A non-invasive diagnosis method and system for deep vein thrombosis based on machine learning

Abstract

The present invention provides a non-invasive diagnosis method and system for deep vein thrombosis based on machine learning, which belongs to the technical field of deep vein thrombosis diagnosis, and includes: step S1, collecting first ultrasonic image data of a preset inspection site of a subject in an uncompressed state and second ultrasonic image data in a compressed state; step S2, inputting the first ultrasonic image data and the second ultrasonic image data into a pre-established deep vein thrombosis prediction model to obtain the compression state of the vein; step S3, determining the deep vein thrombosis diagnosis result according to the compression state. Beneficial effects: Based on the machine learning technology, the compression state of the vein is predicted according to the collected ultrasonic image data before and after compression, and then the deep vein thrombosis diagnosis result is determined, which reduces the patient waiting time, provides rapid diagnosis, reduces the burden on experts, and performs reliable DVT examinations in ordinary medical environments and remote care points.

Description

Technical Field

[0001] The present invention relates to the technical field of deep vein thrombosis diagnosis, and in particular to a non-invasive deep vein thrombosis diagnosis method and system based on machine learning. Background Art

[0002] Deep vein thrombosis (DVT) is a process in which blood coagulates or certain formed elements in the blood aggregate to form solid masses in the deep veins. DVT usually occurs in the lower limbs, such as the pelvis, thigh, and calf, so lower limb DVT is divided into calf DVT and iliofemoral DVT. Most symptoms of calf DVT are mild, with the main clinical manifestations of calf pain and mild swelling, pain when the foot dorsiflexes or compresses the back muscles of the calf, and can generally be ablated or autolyzed with appropriate treatment. Iliofemoral DVT may be confined to the deep iliofemoral veins or may fill the deep veins of the entire lower limb. The main clinical manifestations are swelling below the buttocks, distended superficial veins in the lower limbs, groin, and affected side of the abdominal wall, and increased skin temperature. Once the blood clot breaks and falls off the venous wall, it will move with the blood flow and may attach to the lungs, causing acute pulmonary embolism (PE). Therefore, timely and accurate diagnosis of DVT is crucial for the treatment and prognosis of patients.

[0003] In medical practice, compression ultrasound examination of the leg is the gold standard for diagnosing DVT. This examination method uses ultrasound technology to observe the changes in blood flow when the leg veins are squeezed by external force, thereby determining whether there is thrombosis.

[0004] However, the diagnostic method of DVT mainly relies on compressed ultrasound examinations performed by professional and highly skilled radiologists or cardiologists, resulting in many patients who may have DVT symptoms facing long waiting times when seeking diagnosis, and even delaying the best treatment time. Secondly, with the increasing number of DVT cases, the workload of expert physicians is increasing, so non-expert diagnosis at the point of care is very necessary. However, for ordinary medical environments or non-expert remote care points (such as nursing homes and patients' homes), reliable DVT examinations cannot be performed due to the lack of professional ultrasound equipment and diagnostic physicians, which limits the popularity and convenience of diagnosis. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a non-invasive diagnosis method for deep vein thrombosis based on machine learning; on the other hand, it also provides a non-invasive diagnosis system for deep vein thrombosis based on machine learning.

[0006] The technical problem solved by the present invention can be achieved by adopting the following technical solutions:

[0007] The first aspect of the present invention is to provide a non-invasive diagnosis method for deep vein thrombosis based on machine learning, comprising:

[0008] Step S1, collecting first ultrasonic image data of a preset examination part of a subject in an uncompressed state and second ultrasonic image data in a compressed state;

[0009] Step S2, inputting the first ultrasound image data and the second ultrasound image data into a pre-established deep vein thrombosis prediction model to obtain a compression state of the vein;

[0010] Step S3, determining a deep vein thrombosis diagnosis result according to the compression state.

[0011] Preferably, the step S1 comprises:

[0012] Step S11, placing an ultrasonic probe at the preset inspection site and scanning along the vein direction to obtain a real-time image;

[0013] Step S12, inputting the real-time image into an anatomical position recognition model to recognize the real-time image and obtain an anatomical position recognition result;

[0014] Step S13, performing a preliminary scan according to the anatomical position recognition result, and collecting first ultrasonic image data of veins and arteries in an uncompressed state;

[0015] Step S14, gradually increasing the pressure of the ultrasound probe, and performing scanning again to collect second ultrasound image data of veins and arteries in a compressed state.

[0016] Preferably, the training step of the deep vein thrombosis prediction model includes:

[0017] Step S201, using suspected deep vein thrombosis patients and healthy volunteers as subjects, obtaining first ultrasound image data of a preset examination site of the subjects in an uncompressed state and second ultrasound image data in a compressed state;

[0018] Step S202, preprocessing the first ultrasound image data and the second ultrasound image data, and labeling them to obtain a training data set with labels, wherein the labels include a vein position label, an artery position label, and a vein compression state label;

[0019] Step S203, using the ultrasound image data of the training data set as model input, and the label as model output, to train and obtain the deep vein thrombosis prediction model.

[0020] Preferably, the deep vein thrombosis prediction model comprises:

[0021] Backbone network, used to extract image features;

[0022] The branch network is connected to the trunk network and is used to predict the compression state of the vein.

[0023] Preferably, the deep vein thrombosis prediction model is an improved U-Net model, the trunk network and the branch network constitute a U-Net model, and an enhanced feature representation module is respectively provided after the convolution layer of the encoder and the bottleneck layer of the U-Net model.

[0024] Preferably, the backbone network includes:

[0025] An input layer, configured to receive the first ultrasonic image data and the second ultrasonic image data;

[0026] The encoder includes a plurality of first convolution blocks, each of which includes a first convolution layer, a second convolution layer, an enhanced feature representation module and a maximum pooling layer connected in sequence, and is used to extract image features layer by layer.

[0027] Preferably, the branch network includes:

[0028] The bottleneck layer includes the third convolutional layer, the fourth convolutional layer and the enhanced feature representation module connected in sequence;

[0029] A decoder, comprising a plurality of second convolution blocks, each of which comprises a deconvolution layer, a splicing layer, a fifth convolution layer, and a sixth convolution layer connected in sequence, wherein the splicing layer is jump-connected to the corresponding first convolution block of the encoder, for restoring the position information of the image;

[0030] The output layer, including the seventh convolutional layer, is used to output the compression state of the vein.

[0031] Preferably, the enhanced feature representation module is a multi-scale convolutional layer or an attention layer.

[0032] Preferably, step S3 comprises:

[0033] If the compression state is completely compressed, determining that the deep vein thrombosis diagnosis result is no deep vein thrombosis;

[0034] If the compression state is incomplete compression, the deep vein thrombosis diagnosis result is determined to be suspected deep vein thrombosis.

[0035] The second aspect of the present invention is to provide a non-invasive diagnosis system for deep vein thrombosis based on machine learning, which is used to implement the non-invasive diagnosis method for deep vein thrombosis based on machine learning as described above, and the system includes:

[0036] A data acquisition module, used for acquiring first ultrasonic image data of a preset examination part of a subject in an uncompressed state and second ultrasonic image data in a compressed state;

[0037] a model prediction module, connected to the data acquisition module, and used for inputting the first ultrasound image data and the second ultrasound image data into a pre-established deep vein thrombosis prediction model to obtain a compression state of the vein;

[0038] A diagnosis module is connected to the model prediction module and is used to determine a diagnosis result of deep vein thrombosis according to the compression state.

[0039] The advantages or beneficial effects of the technical solution of the present invention are:

[0040] The present invention collects first ultrasonic image data of the examination part of the subject in an uncompressed state and second ultrasonic image data in a compressed state, predicts the compression state of the vein based on machine learning technology, and then determines the diagnosis result of deep vein thrombosis, reduces patient waiting time, provides rapid diagnosis, reduces the burden on experts, and performs reliable DVT examinations in ordinary medical environments and remote care points. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of a non-invasive method for diagnosing deep vein thrombosis based on machine learning in a preferred embodiment of the present invention;

[0042] Figure 2 This is a flow chart of step S1 in a preferred embodiment of the present invention;

[0043] Figure 3 A schematic diagram of the flow chart of the deep vein thrombosis prediction model training steps in a preferred embodiment of the present invention;

[0044] Figure 4 A schematic diagram of a deep vein thrombosis prediction model in a preferred embodiment of the present invention;

[0045] Figure 5 This is a structural block diagram of a non-invasive deep vein thrombosis diagnosis system based on machine learning in a preferred embodiment of the present invention;

[0046] Figure 6 A schematic diagram of the femoral vein in an uncompressed state in a preferred embodiment of the present invention;

[0047] Figure 7 Schematic diagram of the femoral vein in a compressed state in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0049] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0050] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0051] In a preferred embodiment of the present invention, based on the above problems existing in the prior art, a non-invasive diagnosis method for deep vein thrombosis based on machine learning is provided. In the model application prediction stage, Figure 1 As shown, including:

[0052] Step S1, collecting first ultrasonic image data of a preset examination part of a subject in an uncompressed state and second ultrasonic image data in a compressed state;

[0053] Step S2, inputting the first ultrasound image data and the second ultrasound image data into a pre-established deep vein thrombosis prediction model to obtain a compression state of the vein;

[0054] Step S3, determining a deep vein thrombosis diagnosis result according to the compression state.

[0055] Specifically, first, an ultrasonic examination is performed on the subject using an ultrasonic probe to collect first ultrasonic image data of a preset examination site of the subject, such as the groin or behind the knee, in an uncompressed state, and second ultrasonic image data after external pressure is applied, such as manual compression through the probe.

[0056] The present invention collects ultrasonic image data before and after compression, predicts the compression state of the vein based on machine learning technology, and then determines the diagnosis result of deep vein thrombosis, reduces patient waiting time, provides rapid diagnosis, reduces the burden on experts, and conducts reliable DVT examinations in ordinary medical environments and remote care points.

[0057] As a preferred embodiment, wherein Figure 2 As shown, step S1 includes:

[0058] Step S11, placing the ultrasound probe at a preset inspection site and scanning along the vein direction to obtain a real-time image;

[0059] Step S12, inputting the real-time image into the anatomical position recognition model to recognize the real-time image and obtain the anatomical position recognition result;

[0060] Step S13, performing a preliminary scan according to the anatomical position recognition result, and collecting first ultrasonic image data of veins and arteries in an uncompressed state;

[0061] Step S14, gradually increasing the pressure of the ultrasound probe, and performing scanning again to collect second ultrasound image data of veins and arteries in a compressed state.

[0062] Specifically, the preset inspection sites include the groin and the back of the knee. Ultrasound inspection includes inspection of the groin and the back of the knee.

[0063] Among them, the groin examination includes four steps: position identification, preliminary scanning, vein compression and model judgment. First, the ultrasound probe is placed in the groin area and scanned along the femoral vein. The anatomical position of the femoral vein is identified through real-time image processing; then, a preliminary scan is performed to obtain cross-sectional images of the vein and artery; then, the probe pressure is gradually increased to observe the compression of the vein and record the images before and after compression; finally, the acquired images are input into the anatomical position recognition model pre-built based on machine learning technology to determine the compression state of the vein. If the vein is completely compressed, it is judged as no DVT, otherwise it is suspected DVT.

[0064] Similarly, the examination behind the knee also includes four steps: position identification, preliminary scanning, vein compression, and model judgment. First, the ultrasound probe is placed behind the knee and scanned along the popliteal vein. The anatomical position of the popliteal vein is identified through real-time image processing. Then, a preliminary scan is performed to obtain cross-sectional images of the vein and artery. Next, the probe pressure is gradually increased to observe the compression of the vein and record the images before and after compression. Finally, the acquired images are input into the machine learning model to determine the compression state of the vein. If the vein is completely compressed, it is judged as no DVT, otherwise it is suspected DVT.

[0065] The embodiments of the present invention can automatically identify anatomical locations based on deep learning and squeeze ultrasound imaging, provide guidance for free-hand ultrasound, and help non-experts detect DVT.

[0066] As a preferred embodiment, wherein Figure 3 As shown, the training steps of the deep vein thrombosis prediction model include:

[0067] Step S201, using suspected deep vein thrombosis patients and healthy volunteers as subjects, obtaining first ultrasound image data of a preset examination site of the subjects in an uncompressed state and second ultrasound image data in a compressed state;

[0068] Step S202, preprocessing the first ultrasonic image data and the second ultrasonic image data, and annotating them to obtain a training data set with labels, wherein the labels include a vein position label, an artery position label, and a vein compression state label;

[0069] Step S203, using the ultrasound image data of the training data set as model input and the label as model output, and training to obtain a deep vein thrombosis prediction model.

[0070] Specifically, the model is divided into a model training phase and a model application prediction phase.

[0071] Equipment preparation: Select ultrasound equipment, such as portable ultrasound equipment such as Senz 310 / 330, SonoHealth P5L, Ramzed DP10, etc., with high resolution and real-time image processing capabilities. The ultrasound equipment is equipped with a linear ultrasound probe with a frequency between 5-10MHz to obtain high-quality superficial tissue images.

[0072] The model training phase includes data set preparation, model training, model validation and testing sub-phases.

[0073] In the dataset preparation sub-stage, first, patients suspected of DVT and healthy volunteers are selected. Patients need to have clear DVT symptoms, such as pain and swelling in the lower limbs, etc. Healthy volunteers need to have no DVT symptoms and be in good health as the control group data.

[0074] Then, the images were collected according to the standard two-point or three-point compression ultrasound examination method. The examination sites included the groin (femoral vein area) and behind the knee (popliteal vein area). The ultrasound probe was used for free-hand operation at the target site of the subject, and the video of the entire examination process was recorded. Multiple compressions were required for each examination site to ensure comprehensive data.

[0075] During the groin examination, the subject lies flat with the legs abducted to expose the groin area. The ultrasound probe is placed in the groin area and scanned along the femoral vein. - Perform vein compression and record the state of the vein before and after compression.

[0076] During the behind-the-knee exam, the subject lies flat or sits with the knee bent, exposing the area behind the knee. The ultrasound probe is placed behind the knee and scanned in the direction of the popliteal vein. The vein is compressed, and the state of the vein before and after compression is recorded.

[0077] Next, the collected ultrasound images are preprocessed, including but not limited to cropping, scaling, denoising and other steps. The bilinear interpolation method is used to scale the image to 128×128 pixels to meet the subsequent model input requirements. Image denoising uses a Gaussian filter to reduce noise interference in the image.

[0078] Next, the radiologist manually annotates the image, and the marking results include a first marker, a second marker, and a third marker. The first marker is used to mark the location of the vein, the second marker is used to mark the location of the artery, and the third marker is used to mark the compression state of the vein, such as open or completely compressed.

[0079] Use dedicated annotation tools to precisely depict veins and arteries in each frame to ensure annotation accuracy.

[0080] In the model training sub-stage, first, select an appropriate model architecture, such as using the U-Net model in deep learning as the basic architecture to automatically segment veins and arteries. U-Net is a convolutional neural network commonly used in medical image segmentation, with an encoder and decoder structure.

[0081] The model architecture consists of a backbone network and a branch network. The backbone network is used to extract image features, and the branch networks are used to predict compression states.

[0082] Preferably, a branch network may be added on the basis of U-Net to predict the anatomical position of the image, namely the above-mentioned anatomical position recognition model.

[0083] Then, data augmentation is performed on the training data, including random flipping, translation, rotation, scaling, intensity rescaling and other operations.

[0084] The specific parameters of the data enhancement method are as follows:

[0085] Random flip: flip left to right with a probability of 0.5.

[0086] Translation: Random translation in the range [-15, 15] pixels.

[0087] Rotation: Randomly rotate in the range [-15°, 15°].

[0088] Scale: Randomly scale in the range [0.95, 1.05].

[0089] Intensity Rescaling: Randomly adjust the image intensity in the range [-0.3, 0.3].

[0090] Next, the Adam optimizer was used for model training, and the loss functions included segmentation loss (cross entropy), anatomical location prediction loss (cross entropy), and venous compression state prediction loss (cross entropy).

[0091] The loss function is defined as follows:

[0092] Segmentation loss L segmentation :

[0093]

[0094] Wherein, the subscript i represents, i=1,...,N; N represents; y i means; p(y i )express;

[0095] Anatomical location prediction loss L anatomic_location :

[0096]

[0097] Venous compression state predicts loss L vein_open_or_closed :

[0098] L vein_open_or_closed =-[ylog(p(y))+(1-y)log(1-p(y))]

[0099] The total loss function L total :

[0100] L total =αL segmentation +βL anatomic_location +γL vein_open_or_closed

[0101] Among them, α, β, and γ represent the weights of each loss respectively, where α=100, β=1, and γ=1.

[0102] In the model verification and testing sub-phase, it is divided into internal verification and external verification.

[0103] The model performance is evaluated on the internal validation set, which randomly extracts 10% of the data from the training data and does not participate in model training.

[0104] The evaluation indicators include segmentation accuracy (Dice coefficient), compression state prediction accuracy (F1-score), and anatomical position prediction accuracy (F1-score).

[0105] Segmentation accuracy (Dice coefficient) calculation formula:

[0106]

[0107] Among them, A is the manually annotated segmentation result, and B is the segmentation result predicted by the model.

[0108] The model performance was evaluated on an external validation set with data from different clinical centers to ensure the generalization ability of the model.

[0109] Evaluation indicators included sensitivity, specificity, positive predictive value, and negative predictive value.

[0110] Sensitivity:

[0111]

[0112] Among them, TP is true positive and FN is false negative.

[0113] Specificity:

[0114]

[0115] Among them, TN is true negative and FP is false positive.

[0116] Positive predictive value PPV:

[0117]

[0118] Negative predictive value NPV:

[0119]

[0120] Patients in the external validation set were divided into two groups: one group was examined without software guidance and the other was examined with software guidance.

[0121] By comparing the inspection results of the two groups, the performance of the model in practical application was evaluated.

[0122] As a preferred embodiment, the deep vein thrombosis prediction model includes:

[0123] Backbone network, used to extract image features;

[0124] The branch network, connected to the trunk network, is used to predict the compression state of the vein.

[0125] As a preferred implementation, the deep vein thrombosis prediction model is an improved U-Net model, the trunk network and the branch network constitute the U-Net model, and the convolutional layers of the encoder and the bottleneck layer of the U-Net model are respectively provided with enhanced feature representation modules to construct the improved U-Net model.

[0126] Specifically, adding an enhanced feature representation (Conix) module to the improved U-Net model can enhance the feature extraction capability and robustness of the model.

[0127] In actual clinical prediction tests, it was found that the Conix structure had excellent prediction effect on venous compression status.

[0128] As a preferred embodiment, wherein Figure 4 As shown, the backbone network includes:

[0129] An input layer 100 is used to receive first ultrasonic image data and second ultrasonic image data;

[0130] The encoder 200 includes a plurality of first convolution blocks, each of which includes a first convolution layer, a second convolution layer, an enhanced feature representation module and a maximum pooling layer connected in sequence, for extracting image features layer by layer.

[0131] As a preferred embodiment, the branch network includes:

[0132] The bottleneck layer 300 includes a third convolutional layer, a fourth convolutional layer and an enhanced feature representation module connected in sequence;

[0133] The decoder 400 includes a plurality of second convolution blocks, each of which includes a deconvolution layer, a concatenation layer, a fifth convolution layer, and a sixth convolution layer connected in sequence, and the concatenation layer is jump-connected to the corresponding first convolution block of the encoder, so as to restore the position information of the image;

[0134] The output layer 500 includes a seventh convolutional layer, which is used to output the compression state of the vein.

[0135] Specifically, the following is an example of an improved U-Net network structure including a Conix module.

[0136] The input layer 100 is used to receive a grayscale image with an input size of 256.

[0137] Encoder 200 (downsampling path) progressively extracts image features through multiple convolutional blocks and reduces the spatial resolution.

[0138] Convolutional Block 1

[0139] Convolutional layer: convolution kernel 3x3, 64 filters, ReLU activation;

[0140] Convolutional layer: convolution kernel 3x3, 64 filters, ReLU activation;

[0141] Conix module: enhanced feature representation;

[0142] Max pooling layer: convolution kernel 2x2;

[0143] Convolutional Block 2

[0144] Convolutional layer: convolution kernel 3x3, 128 filters, ReLU activation;

[0145] Convolutional layer: convolution kernel 3x3, 128 filters, ReLU activation;

[0146] Conix module: enhanced feature representation;

[0147] Max pooling layer: convolution kernel 2x2;

[0148] Convolutional Block 3

[0149] Convolutional layer: convolution kernel 3x3, 256 filters, ReLU activation

[0150] Convolutional layer: convolution kernel 3x3, 256 filters, ReLU activation;

[0151] Conix module: enhanced feature representation;

[0152] Max pooling layer: convolution kernel 2x2;

[0153] Convolutional Block 4

[0154] Convolutional layer: convolution kernel 3x3, 512 filters, ReLU activation;

[0155] Convolutional layer: convolution kernel 3x3, 512 filters, ReLU activation;

[0156] Conix module: enhanced feature representation;

[0157] Max pooling layer: convolution kernel 2x2.

[0158] Bottleneck layer 300

[0159] Convolutional layer: convolution kernel 3x3, 1024 filters, ReLU activation;

[0160] Convolutional layer: convolution kernel 3x3, 1024 filters, ReLU activation;

[0161] Conix module: Enhanced feature representation.

[0162] Decoder 400 (Upsampling Path)

[0163] Convolutional Block 5

[0164] Deconvolution layer: convolution kernel 2x2, 512 filters;

[0165] Concatenation layer: concatenated with the output of convolution block 4 of the encoder;

[0166] Convolutional layer: convolution kernel 3x3, 512 filters, ReLU activation;

[0167] Convolutional layer: kernel 3x3, 512 filters, ReLU activation; convolutional block 6

[0168] Deconvolution layer: convolution kernel 2x2, 256 filters;

[0169] Concatenation layer: concatenated with the output of convolution block 3 of the encoder;

[0170] Convolutional layer: convolution kernel 3x3, 256 filters, ReLU activation;

[0171] Convolutional layer: kernel 3x3, 256 filters, ReLU activation; convolutional block 7

[0172] Deconvolution layer: convolution kernel 2x2, 128 filters;

[0173] Concatenation layer: concatenated with the output of convolution block 2 of the encoder;

[0174] Convolutional layer: convolution kernel 3x3, 128 filters, ReLU activation;

[0175] Convolutional layer: kernel 3x3, 128 filters, ReLU activation; convolutional block 8

[0176] Deconvolution layer: convolution kernel 2x2, 64 filters;

[0177] Concatenation: concatenate with the output of convolution block 1 of the encoder;

[0178] Convolutional layer: convolution kernel 3x3, 64 filters, ReLU activation;

[0179] Convolutional layer: convolution kernel 3x3, 64 filters, ReLU activation.

[0180] Output layer 500

[0181] Convolution layer: The convolution kernel is 1x1, the number of output channels depends on the specific task (such as 1 or 2), and the appropriate activation function (such as Softmax or Sigmoid) is selected for binary or multi-classification tasks.

[0182] As a preferred implementation, the enhanced feature representation module is a multi-scale convolutional layer or an attention layer.

[0183] Specifically, the Conix module is a module for enhancing feature representation, which can be implemented through multi-scale convolution or attention mechanism.

[0184] The specific implementation can be customized according to requirements, for example:

[0185] Multi-scale convolution: Use convolution kernels of different sizes (such as 1x1, 3x3, 5x5) to process feature maps in parallel and then concatenate them.

[0186] Attention Mechanism: Use channel attention or spatial attention to enhance important features.

[0187] Compared with the traditional U-Net model, the improved U-Net model of the present invention can better capture the details and complex features in the image and improve the segmentation performance by introducing the Conix structure based on multi-scale convolution or attention mechanism, and is particularly suitable for the identification of small or irregular thrombi in blood vessels. This structure can not only combine more contextual information, which helps to distinguish thrombi from surrounding tissues, thereby improving the accuracy of the model in identifying complex backgrounds, but also significantly enhances the robustness of the model by enhancing feature representation, has better adaptability to different patient variability and image noise, and reduces false detection and missed detection.

[0188] In addition, the improved model showed higher precision and recall in segmentation tasks, ensuring accurate diagnosis in clinical applications. Its multi-scale feature extraction capability enables the model to better handle complex or irregularly shaped thrombi, improving detection sensitivity. At the same time, if the Conix module is designed properly, it can reduce the number of parameters without significantly increasing the amount of calculation, improve model performance, and make it more suitable for resource-constrained environments.

[0189] Table 1 below shows the evaluation results using different model architectures.

[0190] Table 1 Evaluation results of different model architectures

[0191] Model Architecture Sensitivity Specificity AUC Embodiments of the present invention U-NET(conix) 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

[0192] As a preferred embodiment, step S3 includes:

[0193] If the compression state is complete compression, the diagnosis result of deep vein thrombosis is confirmed as no deep vein thrombosis;

[0194] If the compression state is incomplete compression, the diagnosis result of deep vein thrombosis is determined to be suspected deep vein thrombosis.

[0195] The present invention also provides a non-invasive diagnosis system for deep vein thrombosis based on machine learning, which is used to implement the non-invasive diagnosis method for deep vein thrombosis based on machine learning as described above, and the system comprises:

[0196] A data acquisition module 1 is used to acquire first ultrasonic image data of a preset examination part of a subject in an uncompressed state and second ultrasonic image data in a compressed state;

[0197] The model prediction module 2 is connected to the data acquisition module 1 and is used to input the first ultrasound image data and the second ultrasound image data into a pre-established deep vein thrombosis prediction model to obtain the compression state of the vein;

[0198] The diagnosis module 3 is connected to the model prediction module 2 and is used to determine the diagnosis result of deep vein thrombosis according to the compression state.

[0199] Application Example 1: Application in a general medical environment

[0200] Patient A, 65 years old, female, had lower extremity pain and swelling, suspected of DVT. The examination was performed using a portable ultrasound device, and a two-point compression ultrasound examination was guided by the machine learning model of the present invention. The examination results showed complete compression of the vein and no DVT. The patient did not need to be referred and received the diagnosis directly in the general medical setting. The total examination time was about 10 minutes.

[0201] Application Example 2: Application in Remote Care Points

[0202] Patient B, 75 years old, male, with a history of DVT, currently living in a nursing home, had symptoms of lower limb swelling. The nursing staff used a portable ultrasound device to perform the examination, and the machine learning model of the present invention guided the three-point compression ultrasound examination. The examination results showed incomplete compression at the femoral vein, and DVT was suspected. The patient was referred to a specialist hospital for further diagnosis. The total examination time was about 15 minutes.

[0203] Application Example 3: Multicenter Clinical Trial

[0204] A multicenter clinical trial was conducted in two domestic DVT diagnosis centers, recruiting a total of 83 patients suspected of DVT. The machine learning model of the present invention was used for examination, and the results showed that the sensitivity was 0.82-0.96, the specificity was 0.70-0.82, the positive predictive value was 0.65-0.89, and the negative predictive value was 0.99-1.00.

[0205] The test results show that the present invention has good application effects in different clinical environments.

[0206] The advantages or beneficial effects of adopting the above technical solution are: the non-invasive diagnosis method and system of deep vein thrombosis based on machine learning of the present invention are efficient, convenient, accurate and economical, and can significantly reduce the time from initial suspicion to final diagnosis of patients, thereby improving diagnostic efficiency; non-experts can perform DVT screening in ordinary medical environments and remote care points, thereby increasing the convenience of diagnosis; a machine learning model is used to provide highly accurate diagnostic results, with both sensitivity and specificity meeting clinical practical standards; rapid diagnosis and reduction of unnecessary referrals can reduce medical costs and increase net monetary benefits.

[0207] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the contents of this specification and illustrations should be included in the protection scope of the present invention.

Claims

1. A non-invasive diagnosis method for deep vein thrombosis based on machine learning, characterized in that: include: Step S1, collecting first ultrasonic image data of a preset examination part of a subject in an uncompressed state and second ultrasonic image data in a compressed state; Step S2, inputting the first ultrasound image data and the second ultrasound image data into a pre-established deep vein thrombosis prediction model to obtain a compression state of the vein; Step S3, determining a deep vein thrombosis diagnosis result according to the compression state.

2. The non-invasive diagnosis method for deep vein thrombosis based on machine learning according to claim 1, characterized in that: The step S1 comprises: Step S11, placing an ultrasonic probe at the preset inspection site and scanning along the vein direction to obtain a real-time image; Step S12, inputting the real-time image into an anatomical position recognition model to recognize the real-time image and obtain an anatomical position recognition result; Step S13, performing a preliminary scan according to the anatomical position recognition result, and collecting first ultrasonic image data of veins and arteries in an uncompressed state; Step S14, gradually increasing the pressure of the ultrasound probe, and performing scanning again to collect second ultrasound image data of veins and arteries in a compressed state.

3. The non-invasive diagnosis method for deep vein thrombosis based on machine learning according to claim 1, characterized in that: The training steps of the deep vein thrombosis prediction model include: Step S201, using suspected deep vein thrombosis patients and healthy volunteers as subjects, obtaining first ultrasound image data of a preset examination site of the subjects in an uncompressed state and second ultrasound image data in a compressed state; Step S202, preprocessing the first ultrasound image data and the second ultrasound image data, and labeling them to obtain a training data set with labels, wherein the labels include a vein position label, an artery position label, and a vein compression state label; Step S203, using the ultrasound image data of the training data set as model input, and the label as model output, to train and obtain the deep vein thrombosis prediction model.

4. The non-invasive diagnosis method for deep vein thrombosis based on machine learning according to claim 1, characterized in that: The deep vein thrombosis prediction model includes: Backbone network, used to extract image features; The branch network is connected to the trunk network and is used to predict the compression state of the vein.

5. The non-invasive diagnosis method for deep vein thrombosis based on machine learning according to claim 4, characterized in that: The deep vein thrombosis prediction model is an improved U-Net model. The trunk network and the branch network constitute a U-Net model. The encoder and the convolution layer of the bottleneck layer of the U-Net model are respectively provided with enhanced feature representation modules.

6. The non-invasive diagnosis method for deep vein thrombosis based on machine learning according to claim 4, characterized in that: The backbone network includes: An input layer, configured to receive the first ultrasonic image data and the second ultrasonic image data; The encoder includes a plurality of first convolution blocks, each of which includes a first convolution layer, a second convolution layer, an enhanced feature representation module and a maximum pooling layer connected in sequence, and is used to extract image features layer by layer.

7. The non-invasive diagnosis method for deep vein thrombosis based on machine learning according to claim 6, characterized in that: The branch network includes: The bottleneck layer includes the third convolutional layer, the fourth convolutional layer and the enhanced feature representation module connected in sequence; A decoder, comprising a plurality of second convolution blocks, each of which comprises a deconvolution layer, a splicing layer, a fifth convolution layer, and a sixth convolution layer connected in sequence, wherein the splicing layer is jump-connected to the corresponding first convolution block of the encoder, for restoring the position information of the image; The output layer, including the seventh convolutional layer, is used to output the compression state of the vein.

8. The non-invasive diagnosis method for deep vein thrombosis based on machine learning according to claim 5, characterized in that: The enhanced feature representation module is a multi-scale convolutional layer or an attention layer.

9. The non-invasive diagnosis method for deep vein thrombosis based on machine learning according to claim 1, characterized in that: The step S3 comprises: If the compression state is completely compressed, determining that the deep vein thrombosis diagnosis result is no deep vein thrombosis; If the compression state is incomplete compression, the deep vein thrombosis diagnosis result is determined to be suspected deep vein thrombosis.

10. A non-invasive diagnosis system for deep vein thrombosis based on machine learning, characterized in that: A system for implementing the non-invasive diagnosis method for deep vein thrombosis based on machine learning as described in any one of claims 1 to 9, comprising: A data acquisition module, used for acquiring first ultrasonic image data of a preset examination part of a subject in an uncompressed state and second ultrasonic image data in a compressed state; a model prediction module, connected to the data acquisition module, and used for inputting the first ultrasound image data and the second ultrasound image data into a pre-established deep vein thrombosis prediction model to obtain a compression state of the vein; A diagnosis module is connected to the model prediction module and is used to determine a diagnosis result of deep vein thrombosis according to the compression state.

Images