Lower limb deep venous thrombosis ultrasonic diagnosis method and system based on deep learning

Abstract

The invention discloses a lower limb deep venous thrombosis ultrasonic diagnosis method and system based on deep learning. The method comprises the following steps: S1, acquiring basic ultrasonic image data of a lower limb deep vein area in an unpressed state and contrast ultrasonic image data of the lower limb deep vein area in a pressed state; s2, inputting the basic ultrasonic image data and the contrast ultrasonic image data into a pre-trained lower limb deep vein thrombosis diagnosis model to obtain a pressed state of a vein; and S3, determining a diagnosis result of the deep venous thrombosis of the lower limb according to the pressed state of the vein. According to the method, the diagnosis efficiency is improved, the DVT can be quickly and accurately diagnosed through the deep learning technology, and the waiting time of a patient is shortened; according to the invention, the diagnosis convenience is increased: non-experts can carry out DVT screening in a common medical environment and a remote nursing point, so that the diagnosis convenience is increased; diagnosis accuracy is improved, a diagnosis result with high accuracy is provided through a machine learning model, and sensitivity and specificity both reach clinical practical standards.

Description

Technical Field

[0001] This invention relates to the field of medical imaging diagnostic technology, and more specifically, to a deep learning-based ultrasound diagnostic method and system for deep vein thrombosis of the lower extremities. Background Technology

[0002] Deep vein thrombosis (DVT) of the lower extremities is a serious medical condition, referring to the formation of a blood clot in the deep veins of the lower extremities. This condition can cause pain, swelling, and in severe cases, pulmonary embolism, a life-threatening emergency. Therefore, rapid and accurate diagnosis of DVT is crucial for timely treatment and prevention of complications.

[0003] Traditional DVT diagnosis relies on the physician's experience and the use of ultrasound equipment. This method is not only time-consuming but also requires a high level of skill from the operator. With the increasing number of DVT cases, the demand for specialist physicians is also increasing, making DVT diagnosis in non-specialist settings particularly important. However, non-specialist settings often lack professional ultrasound equipment and diagnostic physicians, limiting the accessibility and convenience of DVT diagnosis.

[0004] Therefore, it is indeed necessary to provide a deep learning-based ultrasound diagnostic method and system for deep vein thrombosis of the lower extremities. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based ultrasound diagnostic method and system for deep vein thrombosis of the lower extremities, so as to overcome the defects of the existing technology.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A deep learning-based ultrasound diagnostic method for lower extremity deep vein thrombosis includes the following steps:

[0008] S1. Collect baseline ultrasound imaging data of the deep vein region of the lower extremities under uncompressed conditions and comparative ultrasound imaging data under compressed conditions;

[0009] S2. Input the basic ultrasound image data and the contrast ultrasound image data into the pre-trained lower extremity deep vein thrombosis diagnostic model to obtain the compression status of the veins.

[0010] S3. Determine the diagnosis of deep vein thrombosis in the lower extremities based on the compression status of the veins.

[0011] Furthermore, the training steps for the lower extremity deep vein thrombosis diagnostic model in step S2 include:

[0012] S20. Collect baseline ultrasound imaging data of the deep vein region of the lower extremities in an uncompressed state and comparative ultrasound imaging data in a compressed state from patients with deep vein thrombosis of the lower extremities and healthy controls.

[0013] S21. Preprocess the basic ultrasound image data and the contrast ultrasound image data, and annotate the location of veins, arteries and the state of vein compression to form an annotated training dataset.

[0014] S22. Use the training dataset to train a diagnostic model for deep vein thrombosis in the lower extremities.

[0015] Furthermore, the diagnostic model for deep vein thrombosis in the lower extremities includes:

[0016] The backbone network is used to extract features from ultrasound images;

[0017] Branch networks connecting the main network are used to predict the compression status of veins;

[0018] The diagnostic model for deep vein thrombosis in the lower extremities is an improved U-Net model. The U-Net model consists of a backbone network and branch networks. The U-Net model has feature enhancement modules after the encoder and the convolutional layer of the bottleneck layer.

[0019] Furthermore, the preprocessing step S21 for the basic ultrasound image data and the contrast ultrasound image data includes:

[0020] S210. Clean and format the basic ultrasound image data of the lower extremity deep vein region under uncompressed conditions and the comparative ultrasound image data under compressed conditions.

[0021] S211. Use an automatic or semi-automatic image registration algorithm to spatially align images of the same region sequence and perform region segmentation according to the set vein and artery ROIs;

[0022] S212. Perform image enhancement on the image after region segmentation;

[0023] S213. Perform frame-by-frame annotation on the enhanced image. The frame-by-frame annotation includes the outline of veins and arteries, the direction of multi-level blood vessels, the compression status of veins, the presence of thrombi, and echo characteristics.

[0024] S214. Enhance the training data labeled frame by frame in various ways or use virtual compressed images under simulated different probe pressure conditions to enhance sample diversity, and standardize the resolution and pixel intensity values ​​of all samples.

[0025] Furthermore, the lower extremity deep vein thrombosis diagnostic model in step S22 includes:

[0026] The input layer is designed such that each pair of training samples contains two input frames, one unpressed and one pressurized, which are used as dual-channel or sequence fusion inputs. Alternatively, the two frames can be input into the network separately and feature convergence can be performed through early fusion or late fusion structures.

[0027] The backbone network adopts a multi-level convolutional structure, downsampling layer by layer to extract spatial and textural features of the image. Multi-scale convolutional modules or attention mechanisms are integrated after the encoder and bottleneck layer to effectively capture the direction and morphological changes of blood vessels at different resolutions.

[0028] By utilizing the anatomical structure of the outgoing veins, arteries, and vessel walls, and employing the discriminant branch fusion of pre- and post-compression characteristics, as well as the probability distribution of whether the outgoing veins are completely compressed, a quantitative classification of venous compression status is achieved, including complete compression, incomplete compression, no compression, suspicious, and thrombosis.

[0029] The Dice loss, cross-entropy loss, and pair consistency loss designed specifically for dynamic sequence discrimination are used in combination.

[0030] Choose AdamW or Ranger as the optimizer and use multi-fold cross-validation.

[0031] The present invention also provides a system based on the above-described deep learning-based ultrasound diagnostic method for deep vein thrombosis of the lower extremities, comprising:

[0032] The acquisition module is used to acquire basic ultrasound imaging data of the deep vein region of the lower extremities under uncompressed conditions and comparative ultrasound imaging data under compressed conditions.

[0033] The diagnostic module is used to input basic ultrasound imaging data and contrast ultrasound imaging data into a pre-trained diagnostic model for deep vein thrombosis in the lower extremities to obtain the compression status of the veins.

[0034] The output module is used to determine the diagnostic result of deep vein thrombosis in the lower extremities based on the compression status of the veins.

[0035] Compared with existing technologies, the advantages of this invention are as follows: This invention improves diagnostic efficiency by using deep learning technology to quickly and accurately diagnose DVT, reducing patient waiting time; This invention increases diagnostic convenience by allowing non-experts to perform DVT screening in ordinary medical environments and remote care points, increasing diagnostic convenience; This invention improves diagnostic accuracy by providing highly accurate diagnostic results through machine learning models, with sensitivity and specificity meeting clinically applicable standards; This invention reduces medical costs by reducing unnecessary referrals and increasing net monetary benefits. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of the ultrasound diagnostic method for deep vein thrombosis in the lower extremities based on deep learning, as described in this invention.

[0038] Figure 2 This is a schematic diagram of the deep learning-based ultrasound diagnostic system for deep vein thrombosis in the lower extremities, based on the present invention. Detailed Implementation

[0039] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0040] See Figure 1 As shown in the figure, this embodiment discloses a deep learning-based ultrasound diagnostic method for lower extremity deep vein thrombosis, including the following steps:

[0041] Step S1: Collect baseline ultrasound imaging data of the deep vein region of the lower extremities under uncompressed conditions and comparative ultrasound imaging data under compressed conditions.

[0042] In this embodiment, an ultrasound probe is used to perform an ultrasound examination on the subject, and to collect basic ultrasound image data of the subject's lower limb deep vein areas, such as the groin and behind the knee, in an uncompressed state, as well as comparative ultrasound image data after applying external pressure, such as manually compressing the probe.

[0043] Step S2: Input the basic ultrasound image data and the contrast ultrasound image data into the pre-trained lower extremity deep vein thrombosis diagnostic model to obtain the compression status of the veins.

[0044] The training steps for the lower extremity deep vein thrombosis diagnostic model include:

[0045] Step S20: Collect baseline ultrasound imaging data of the deep vein region of the lower extremities in the uncompressed state and comparative ultrasound imaging data in the compressed state of patients with deep vein thrombosis of the lower extremities and healthy controls.

[0046] In this embodiment, baseline ultrasound images of the lower extremity deep vein region in an uncompressed state and contrasting ultrasound images in a compressed state were collected from suspected DVT patients and healthy controls. These ultrasound images were preprocessed and labeled to form a labeled training dataset. The labels included vein location, artery location, and vein compression status. This training dataset was used to train a lower extremity deep vein thrombosis diagnostic model to identify venous compression status.

[0047] The data preprocessing process includes:

[0048] Step S210: Clean and format the basic ultrasound image data of the lower extremity deep vein region under uncompressed conditions and the comparative ultrasound image data under compressed conditions. Remove and clean up some image samples with poor imaging quality, blurriness, obvious artifacts or missing information to ensure that the data files of each subject are complete and uniformly stored as standard DICOM or PNG / JPG format files for subsequent analysis.

[0049] Step S211: To ensure that the same anatomical structure can be aligned in both uncompressed and compressed states, an automatic or semi-automatic image registration algorithm is used to spatially align the sequence of images of the same region. Based on clinical experience, the region is cut according to the set ROI of veins and arteries to remove irrelevant parts and highlight the target structure to the maximum extent.

[0050] Step S212: Image enhancement is performed on the segmented image. Ultrasound imaging is prone to low signal-to-noise ratio due to probe pressure, depth, etc. Image histogram equalization, adaptive contrast enhancement, and salt-and-pepper noise removal techniques are used to improve the clarity of blood vessel and thrombus boundaries. For dark-field images, brightness and contrast can be adjusted appropriately to make the boundaries between arteries and veins and surrounding tissues more distinct. This embodiment can use medical image enhancement algorithms based on Retinex or GAN to automatically compensate for uneven illumination and missing structural details.

[0051] Step S213: A professional ultrasound physician annotates the enhanced image frame by frame. This annotation includes the outline of veins and arteries, the course of multi-level blood vessels, the compression status of veins (completely collapsed / partially collapsed / not collapsed), the presence of thrombi (present / absent / suspicious), and echo characteristics (increased echo / hypoechoic / foreign body echo, etc.). All annotations are output as binary masks using medical annotation tools such as LabelMe, ITK-SNAP, or MedSeg and paired with the original images.

[0052] Step S214: To improve the robustness of the model, various enhancements are performed on the frame-by-frame labeled training data (such as geometric translation, rotation, scale transformation, horizontal mirroring, random cropping, noise addition, etc.). Virtual compressed images simulating different probe pressure conditions can also be used to enhance sample diversity. All sample resolution and pixel intensity values ​​are standardized (Z-score normalization) so that the deep network receives consistent input.

[0053] This embodiment designs a custom edge enhancement or vascular direction filtering module to enhance the contrast of subtle changes in blood vessels, targeting the high and low echo distribution of ultrasound images. Dynamic pairwise annotation is introduced to automatically / semi-automatically pair images of the same location before and after compression, driving the model to actively focus on dynamic changes under compression rather than static structures, which better meets the actual needs of compression ultrasound scenarios.

[0054] Step S21: Preprocess the basic ultrasound image data and the contrast ultrasound image data, and annotate the vein location, artery location and vein compression status to form an annotated training dataset.

[0055] Step S22: Train the lower extremity deep vein thrombosis diagnostic model using the training dataset.

[0056] In this embodiment, the diagnostic model for deep vein thrombosis in the lower extremities includes:

[0057] The backbone network is used to extract features from ultrasound images;

[0058] Branch networks connecting the main network are used to predict the compression status of veins;

[0059] The diagnostic model for deep vein thrombosis in the lower extremities is an improved U-Net model. The U-Net model consists of a backbone network and branch networks. The U-Net model has feature enhancement modules after the encoder and the convolutional layer of the bottleneck layer.

[0060] In this embodiment, the lower extremity deep vein thrombosis diagnostic model in step S22 includes:

[0061] The input layer is designed such that each pair of training samples contains two input frames, one unpressed and one pressurized, which are used as dual-channel or sequence fusion inputs. Alternatively, the two frames can be input into the network separately and feature convergence can be performed through early fusion or late fusion structures.

[0062] For feature extraction and enhancement, the backbone network adopts a multi-level convolutional structure (U-Net encoder), which downsamples layer by layer to extract spatial and textural features of the image. Multi-scale convolutional modules (such as Inception block, Atrous spatial pyramid pooling) or attention mechanisms (SE-Block, CBAM, Self-attention) are integrated after the encoder and bottleneck layer to effectively capture the direction and morphological changes of blood vessels at different resolutions, and are particularly sensitive to dynamic compression details.

[0063] Branching and segmentation discrimination uses the anatomical structure of the output veins, arteries, and vessel walls of the structural branches, and uses the discriminative branches to fuse pre- and post-compression features and the probability distribution of whether the output vein is completely compressed, to achieve quantitative classification of vein compression status, including complete compression, incomplete compression, no compression, suspicious, and thrombosis;

[0064] The loss function is set by jointly employing Dice loss, cross-entropy loss, and pairwise consistency loss, which is specifically designed for dynamic sequence discrimination. This ensures both the accuracy of structural segmentation and enhances the model's ability to discriminate dynamic stress. Innovatively, a temporal consistency loss based on sequence differences can be explored, constraining the model to identify "the region with the greatest change before and after compression as the high incidence point of DVT".

[0065] Training details: The optimizer can be AdamW or Ranger, and the learning rate is dynamically adjusted using the Cosine Annealing Schedule.

[0066] Folded cross-validation is used to ensure the model's generalization ability.

[0067] It dynamically monitors multiple metrics such as sensitivity, specificity, and AUC, as well as the mIoU and Dice scores for structural segmentation. An early stopping strategy can also be enabled to prevent overfitting.

[0068] This embodiment introduces a "Hard Example Mining" mechanism, which increases the training weights for collapsed and non-collapsed samples that the model misclassifies, thereby strengthening the learning of features of marginal cases.

[0069] By combining a small amount of unlabeled data, semi-supervised or self-supervised methods can be used to improve the model's data utilization and adaptability to rare clinical situations.

[0070] The model output can be used in conjunction with interpretability analysis (such as Grad-CAM and significance plots) to help doctors understand the basis for their judgments.

[0071] Model validation and deployment involve testing on independent validation sets and real clinical samples, outputting clinically relevant metrics such as accuracy, sensitivity, specificity, and AUC. The final deployment is as an API or integrated into medical systems, supporting batch interpretation and real-time feedback, facilitating application in remote or general medical institutions.

[0072] Step S3: Determine the diagnosis of deep vein thrombosis in the lower extremities based on the compression status of the veins.

[0073] The collected baseline and contrast ultrasound images were input into a pre-trained lower extremity deep vein thrombosis (DVT) diagnostic model to obtain the venous compression status. The diagnosis of DVT was determined based on the venous compression status. Table 1 below shows the evaluation results of different model architectures.

[0074] Table 1 Evaluation results of different model architectures

[0075] Model Architecture Sensitivity Specificity AUC Embodiments of the present invention Improved U-NET 0.92 0.91 0.95 Comparative Example 1 U-NET 0.88 0.89 0.90 Comparative Example 2 NN-U-NET 0.87 0.88 0.89

[0076] This invention improves diagnostic efficiency by using deep learning technology to quickly and accurately diagnose deep vein thrombosis (DVT), reducing patient waiting time. It also increases diagnostic convenience: non-experts can perform DVT screening in ordinary medical environments and remote care points, enhancing diagnostic accessibility. Furthermore, it improves diagnostic accuracy by providing highly accurate diagnostic results through machine learning models, with sensitivity and specificity meeting clinically applicable standards. Finally, it reduces healthcare costs by enabling rapid diagnosis and minimizing unnecessary referrals, thereby increasing net monetary benefits.

[0077] See Figure 2 As shown, the present invention also provides a system based on the above-described deep learning-based ultrasound diagnostic method for deep vein thrombosis of the lower extremities, comprising:

[0078] Acquisition module 1 is used to acquire baseline ultrasound imaging data of the deep vein region of the lower extremities under uncompressed conditions and comparative ultrasound imaging data under compressed conditions.

[0079] Diagnostic module 2 is used to input basic ultrasound image data and contrast ultrasound image data into a pre-trained lower extremity deep vein thrombosis diagnostic model to obtain the compression status of the veins.

[0080] Output module 3 is used to determine the diagnostic result of deep vein thrombosis in the lower extremities based on the compression status of the vein.

[0081] This invention improves diagnostic efficiency by using deep learning technology to quickly and accurately diagnose deep vein thrombosis (DVT), reducing patient waiting time. It also increases diagnostic convenience: non-experts can perform DVT screening in ordinary medical environments and remote care points, enhancing diagnostic accessibility. Furthermore, it improves diagnostic accuracy by providing highly accurate diagnostic results through machine learning models, with sensitivity and specificity meeting clinically applicable standards. Finally, it reduces healthcare costs by enabling rapid diagnosis and minimizing unnecessary referrals, thereby increasing net monetary benefits.

[0082] The invention will be further described below with reference to clinical application cases.

[0083] Application Example 1: Application in a General Medical Environment

[0084] Patient C, a 60-year-old male, presented with lower extremity pain and swelling, suspected of having deep vein thrombosis (DVT). An examination was performed using a portable ultrasound device, guided by the deep learning model of this invention, using a two-point compression ultrasound examination. The examination results showed complete venous compression, with no DVT. The patient received the diagnosis directly in a standard medical setting without referral. The total examination time was approximately 10 minutes.

[0085] Application Example 2: Application in Remote Care Points

[0086] Patient D, a 70-year-old female with a history of deep vein thrombosis (DVT), currently residing in a nursing home, presented with lower extremity swelling. Nursing staff performed an examination using a portable ultrasound device, guided by the deep learning model of this invention, using a three-point compression ultrasound examination. The examination results showed incomplete compression of the femoral vein, suggesting DVT. The patient was referred to a specialist hospital for further diagnosis. The total examination time was approximately 15 minutes.

[0087] Application Example 3: Multicenter Clinical Trials

[0088] A multicenter clinical trial was conducted at three DVT diagnostic centers in China, recruiting a total of 90 patients suspected of having DVT. The deep learning model of this invention was used for examination, and the results showed a sensitivity of 0.85-0.98, a specificity of 0.75-0.85, a positive predictive value of 0.70-0.90, and a negative predictive value of 0.95-1.00.

[0089] The experimental results show that the present invention has good application effects in different clinical environments and can significantly improve the diagnostic accuracy and efficiency of DVT.

[0090] Although embodiments of the present invention have been described in conjunction with the accompanying drawings, the patent owner may make various modifications or alterations within the scope of the appended claims, as long as they do not exceed the protection scope described in the claims of the present invention, they shall be within the protection scope of the present invention.

Claims

1. A deep learning-based ultrasound diagnostic method for deep vein thrombosis in the lower extremities, characterized in that, Includes the following steps: S1. Collect baseline ultrasound imaging data of the deep vein region of the lower extremities under uncompressed conditions and comparative ultrasound imaging data under compressed conditions; S2. Input the basic ultrasound image data and the contrast ultrasound image data into the pre-trained lower extremity deep vein thrombosis diagnostic model to obtain the compression status of the veins. S3. Determine the diagnosis of deep vein thrombosis in the lower extremities based on the compression status of the veins.

2. The ultrasound diagnostic method for lower extremity deep vein thrombosis based on deep learning according to claim 1, characterized in that, The training steps for the lower extremity deep vein thrombosis diagnostic model in step S2 include: S20. Collect baseline ultrasound imaging data of the deep vein region of the lower extremities in an uncompressed state and comparative ultrasound imaging data in a compressed state from patients with deep vein thrombosis of the lower extremities and healthy controls. S21. Preprocess the basic ultrasound image data and the contrast ultrasound image data, and annotate the location of veins, arteries and the state of vein compression to form an annotated training dataset. S22. Use the training dataset to train a diagnostic model for deep vein thrombosis in the lower extremities.

3. The ultrasound diagnostic method for lower extremity deep vein thrombosis based on deep learning according to claim 1, characterized in that, The diagnostic model for deep vein thrombosis in the lower extremities includes: The backbone network is used to extract features from ultrasound images; Branch networks connecting the main network are used to predict the compression status of veins; The diagnostic model for deep vein thrombosis in the lower extremities is an improved U-Net model. The U-Net model consists of a backbone network and branch networks. The U-Net model has feature enhancement modules after the encoder and the convolutional layer of the bottleneck layer.

4. The ultrasound diagnostic method for lower extremity deep vein thrombosis based on deep learning according to claim 2, characterized in that, The preprocessing steps for the basic ultrasound image data and the contrast ultrasound image data in step S21 include: S210. Clean and format the basic ultrasound image data of the lower extremity deep vein region under uncompressed conditions and the comparative ultrasound image data under compressed conditions. S211. Use an automatic or semi-automatic image registration algorithm to spatially align images of the same region sequence and perform region segmentation according to the set vein and artery ROIs; S212. Perform image enhancement on the image after region segmentation; S213. Perform frame-by-frame annotation on the enhanced image. The frame-by-frame annotation includes the outline of veins and arteries, the direction of multi-level blood vessels, the compression status of veins, the presence of thrombi, and echo characteristics. S214. Enhance the training data labeled frame by frame in various ways or use virtual compressed images under simulated different probe pressure conditions to enhance sample diversity, and standardize the resolution and pixel intensity values ​​of all samples.

5. The ultrasound diagnostic method for lower extremity deep vein thrombosis based on deep learning according to claim 2, characterized in that, The lower extremity deep vein thrombosis diagnostic model in step S22 includes: The input layer is designed such that each pair of training samples contains two input frames, one unpressed and one pressurized, which are used as dual-channel or sequence fusion inputs. Alternatively, the two frames can be input into the network separately and feature convergence can be performed through early fusion or late fusion structures. The backbone network adopts a multi-level convolutional structure, downsampling layer by layer to extract spatial and textural features of the image. Multi-scale convolutional modules or attention mechanisms are integrated after the encoder and bottleneck layer to effectively capture the direction and morphological changes of blood vessels at different resolutions. By utilizing the anatomical structure of the outgoing veins, arteries, and vessel walls, and employing the discriminant branch fusion of pre- and post-compression characteristics, as well as the probability distribution of whether the outgoing veins are completely compressed, a quantitative classification of venous compression status is achieved, including complete compression, incomplete compression, no compression, suspicious, and thrombosis. The Dice loss, cross-entropy loss, and pair consistency loss designed specifically for dynamic sequence discrimination are used in combination. Choose AdamW or Ranger as the optimizer and use multi-fold cross-validation.

6. A system for ultrasound diagnosis of lower extremity deep vein thrombosis based on deep learning according to any one of claims 1-5, characterized in that, include: The acquisition module is used to acquire basic ultrasound imaging data of the deep vein region of the lower extremities under uncompressed conditions and comparative ultrasound imaging data under compressed conditions. The diagnostic module is used to input basic ultrasound imaging data and contrast ultrasound imaging data into a pre-trained diagnostic model for deep vein thrombosis in the lower extremities to obtain the compression status of the veins. The output module is used to determine the diagnostic result of deep vein thrombosis in the lower extremities based on the compression status of the veins.

Images