnetic resonance images based on deep learning according to claim 1, characterized in that, The quality scoring results for each dimension in step S7 include: The seven major vascular segments in the 3D whole-brain vascular wall imaging were scored independently. Local enhancement assessment of the target vascular segment in 2D vessel wall imaging; Output the classification probability distribution for each rating dimension and the final rating suggestion.

Images