Apparatus for prostate cancer diagnosis model based on multi-parameter ultrasound images and training method thereof

Abstract

The application discloses a prostate cancer diagnosis model based on multi-parameter ultrasonic images and a training method thereof. The prostate cancer diagnosis model comprises a convolutional neural network for extracting image features, a feature reconstruction algorithm module, a long short-term memory neural network for extracting ultrasonic contrast semantic features, and a decision fusion classification network. The training method comprises establishing a training sample set, then inputting the training sample into the prostate cancer lesion diagnosis model for training until the error between the predicted value output by the prostate cancer lesion diagnosis model and the real value of the lesion target classification converges, and a qualified prostate cancer lesion diagnosis model is obtained. The application can extract the optimal features in different parameter ultrasonic images by virtue of the advantages of multiple feature extraction networks, and can obtain a prostate cancer diagnosis model capable of extracting rich features, recognizing accurate results and being robust by processing the time sequence features of ultrasonic contrast through the long short-term memory network.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for establishing a multi-parameter ultrasound intelligent diagnostic model for prostate cancer. Background Technology

[0002] Prostate cancer is one of the most common cancers in men, and it is currently the second most common and fifth leading cause of cancer death in men worldwide. In China, its incidence rate is also on the rise, now ranking sixth and its mortality rate seventh. Traditional methods for diagnosing prostate cancer mainly rely on biopsy, but this method has drawbacks such as requiring numerous needles and low accuracy. Furthermore, unlike other parts of the body, the prostate cannot be directly classified using ultrasound images to determine whether it is cancerous; the lesion area needs to be delineated. This places higher demands on the accuracy of imaging in identifying prostate cancer and precise targeted biopsy localization.

[0003] With advancements in imaging techniques, multiparameter ultrasound imaging has become a promising diagnostic tool for prostate cancer. However, traditional prostate cancer diagnostic methods struggle to extract effective features, exhibiting poor robustness and low stability, resulting in low accuracy and failing to meet practical medical needs. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a device and training method for a prostate cancer diagnostic model based on multi-parameter ultrasound imaging, so as to solve the problem of obtaining a prostate cancer diagnostic model that can extract rich features, has good robustness and accurate identification results.

[0005] The device for the prostate cancer diagnostic model based on multi-parameter ultrasound imaging of the present invention includes a convolutional neural network for extracting image features, a feature reconstruction algorithm module, a long short-term memory neural network for extracting semantic features of ultrasound contrast imaging, and a decision fusion classification network.

[0006] The convolutional neural network includes a first convolutional neural network that takes the dynamic ultrasound contrast image of the prostate puncture target area as input, a second convolutional neural network that takes the grayscale ultrasound image of the prostate puncture target area as input, a third convolutional neural network that takes the color Doppler blood flow image of the prostate puncture target area as input, and a fourth convolutional neural network that takes the SMI image of the prostate puncture target area as input. The outputs of the first convolutional neural network, the second convolutional neural network, the third convolutional neural network, and the fourth convolutional neural network are used as inputs to the feature reconstruction algorithm module.

[0007] The long short-term memory neural network uses ultrasound contrast dynamic images as input;

[0008] The decision fusion classification network includes a softmax activation function layer and a decision fusion network. The output of the feature reconstruction algorithm module and the output of the long short-term memory neural network are used as inputs to the softmax activation function layer. The decision fusion network fuses the classification results of the softmax activation function layer and outputs them.

[0009] Furthermore, the feature reconstruction algorithm module is defined as follows: the outputs of the first, second, third, and fourth convolutional neural networks are sequentially used as a feature input matrix. Non-ideal feature values ​​in each row of the feature matrix are removed. The non-ideal feature value in each row of the feature matrix has the largest difference with the other three feature values ​​in that row. Let the four feature values ​​in the nth row of the feature matrix be A... n1 A n2 A n3 A n4 ,

[0010] A n1 The difference between the value of the feature and the other three feature values ​​in the same row is ,

[0011] A n2 The difference between the value of the feature and the other three feature values ​​in the same row is ,

[0012] A n3 The difference between the value of the feature and the other three feature values ​​in the same row is ,

[0013] A n4 The difference between the value of the feature and the other three feature values ​​in the same row is ,

[0014] Compare the four differences R n1 R n2 R n3 R n4 The size of the feature matrix is ​​determined by the maximum value of the feature matrix, which is the non-ideal feature value. Feature extraction is performed on the new feature matrix obtained after removing the non-ideal feature values ​​from each row, and the extracted features are used as the output of the feature reconstruction algorithm module.

[0015] Furthermore, the first and second convolutional neural networks are shallower than the third and fourth convolutional neural networks.

[0016] The training method for the prostate cancer diagnostic model based on multi-parameter ultrasound imaging described in this invention includes the following steps:

[0017] S1: Establish the training sample set:

[0018] S11: Acquire multi-parameter ultrasound imaging data of the patient, including contrast-enhanced ultrasound images, grayscale ultrasound images, color Doppler blood flow images, and SMI images;

[0019] S12: The acquired grayscale ultrasound images, color Doppler blood flow images and SMI images are preprocessed to increase the image quality and resolvability. The preprocessing includes extraction of the prostate boundary region, cropping of the puncture target area image, and denoising, enhancement and contrast adjustment of the target area image to obtain grayscale ultrasound image training samples, color Doppler blood flow image training samples and SMI image training samples.

[0020] S13: The time-intensity curve algorithm based on correction is used to extract the time features of the preprocessed ultrasound contrast imaging dynamic image. Then, the target area features are enhanced. The prostate area is cropped using the manual boundary localization method. The cropped prostate area is scaled to a fixed size. The semantic features of the puncture target area are extracted from the scaled image. Each semantic feature is normalized and the semantic features are combined into corresponding feature vectors according to the time series. At the same time, the ultrasound contrast imaging dynamic image training samples are obtained.

[0021] S14: Perform mirror flipping, random cropping and rotation on the training samples obtained after steps S12 and S13 to augment the training data.

[0022] S2: Training a prostate cancer lesion diagnostic model:

[0023] S21: Ultrasound contrast-enhanced dynamic image training samples are extracted from the training sample set and input into the first convolutional neural network. The first convolutional neural network performs feature extraction and outputs a first feature vector. Grayscale ultrasound image training samples are extracted from the training sample set and input into the second convolutional neural network. The second convolutional neural network performs feature extraction and outputs a second feature vector. Color Doppler blood flow image training samples are extracted from the training sample set and input into the third convolutional neural network. The third convolutional neural network performs feature extraction and outputs a third feature vector. SMI image training samples are extracted from the training sample set and input into the fourth convolutional neural network. The fourth convolutional neural network performs feature extraction and outputs a fourth feature vector. The first, second, third, and fourth feature vectors are each used as a column of features and input into the feature reconstruction algorithm module. The feature reconstruction algorithm module performs irrational feature removal and feature extraction and outputs a fifth feature vector.

[0024] The training samples of ultrasound contrast dynamic images are extracted from the training sample set and input into the long short-term memory neural network. After the long short-term memory neural network extracts features, it outputs the sixth feature vector.

[0025] The training samples for the first convolutional neural network, the second convolutional neural network, the third convolutional neural network, the fourth convolutional neural network, and the long short-term memory neural network were all from the same target area of ​​the patient at the same time.

[0026] S22: Input the fifth and sixth feature vectors originating from the same target area of ​​the patient into the decision fusion classification network, and the decision fusion classification network outputs the predicted value of the target area classification of prostate cancer lesions;

[0027] S23: Repeat steps S21 and S22 until the error between the predicted value output by the prostate cancer lesion diagnosis model and the true value of the lesion target area classification converges, thus obtaining a qualified prostate cancer lesion diagnosis model.

[0028] Furthermore, the time intensity curve algorithm based on correction described in step S13 is as follows:

[0029]

[0030] In the formula, [ ] is the floor function, N Cor N represents the corrected sequence value of peak frame count for contrast microbubbles. Cor It also represents the end frame sequence value extracted from ultrasound contrast imaging dynamic image segments, N. Max C represents the frame sequence value corresponding to the maximum peak value. Max The grayscale value corresponding to the maximum peak value; j is the number of peaks before the maximum peak value, C i The grayscale value corresponding to each peak, N i fps represents the frame rate sequence value, where fps is the frame rate of the ultrasound contrast video.

[0031] Furthermore, the enhancement processing of the ultrasound contrast imaging dynamic image in S13 includes:

[0032] (1) The dynamic images of ultrasound contrast imaging are processed by the inter-frame difference method to obtain the frames in which the initial microbubbles reach the target area and the frames in which the microbubbles appear in both the internal and external glands. The average gray value of the internal gland target area and the average gray value of the external gland target area in the frame are calculated. When the average gray value of the external gland target area is greater than that of the internal gland target area, the external gland target area feature enhancement is performed.

[0033] (2) For each frame image processed by the inter-frame difference method, compare the average gray value of the left leaf target area with the average gray value of the right leaf target area. If the difference in average gray value exceeds 5, then perform target area feature enhancement on the target area with the larger average gray value.

[0034] (3) Calculate the average gray value of all target areas in each frame image, and perform target area feature enhancement on target areas whose gray values ​​are less than the average gray value of all target areas.

[0035] Furthermore, S13 also includes using a domain knowledge-guided time attention mechanism to guide the enhancement processing of ultrasound contrast imaging dynamic images. The domain knowledge-guided time attention mechanism includes: collecting the experience of imaging experts in observing ultrasound contrast imaging sequence data, and performing different weighting processing on the sequence data to focus more attention on the time segments that the imaging experts consider important. The time segments that the imaging experts consider important include the time segment from the beginning of the contrast microbubbles to the appearance of both the internal and external glands, and the time segment from the time segment when the microbubbles reach their peak and gradually weaken.

[0036] The beneficial effects of this invention are:

[0037] This invention relates to a device and training method for a prostate cancer diagnostic model based on multi-parameter ultrasound imaging. It employs ultrasound images of multiple modalities and utilizes neural networks of varying depths for feature extraction from different images, resulting in better image features. The fused and reconstructed features from these multiple image types compensate for the feature deficiencies of each individual modality. Furthermore, when processing dynamic ultrasound contrast imaging, a time-intensity curve correction algorithm is employed. This algorithm is adaptive and can automatically extract the time segment from the appearance to the peak of arterial microbubbles in the ultrasound contrast imaging video, making it more sensitive to semantic feature extraction. The established prostate cancer diagnostic model, leveraging the advantages of multiple different feature extraction networks, can richly extract image and semantic features from the training samples. It is adaptable to the diagnosis of prostate cancer images from different doctors and different machines, exhibiting good robustness. Simultaneously, it utilizes the dynamic characteristics of ultrasound contrast imaging, thereby improving the accuracy of prostate cancer prediction. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the technical route for constructing a model for diagnosing prostate cancer;

[0039] Figure 2 This is a schematic diagram of the network structure of a prostate cancer lesion diagnostic model;

[0040] Figure 3 This is a schematic diagram of a relatively shallow convolutional neural network structure;

[0041] Figure 4 This is a schematic diagram of a relatively deep convolutional neural network structure;

[0042] Figure 5 This is a schematic diagram of the feature reconstruction algorithm structure of a convolutional neural network;

[0043] Figure 6 This is a schematic diagram of the time attention mechanism guided by domain knowledge. Detailed Implementation

[0044] The present invention will be further described below with reference to embodiments.

[0045] The device for the prostate cancer diagnostic model based on multi-parameter ultrasound images in this embodiment includes a convolutional neural network for extracting image features, a feature reconstruction algorithm module, a long short-term memory neural network for extracting semantic features of ultrasound contrast imaging, and a decision fusion classification network.

[0046] The convolutional neural network includes a first convolutional neural network that takes the dynamic ultrasound contrast image of the prostate puncture target area as input, a second convolutional neural network that takes the grayscale ultrasound image of the prostate puncture target area as input, a third convolutional neural network that takes the color Doppler blood flow image of the prostate puncture target area as input, and a fourth convolutional neural network that takes the SMI image of the prostate puncture target area as input. The outputs of the first, second, third, and fourth convolutional neural networks serve as the inputs to the feature reconstruction algorithm module.

[0047] The long short-term memory neural network uses ultrasound contrast images as input.

[0048] The decision fusion classification network includes a softmax activation function layer and a decision fusion network. The output of the feature reconstruction algorithm module and the output of the long short-term memory neural network are used as inputs to the softmax activation function layer. The decision fusion network fuses the classification results of the softmax activation function layer and outputs them.

[0049] In this embodiment, the feature reconstruction algorithm module is defined as follows: the outputs of the first convolutional neural network, the second convolutional neural network, the third convolutional neural network, and the fourth convolutional neural network are sequentially used as a feature input feature matrix. Non-ideal feature values ​​in each row of the feature matrix are removed. The non-ideal feature value in each row of the feature matrix has the largest difference with the other three feature values ​​in that row. Let the four feature values ​​in the nth row of the feature matrix be A... n1 A n2 A n3 A n4 ,

[0050] A n1 The difference between the value of the feature and the other three feature values ​​in the same row is ,

[0051] A n2 The difference between the value of the feature and the other three feature values ​​in the same row is ,

[0052] A n3 The difference between the value of the feature and the other three feature values ​​in the same row is ,

[0053] A n4 The difference between the value of the feature and the other three feature values ​​in the same row is ,

[0054] Compare the four differences R n1 R n2 R n3 R n4 The size of the feature matrix is ​​determined by the maximum value of the feature matrix, which is the non-ideal feature value. Feature extraction is performed on the new feature matrix obtained after removing the non-ideal feature values ​​from each row, and the extracted features are used as the output of the feature reconstruction algorithm module.

[0055] As an improvement to the above embodiments, the first and second convolutional neural networks are shallower than the third and fourth convolutional neural networks. This improvement makes it easier for each convolutional neural network to extract different features.

[0056] The training method for the prostate cancer diagnostic model based on multi-parameter ultrasound imaging described in the above embodiments includes the following steps:

[0057] S1: Establish the training sample set:

[0058] S11: Acquire multi-parameter ultrasound imaging data of the patient, including contrast-enhanced ultrasound images, grayscale ultrasound images, color Doppler blood flow images, and SMI images.

[0059] S12: Preprocess the acquired grayscale ultrasound images, color Doppler blood flow images, and SMI images to increase image quality and resolvability. The preprocessing includes extracting the prostate boundary region, scaling the prostate region to a fixed size of 600 (width) * 480 (height), cropping the puncture target area image, with each target area image size being 64 (image width) * 64 (image height) * 3 (image channel number), and denoising, enhancement, and contrast adjustment of the target area images to obtain grayscale ultrasound image training samples, color Doppler blood flow image training samples, and SMI image training samples.

[0060] S13: A correction-based temporal intensity curve algorithm is used to extract temporal features from the preprocessed contrast-enhanced ultrasound images. For the clinically retained training sample videos of prostate contrast-enhanced ultrasound cases, the entire video contains not only the time segment from the observation of arterial microbubbles to the peak arterial flow during diagnosis, but also other irrelevant time segments. The correction-based temporal intensity curve algorithm automatically selects the time nodes in the contrast-enhanced ultrasound video, eliminating these irrelevant time segments.

[0061] The time intensity curve algorithm based on correction described in this step is as follows:

[0062]

[0063] In the formula, [ ] is the floor function, N Cor N represents the corrected sequence value of peak frame count for contrast microbubbles. Cor It also represents the end frame sequence value extracted from ultrasound contrast imaging dynamic image segments, N. Max C represents the frame sequence value corresponding to the maximum peak value. Max The grayscale value corresponding to the maximum peak value; j is the number of peaks before the maximum peak value, C i The grayscale value corresponding to each peak, N i Here, fps represents the frame rate of the ultrasound contrast imaging video. In this embodiment, a maximum value is selected every 5 units (frames) in the curve, i.e., fps=5, because each fluctuation in grayscale value indicates that venous microbubbles have an impact, and it is necessary to weaken the influence of venous blood vessels.

[0064] Then, a domain-knowledge-guided temporal attention mechanism is used to enhance the dynamic ultrasound contrast-enhanced images. This mechanism includes: leveraging the experience of imaging experts in observing ultrasound contrast-enhanced sequence data; applying different weightings to the sequence data to focus more attention on time segments deemed important by the imaging experts. These important time segments include the period from the initial emergence of contrast-enhanced microbubbles to their appearance in both internal and external glands, and the period from the microbubbles reaching their peak and gradually diminishing. Enhancement processing of target area features in the dynamic ultrasound contrast-enhanced images includes:

[0065] (1) The dynamic images of ultrasound contrast imaging are processed by the inter-frame difference method to obtain the frames in which the initial microbubbles reach the target area and the frames in which the microbubbles appear in both the internal and external glands. The average gray value of the internal gland target area and the average gray value of the external gland target area in the frame are calculated. When the average gray value of the external gland target area is greater than that of the internal gland target area, the external gland target area feature enhancement is performed.

[0066] (2) For each frame image processed by the inter-frame difference method, compare the average gray value of the left leaf target area with the average gray value of the right leaf target area. If the difference in average gray value exceeds 5, then perform target area feature enhancement on the target area with the larger average gray value.

[0067] (3) Calculate the average gray value of all target areas in each frame image, and perform target area feature enhancement on target areas whose gray values ​​are less than the average gray value of all target areas.

[0068] Then, the prostate region is cropped using a manual boundary localization method. The cropped prostate region is then scaled to a fixed size. In this embodiment, the size is scaled to a fixed size of 600 (width) * 480 (height). Semantic features of the puncture target area are extracted from the scaled image. The semantic features include the average gray value, contrast, variance, and shape factor of each target area. Each semantic feature is normalized, and the semantic features are combined into a 1*128 feature vector according to the time series to obtain the ultrasound contrast imaging dynamic image training sample.

[0069] S14: Perform mirror flipping, random cropping and rotation processing on the training samples obtained after steps S12 and S13 to achieve training data augmentation.

[0070] S2: Training a prostate cancer lesion diagnostic model:

[0071] S21: Ultrasound contrast-enhanced dynamic image training samples are extracted from the training sample set and input into a first convolutional neural network. The first convolutional neural network performs feature extraction and outputs a first feature vector. Grayscale ultrasound image training samples are extracted from the training sample set and input into a second convolutional neural network. The second convolutional neural network performs feature extraction and outputs a second feature vector. The first and second convolutional neural networks can be implemented as follows: Figure 3 The shallow neural network structure shown takes a 64*64*3 feature map as input and outputs a 1*128 feature vector after passing through the shallow neural network structure.

[0072] Color Doppler blood flow images are extracted from the training sample set and input into a third convolutional neural network (CNN). The CNN performs feature extraction and outputs a third feature vector. SMI images are extracted from the training sample set and input into a fourth convolutional neural network (CNN). The CNN performs feature extraction and outputs a fourth feature vector. The third and fourth CNNs can be implemented using methods such as... Figure 4 The deep neural network structure shown takes a 64*64*3 feature map as input and outputs a 1*128 feature vector after passing through the deep neural network structure.

[0073] The first, second, third, and fourth feature vectors are each used as a column of features input to the feature reconstruction algorithm module. After removing irrational features and extracting other features, the feature reconstruction algorithm module outputs the fifth feature vector. For example... Figure 5 As shown, the four 1*128 feature vectors are used as a column of feature input to the feature matrix to obtain a 4*128 feature matrix. The non-ideal feature value with the largest difference from the other three ideal feature values ​​in each row of the feature matrix is ​​removed to obtain a new 3*128 feature matrix. Then, feature extraction is performed to obtain a 1*32 feature vector, which is denoted as the fifth feature vector.

[0074] The training samples are extracted from the training sample set, and then input into a Long Short-Term Memory (LSTM) neural network. After extracting features, the LSM outputs a sixth feature vector. For example... Figure 6 As shown, the temporal intensity curve is used as a reference for the weight allocation of the attention mechanism. The weight parameters are increased for the periods of initial microbubble emergence and the slowdown of microbubble emergence, thus paying more attention to the semantic features in these segments. The 1*128 semantic features are used as input to the Long Short-Term Memory (LSTM) neural network, outputting a 1*32 feature vector, denoted as the sixth feature vector. Introducing a domain-knowledge-guided temporal attention mechanism into the LSTM neural network can also avoid the problems of wasted training resources and training difficulties.

[0075] The training samples for the first, second, third, fourth convolutional neural networks and the long short-term memory neural network were all from the same target area of ​​the patient.

[0076] S22: Input the fifth and sixth feature vectors originating from the same target area of ​​the patient into the decision fusion classification network. The decision fusion classification network outputs the predicted value for the classification of the prostate cancer lesion target area.

[0077] S23: Repeat steps S21 and S22 until the error between the predicted value output by the prostate cancer lesion diagnosis model and the true value of the lesion target area classification converges, thus obtaining a qualified prostate cancer lesion diagnosis model.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A device for a prostate cancer diagnostic model based on multi-parameter ultrasound imaging, characterized in that: It includes a convolutional neural network for extracting image features, a feature reconstruction algorithm module, a long short-term memory neural network for extracting semantic features of ultrasound contrast imaging, and a decision fusion classification network; The convolutional neural network includes a first convolutional neural network that takes the dynamic ultrasound contrast image of the prostate puncture target area as input, a second convolutional neural network that takes the grayscale ultrasound image of the prostate puncture target area as input, a third convolutional neural network that takes the color Doppler blood flow image of the prostate puncture target area as input, and a fourth convolutional neural network that takes the SMI image of the previous prostate puncture target area as input. The outputs of the first convolutional neural network, the second convolutional neural network, the third convolutional neural network, and the fourth convolutional neural network are used as inputs to the feature reconstruction algorithm module. The long short-term memory neural network uses ultrasound contrast dynamic images as input; The decision fusion classification network includes a softmax activation function layer and a decision fusion network. The output of the feature reconstruction algorithm module and the output of the long short-term memory neural network are used as the input of the softmax activation function layer. The decision fusion network fuses the classification results of the softmax activation function layer and outputs them. The feature reconstruction algorithm module is defined as follows: The outputs of the first, second, third, and fourth convolutional neural networks are sequentially used as a feature input matrix. Non-ideal feature values ​​in each row of the feature matrix are removed. The non-ideal feature value in each row is maximized by the difference between it and the other three feature values ​​in that row. Let A be the four feature values ​​in the nth row of the feature matrix. n1 A n2 A n3 A n4 , A n1 The difference between the value of the feature and the other three feature values ​​in the same row is , A n2 The difference between the value of the feature and the other three feature values ​​in the same row is , A n3 The difference between the value of the feature and the other three feature values ​​in the same row is , A n4 The difference between the value of the feature and the other three feature values ​​in the same row is , Compare the four differences R n1 R n2 R n3 R n4 The size of the feature matrix is ​​determined, where the feature value corresponding to the maximum value is the non-ideal feature value. Feature extraction is performed on the new feature matrix obtained after removing the non-ideal feature values ​​from each row, and the extracted features are used as the output of the feature reconstruction algorithm module.

2. The device for prostate cancer diagnosis model based on multi-parameter ultrasound imaging according to claim 1, characterized in that: The first and second convolutional neural networks are shallower than the third and fourth convolutional neural networks.

3. A training method for a prostate cancer diagnostic model based on multi-parameter ultrasound imaging as described in claim 1 or 2, characterized in that: Includes the following steps: S1: Establish the training sample set: S11: Acquire multi-parameter ultrasound imaging data of the patient, including contrast-enhanced ultrasound images, grayscale ultrasound images, color Doppler blood flow images, and SMI images; S12: The acquired grayscale ultrasound images, color Doppler blood flow images and SMI images are preprocessed to increase the image quality and resolvability. The preprocessing includes extraction of the prostate boundary region, cropping of the puncture target area image, and denoising, enhancement and contrast adjustment of the target area image to obtain grayscale ultrasound image training samples, color Doppler blood flow image training samples and SMI image training samples. S13: The time-intensity curve algorithm based on correction is used to extract the time features of the preprocessed ultrasound contrast imaging dynamic image. Then, the target area features are enhanced. The prostate area is cropped using the manual boundary localization method. The cropped prostate area is scaled to a fixed size. The semantic features of the puncture target area are extracted from the scaled image. Each semantic feature is normalized and the semantic features are combined into corresponding feature vectors according to the time series. At the same time, the ultrasound contrast imaging dynamic image training samples are obtained. S14: Perform mirror flipping, random cropping and rotation on the training samples obtained after steps S12 and S13 to augment the training data. S2: Training a prostate cancer lesion diagnostic model: S21: Ultrasound contrast-enhanced dynamic image training samples are extracted from the training sample set and input into the first convolutional neural network. The first convolutional neural network performs feature extraction and outputs a first feature vector. Grayscale ultrasound image training samples are extracted from the training sample set and input into the second convolutional neural network. The second convolutional neural network performs feature extraction and outputs a second feature vector. Color Doppler blood flow image training samples are extracted from the training sample set and input into the third convolutional neural network. The third convolutional neural network performs feature extraction and outputs a third feature vector. SMI image training samples are extracted from the training sample set and input into the fourth convolutional neural network. The fourth convolutional neural network performs feature extraction and outputs a fourth feature vector. The first, second, third, and fourth feature vectors are each used as a column of features and input into the feature reconstruction algorithm module. The feature reconstruction algorithm module performs irrational feature removal and feature extraction and outputs a fifth feature vector. The training samples of ultrasound contrast dynamic images are extracted from the training sample set and input into the long short-term memory neural network. After the long short-term memory neural network extracts features, it outputs the sixth feature vector. The training samples for the first convolutional neural network, the second convolutional neural network, the third convolutional neural network, the fourth convolutional neural network, and the long short-term memory neural network were all from the same target area of ​​the patient at the same time. S22: Input the fifth and sixth feature vectors originating from the same target area of ​​the patient into the decision fusion classification network, and the decision fusion classification network outputs the predicted value of the target area classification of prostate cancer lesions; S23: Repeat steps S21 and S22 until the error between the predicted value output by the prostate cancer lesion diagnosis model and the true value of the lesion target area classification converges, thus obtaining a qualified prostate cancer lesion diagnosis model.

4. The training method for the prostate cancer diagnostic model based on multi-parameter ultrasound imaging according to claim 3, characterized in that: The time intensity curve algorithm based on correction mentioned in step S12 is as follows: In the formula, [ ] is the floor function, N Cor N represents the corrected sequence value of peak frame count for contrast microbubbles. Cor It also represents the end frame sequence value extracted from ultrasound contrast imaging dynamic image segments, N. Max C represents the frame sequence value corresponding to the maximum peak value. Max The grayscale value corresponding to the maximum peak value; j is the number of peaks before the maximum peak value, C i The grayscale value corresponding to each peak, N i fps represents the frame rate sequence value, where fps is the frame rate of the ultrasound contrast video.

5. The training method for the prostate cancer diagnostic model based on multi-parameter ultrasound imaging according to claim 3, characterized in that: The enhancement process in S13 includes: (1) The dynamic images of ultrasound contrast imaging are processed by the inter-frame difference method to obtain the frames in which the initial microbubbles reach the target area and the frames in which the microbubbles appear in both the internal and external glands. The average gray value of the internal gland target area and the average gray value of the external gland target area in the frame are calculated. When the average gray value of the external gland target area is greater than that of the internal gland target area, the external gland target area feature enhancement is performed. (2) For each frame image processed by the inter-frame difference method, compare the average gray value of the left leaf target area with the average gray value of the right leaf target area. If the difference in average gray value exceeds 5, then perform target area feature enhancement on the target area with the larger average gray value. (3) Calculate the average gray value of all target areas in each frame image, and perform target area feature enhancement on target areas whose gray values ​​are less than the average gray value of all target areas.

6. The training method for the prostate cancer diagnostic model based on multi-parameter ultrasound imaging according to claim 3, characterized in that: S13 further includes using a domain knowledge-guided time attention mechanism to guide the enhancement processing of ultrasound contrast imaging dynamic images. The domain knowledge-guided time attention mechanism includes: collecting the experience of imaging experts in observing ultrasound contrast imaging sequence data, and performing different weighting processing on the sequence data to focus more attention on the time segments that the imaging experts consider important. The time segments that the imaging experts consider important include the time segment from the beginning of the contrast microbubbles to the appearance of both the internal and external glands, and the time segment from the time segment when the microbubbles reach their peak and gradually weaken.

Images