Generalized localization segmentation method and system for grayscale ultrasound image lesion sites

Abstract

The application provides a general positioning segmentation method and system for grayscale ultrasonic image lesion sites, and the segmentation task is divided into two stages according to similar visual features of the lesion sites in the image: in the first stage, a deep learning model is used to detect the position of the lesion site in the ultrasonic image and preliminarily draw an outline; in the second stage, the original image is cropped based on the position of the detected lesion site in the image to obtain a partial image including only the lesion site, and then a deep learning model is used to extract the lesion site features in the obtained partial image, so that a more fine and accurate outline is obtained. By fusing the two deep learning models used in the above two stages into one model for unified architecture design and learning, the function of end-to-end automatic positioning and segmentation of the lesion site in the ultrasonic image is realized, and the accurate segmentation capability of the lesion site of different sizes in the ultrasonic image of different tissues and organs is possessed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical image processing, convolutional neural network, deep learning and artificial intelligence, and specifically relates to a general positioning and segmentation method and system for lesion sites in grayscale ultrasound images. BACKGROUND

[0002] Artificial intelligence methods based on deep learning have achieved great success in applications such as machine vision and natural language processing in recent years; for example, in the process of clinical diagnosis based on medical images, artificial intelligence methods can realize automatic detection and segmentation of lesion sites for given medical ultrasound images. However, most artificial intelligence methods are only developed and verified for ultrasound images of a specific organ; so far, there is no universal positioning and segmentation method for lesion sites in medical grayscale ultrasound images of various human tissue organs. Possible reasons include the great differences in size, shape and texture features of lesion sites in grayscale ultrasound images of different tissue organs, and the difficulty in obtaining ultrasound image databases across multiple tissue organs and high-quality artificial annotations, which pose great challenges to the training and performance evaluation of deep learning methods.

[0003] Considering the imaging principle of medical ultrasound images, i.e. the lesion site usually presents visual features of different ultrasound echo intensity or mode from the surrounding normal tissue organs in the image, in theory, a universal lesion site segmentation algorithm (Universal Lesion Segmentation, hereinafter referred to as ULS) can be developed to meet the needs of positioning and segmenting lesion sites from ultrasound images of different tissue organs. In fact, in recent years, various ULS algorithms based on deep learning technology developed for CT images have been successfully applied to clinical auxiliary diagnosis scenarios for multiple diseases. The application of these algorithms shows that ULS algorithms based on multiple tissue organs may have better performance and generalization ability than algorithms related to single organ lesion sites, and combined with transfer learning technology, it is helpful to improve the quantity and quality of artificial annotation data in medical images.

[0004] In addition, the size and shape of the lesion region in medical ultrasound images of different tissues and organs vary greatly. According to statistics, more than 60% of the lesion regions in ultrasound images account for 5% or less of the entire image area, but nearly 5% of the lesion regions account for more than 50% of the entire image area. For deep learning segmentation models, a large number of small targets to be segmented and a small number of large targets exist simultaneously in ultrasound images, which poses great challenges to the architecture and objective function design of deep learning models and training methods. For example, the commonly used Dice-based objective function has large numerical changes and slow convergence speed in the training process of deep learning models, and is not suitable for small target segmentation tasks. The binary cross-entropy objective function with weight designed for small target segmentation tasks may be more stable and easy to converge, but the weight parameter setting is extremely challenging in the case of large changes in lesion area. For example, the commonly used data enhancement technique (copy-and-paste) in traditional small target segmentation problems is difficult to apply to medical ultrasound images because the lesion site usually forms a continuous texture feature with the surrounding tissue.

[0005] Therefore, although theoretically feasible and with the precedent of successful application of ULS algorithm for CT images, the implementation of a general positioning and segmentation method for lesion sites in medical ultrasound images still faces great challenges, and a general automated technique is needed to achieve efficient and accurate segmentation of lesion regions in medical ultrasound images. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a general positioning and segmentation method and system for lesion sites in grayscale ultrasound images, for end-to-end automated positioning and segmentation of lesion sites in ultrasound images.

[0007] The technical solution adopted by the present application to solve the above technical problem is: a general positioning and segmentation method for lesion sites in grayscale ultrasound images, comprising the following steps:

[0008] S0: build a general positioning segmentation system for grayscale ultrasound image lesion site; including a dual-input dual-output deep learning segmentation model, an adaptive target function module focusing on the size of the lesion site, a two-stage lesion site perception algorithm module and a weight addition algorithm module fusing the outputs of the two branches; the dual-input dual-output deep learning segmentation model uses a convolutional neural network as the backbone, including a first input branch, a second input branch, a first encoder, a second encoder, a decoder, a first output branch and a second output branch; the first input branch is connected to the first encoder; the second input branch is connected to the second encoder; the outputs of the first encoder and the second encoder are merged and then connected to the decoder; the first encoder and the second encoder are connected to the decoder through a jump connection respectively; the decoder is connected to the first output branch and the second output branch respectively; the adaptive target function module focusing on the size of the lesion site includes two binary cross entropy loss functions with weights and a standard binary cross entropy loss function;

[0009] S1: train the model in combination with the dual-input dual-output deep learning segmentation model and the adaptive target function module focusing on the size of the lesion site, and extract features related to the lesion site from the original ultrasound image;

[0010] S2: the two-stage lesion site perception algorithm module performs progressive detection and lesion site positioning through multiple iterations, and updates the part of the ultrasound image including the lesion site input to the second input branch of the dual-input dual-output deep learning segmentation model after being intercepted;

[0011] S3: the weight addition algorithm module fusing the outputs of the two branches obtains more ideal final prediction results by comprehensively considering the learning effects of the dual-input dual-output deep learning segmentation model in different branches.

[0012] According to the above scheme, in step S1, the specific steps are as follows:

[0013] S11: let y be the artificial lesion annotation, is the model prediction result, N is the number of images input to the model, and λ is the weight coefficient; then the definition of the binary cross entropy loss function with weight is as follows:

[0014]

[0015] For the first input branch, let λ>0.5, which is used to regard the lesion site segmentation problem as a typical small target segmentation problem, and pay attention to suppressing false negatives in the prediction result in the training process of the model;

[0016] For the second input branch, let λ<0.5, which is used to regard the lesion site segmentation problem as a general target segmentation problem, and pay attention to suppressing false positives in the prediction result in the training process of the model;

[0017] For the first output branch and the second output branch, λ = 0.5 is set for measuring the similarity between the prediction results of the lesion site;

[0018] S12: Set y is the prediction result of the first output branch, y is the prediction result of the second output branch, OB1 y is the artificial annotation corresponding to the prediction result of the first output branch, OB2 y is the artificial annotation corresponding to the prediction result of the second output branch, y is the intercepted part image containing the lesion part in the prediction result of the first output branch; then the expression formula of the adaptive target function focusing on the size of the lesion site is:

[0019]

[0020] Further, in step S2, the specific steps are:

[0021] S21: Set the lateral resolution and longitudinal resolution of the original ultrasound image US1 as α and β respectively, and the artificial annotation image corresponding to the original ultrasound image US1 as A1; set the magnification parameter as EL, EL ≥ 1 and EL ∈ R+; set the gradual cycle parameter as NL, NL ≥ 1 and NL ∈ Z + ; randomly intercept part images US2 with sizes of α / 2 and β / 2 from the original ultrasound image US1 respectively, and obtain the annotation images A2 corresponding to the part images US2 from the artificial annotation image A1 according to the coordinates of the interception;

[0022] S22: Take US1, US2, A1 and A2 as training data to train the double-input double-output deep learning segmentation model and update the model weight;

[0023] S23: Use the trained model to infer US1 and obtain the prediction output FM1;

[0024] S24: Perform a binary operation on FM1 and find the largest contour for calculating the minimum bounding rectangle BB corresponding to the contour, and enlarge BB by a specified multiple EL; if the largest contour cannot be obtained, skip step S25;

[0025] S25: Based on the coordinates of the enlarged BB, intercept US1 / A1, and rescale the intercepted image to the size of α / 2 and β / 2, thereby obtaining new US2 and A2;

[0026] S26: Repeat steps S22-S25 NL times.

[0027] Further, in step S2, set the magnification parameter EL = 1.1 and the gradual cycle parameter NL = 5.

[0028] Further, in step S3, the specific steps are:

[0029] S31: Set the true positive TP1, false positive FP1 and false negative FN1 obtained by comparing the prediction output FM1 of the first output branch with the artificially labeled image A1; calculate the Dice index value D1 of the first output branch as:

[0030] D1 = 2 x TP1 / (2 x TP1 + FP1 + FN1);

[0031] Set the prediction output FM2 obtained by using the trained model to infer US2; Set the true positive TP2, false positive FP2 and false negative FN2 obtained by comparing the prediction output FM2 of the second output branch with the artificially labeled image A2; Calculate the Dice index value D2 of the first output branch as:

[0032] D2 = 2 x TP2 / (2 x TP2 + FP2 + FN2);

[0033] S32: Scale FM2 by a ratio of 1 / EL, and replace the corresponding part of FM1 with the scaled FM2 based on the BB coordinates obtained in step S25, to obtain a new FM2;

[0034] S33: Calculate the model prediction result of the fusion of the outputs of the two branches:

[0035] FMF = D1 x FM1 + D2 x FM2.

[0036] A general positioning and segmentation system for gray-scale ultrasound image lesion sites, comprising a double-input double-output deep learning segmentation model, an adaptive target function module focusing on lesion site size, a two-stage lesion site perception algorithm module, and a weight addition algorithm module for fusing the outputs of the two branches; The double-input double-output deep learning segmentation network and the adaptive target function module focusing on the size of the lesion site are used to efficiently extract features related to the lesion site from the ultrasound image; The two-stage lesion site perception algorithm module is used to realize gradual detection and positioning of the lesion site through multiple iterations; The weight addition algorithm module for fusing the outputs of the two branches is used to provide the final prediction result of the network.

[0037] Further, the dual-input dual-output deep learning segmentation model adopts a convolutional neural network as a backbone, and includes a first input branch, a second input branch, a first encoder, a second encoder, a decoder, a first output branch and a second output branch; the first input branch is connected to the first encoder; the second input branch is connected to the second encoder; the outputs of the first encoder and the second encoder are merged and connected to the decoder; the first encoder and the second encoder are respectively connected to the decoder in a manner of jump connection; the decoder is connected to the first output branch and the second output branch; the first input branch is used for receiving an original ultrasound image as input; the second input branch is used for receiving a partial ultrasound image including a lesion site as input; the first encoder and the second encoder are respectively used for layer-by-layer encoding the images output by the first input branch and the second input branch and obtaining two feature maps with the same size and channel number, and then splicing the two feature maps in the channel dimension and inputting them into the decoder; the decoder is used for accepting the feature maps from the first encoder and the second encoder at different convolutional layers through jump connection, and splicing the feature maps with the internally decoded feature maps in the channel dimension, to generate two feature maps with the same size as the input image; the first output branch is used for outputting a probability map of a segmentation prediction result of the original ultrasound image in the feature map output by the decoder; and the second output branch is used for outputting a probability map of a segmentation prediction result of the partial ultrasound image including the lesion site in the feature map output by the decoder.

[0038] Further, the adaptive target function module focusing on the size of the lesion site includes two binary cross entropy loss functions with weights and a standard binary cross entropy loss function; the binary cross entropy loss functions with weights are used for evaluating the errors of the dual-branch outputs of the dual-input dual-output deep learning segmentation network and the corresponding artificial lesion annotations; and the standard binary cross entropy loss function is used for evaluating the similarity between the predictions of the dual-branch outputs for the lesion site.

[0039] A computer storage medium has a computer program stored therein, which can be executed by a computer processor, and the computer program executes a general positioning segmentation method for a lesion site of a grayscale ultrasound image.

[0040] The present application has the following beneficial effects:

[0041] 1.The general positioning and segmentation method and system for the lesion part of the gray-scale ultrasound image of the present application, aiming at the similar visual features presented by the lesion part in the image, divides the segmentation task into two stages: in the first stage, a deep learning model is used to detect the position of the lesion part in the ultrasound image and preliminarily draw the contour, that is, the segmentation of the lesion part is approximated as a typical small target segmentation problem; in the second stage, based on the position of the detected lesion part in the image, the original image is cropped to obtain a partial image including only the lesion part, and then a deep learning model is used to extract the lesion part features in the obtained partial image, that is, the lesion part segmentation is converted into a relatively simple normal or even large target segmentation problem, so as to obtain a more detailed and accurate contour. By fusing the two deep learning models used in the above two stages into one model for unified architecture design and learning, the function of end-to-end automatic positioning and segmentation of the lesion part in the ultrasound image is realized.

[0042] 2.In view of the imaging principle of medical ultrasound images, that is, the lesion part usually presents visual features different from the surrounding normal tissue organs in terms of ultrasonic echo intensity or mode, the general recognition and segmentation method proposed by the present application has the accurate segmentation capability for lesion parts of different sizes in ultrasound images of different tissue organs. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a flowchart of an embodiment of the present application.

[0044] Figure 2 is a deep learning segmentation model diagram based on double-input and double-output of an embodiment of the present application. DETAILED DESCRIPTION

[0045] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0046] Referring to Figure 1 , the embodiment of the present application includes three parts, including a deep learning segmentation network based on double-input and double-output and an adaptive target function focusing on the size of the lesion part, which is used to efficiently extract features related to the lesion part from the ultrasound image; a two-stage lesion part perception algorithm, which realizes progressive detection and positioning of the lesion part through multiple iterations; and a weight addition algorithm fusing the outputs of the two branches of the deep learning segmentation network, which is used to provide the final prediction result of the network.

[0047] The deep learning segmentation model based on dual-input dual-output is composed of two input branches, a deep learning network based on an encoder-decoder structure and a skip-connection, and two output branches. The two input branches respectively accept the original medical ultrasound image and the partial original ultrasound image including the lesion site as input, and the two output branches correspond to the segmentation prediction results of the deep learning network for the original medical ultrasound image and the partial original ultrasound image. The deep learning network adopts a convolutional neural network as the backbone, and designs two encoders to respectively perform layer-by-layer encoding on the images from the two input branches and obtain two feature maps with the same size and channel number. Then the two feature maps are spliced in the channel dimension and input into the decoder. The decoder simultaneously accepts the feature maps from the two encoders at different convolution layers through the skip connection and splices them with the internally decoded feature maps in the channel dimension to generate two feature maps with the same size as the input image. The two feature maps are respectively connected to the dual-output branch and finally obtain the probability map of the segmentation result. The specific architecture of the model is shown in Figure 2

[0048] The adaptive target function focusing on the size of the lesion site is composed of three parts, including two weighted binary cross-entropy loss functions respectively used to evaluate the errors of the dual-branch output of the deep learning segmentation model and the corresponding artificial lesion annotation, and a standard binary cross-entropy loss function used to evaluate the similarity between the predictions of the dual-branch output for the lesion site. The weighted binary cross-entropy loss function is defined as follows,

[0049]

[0050] where y and ​corresponding to the manual annotation and the model prediction result respectively, N is the number of images input to the model, and λ is the weight coefficient. For the first branch of the deep learning segmentation model taking the original ultrasound image as input, we approximate the lesion segmentation as a typical small target segmentation problem, so the model pays more attention to suppressing false negatives in the prediction result during the training process, which is realized by setting the value of λ in formula (1) to be greater than 0.5. For the second branch of the deep learning segmentation model taking the partial original image containing the lesion as input, we approximate the lesion segmentation as a relatively simple and common target segmentation problem and hope to obtain a relatively accurate contour, so the model pays more attention to suppressing false positives in the prediction result during the training process, which is realized by setting the value of λ in formula (1) to be less than 0.5. For the similarity measure between the predictions of the two branches of the model on the lesion, we realize it by setting the value of λ in formula (1) to be equal to 0.5. Therefore, the final expression formula of the adaptive objective function is as follows

[0051]

[0052] wherein and are the prediction results of the two branches of the deep learning segmentation model respectively, y OB1 and y OB2 are the manual annotations corresponding to the prediction results of the two branches of the model respectively. is the partial image containing the lesion part of the prediction result of the branch of the model corresponding to the original ultrasound image. Combined with the above deep learning segmentation model with double input and double output and the adaptive objective function focusing on the size of the lesion, we can efficiently extract the features related to the lesion from the ultrasound image. The combination of the above deep learning segmentation model with double input and double output and the adaptive objective function can complete the process of model training, which is used to efficiently extract the features related to the lesion from the medical ultrasound image.

[0053] The two-stage lesion perception algorithm realizes progressive detection and positioning of the lesion through multiple iterations, which is used to update the input image of the second branch of the aforementioned deep learning segmentation model, i.e. the partial original ultrasound image containing the lesion as input. Assuming that the transverse and longitudinal resolutions of the original medical ultrasound image (US1) are α and β respectively, and the manual annotation image corresponding to the image is A1, the magnification parameter EL (EL≥1 and EL∈R + ) and the progressive cycle parameter NL (NL≥1 and NL∈Z + ) involved in the algorithm, the specific process is as follows:

[0054] (1) randomly crop a partial image US2 with size of a / 2 and b / 2 from the original medical ultrasound image, and obtain the labeled image A2 corresponding to US2 from A1 according to the cropped coordinates;

[0055] (2) learn the aforementioned deep learning segmentation model with US1, US2, A1 and A2 as training data and update the model weight;

[0056] (3) infer US1 based on the trained model and obtain the predicted output FM1 of the model;

[0057] (4) perform a binarization operation on FM1 and find the largest contour for calculating the minimum bounding box (BB) corresponding to the contour, and enlarge the BB by a specified multiple EL; if the largest contour cannot be obtained, skip step (5);

[0058] (5) crop US1 / A1 based on the coordinates of the enlarged BB, and rescale the cropped image to a / 2 and b / 2 to obtain new US2 and A2;

[0059] (6) repeat the above steps (2)-(5) for NL times;

[0060] For the setting of the magnification parameter EL and the gradual loop parameter NL, experiments show that when EL = 1.1 and NL = 5, the aforementioned deep learning segmentation model can better and quickly locate the lesion site in the image and obtain satisfactory segmentation results.

[0061] The weight addition algorithm for fusing the outputs of the double-branch deep learning segmentation model is used to provide the final prediction result of the model. Assuming that the prediction results of the double-branch outputs are FM1 and FM2, and the corresponding artificial labeled images are A1 and A2, the specific calculation formula is as follows:

[0062] (1) calculate the Dice index values of the two output branches, D1 and D2, respectively, as follows:

[0063] Dice = 2 x TP / (2 x TP + FP + FN)

[0064] where TP, FP, and FN are the true positive, false positive, and false negative, respectively, compared with the model prediction result and the artificial labeling result;

[0065] (2) scale FM2 by 1 / EL, and replace the corresponding part of FM1 with the scaled FM2 based on the BB coordinates obtained in step (5) of the aforementioned lesion site perception algorithm to obtain a new FM2;

[0066] (3) calculate the model prediction result of fusing the double-branch outputs:

[0067] FMF = D1 x FM1 + D2 x FM2

[0068] The fusion algorithm can comprehensively consider the learning effects of the aforementioned deep learning segmentation model in different branches, and is expected to obtain a more ideal final prediction result.

[0069] The above examples are only used to illustrate the design idea and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the content of the present application and to implement it, and the protection scope of the present application is not limited to the above examples. Therefore, any equivalent changes or modifications made in accordance with the principles and design ideas disclosed by the present application are within the scope of protection of the present application.

Claims

1. A general-purpose localization segmentation method for gray-scale ultrasound image lesion sites, characterized by: The method comprises the following steps: S0: building a general positioning segmentation system for the lesion part of the grayscale ultrasound image; comprising a double-input double-output deep learning segmentation model, an adaptive target function module focusing on the size of the lesion part, a two-stage lesion part perception algorithm module, and a weight addition algorithm module fusing the outputs of the two branches; The double-input double-output deep learning segmentation model adopts a convolutional neural network as the backbone, comprising a first input branch, a second input branch, a first encoder, a second encoder, a decoder, a first output branch, and a second output branch; The first input branch is connected to the first encoder; The second input branch is connected to the second encoder; The outputs of the first encoder and the second encoder are merged and then connected to the decoder; the first encoder and the second encoder are connected to the decoder through a skip connection respectively; the decoder is connected to the first output branch and the second output branch respectively; the adaptive target function module focusing on the size of the lesion part comprises two binary cross-entropy loss functions with weights and a standard binary cross-entropy loss function; S1: training the model in combination with the double-input double-output deep learning segmentation model and the adaptive target function module focusing on the size of the lesion part, and extracting features related to the lesion part from the original ultrasound image; S2: the two-stage lesion part perception algorithm module performs progressive detection and lesion part positioning through multiple iterations, and updates the part of the ultrasound image including the lesion part after being intercepted by the second input branch input of the double-input double-output deep learning segmentation model; S3: the weight addition algorithm module fusing the outputs of the two branches comprehensively considers the learning effects of the double-input double-output deep learning segmentation model in different branches to obtain a more ideal final prediction result.

2. The method for universal positioning segmentation of gray-scale ultrasound image lesion site according to claim 1, characterized in that: In the step S1, the specific steps are as follows: S11: set y as artificial lesion annotation, For the model prediction result, N is the number of images input to the model, and λ is the weight coefficient; the definition of the weighted binary cross-entropy loss function is as follows: For the first input branch, set λ>0.5, which is used to regard the lesion part segmentation problem as a typical small target segmentation problem, and pay attention to suppressing false negatives in the prediction result during the training process of the model; For the second input branch, set λ<0.5, which is used to regard the lesion part segmentation problem as a general target segmentation problem, and pay attention to suppressing false positives in the prediction result during the training process of the model; For the first output branch and the second output branch, set λ=0.5, which is used to measure the similarity between the prediction results of the lesion part; S12: Set the prediction result of the first output branch, the prediction result of the second output branch, y OB1 the artificial label corresponding to the prediction result of the first output branch, y OB2 the artificial label corresponding to the prediction result of the second output branch, the intercepted partial image containing the lesion part in the prediction result of the first output branch; then the expression formula of the adaptive target function focusing on the size of the lesion site is:

3. The method for universal positioning segmentation of gray-scale ultrasound image lesion site according to claim 2, characterized in that: In the step S2, the specific steps are as follows: S21: assuming that the lateral resolution and the longitudinal resolution of the original ultrasound image US1 are α and β respectively, and the artificial annotation image corresponding to the original ultrasound image US1 is A1; assuming that the magnification parameter is EL, EL≥1 and EL∈R + ; assuming that the gradual cycle parameter is NL, NL≥1 and NL∈Z + ; a partial image US2 with a size of α / 2 and β / 2 is randomly cut from the original ultrasound image US1 respectively, and the annotation image A2 corresponding to the partial image US2 is obtained from the artificial annotation image A1 according to the coordinates of the cut. S22: training the double-input double-output deep learning segmentation model with US1, US2, A1, and A2 as training data and updating the model weight; S23: using the trained model to infer US1 and obtaining the prediction output FM1; S24: performing a binary operation on FM1 and finding the maximum contour to calculate the minimum bounding rectangle BB corresponding to the contour, and enlarging the BB by a specified multiple EL; if the maximum contour cannot be obtained, skip step S25; S25: based on the enlarged BB coordinates, intercept US1 / A1, and rescale the intercepted image to the size of α / 2 and β / 2, thereby obtaining new US2 and A2; S26: repeat steps S22-S25 for NL times.

4. The method for universal positioning segmentation of gray-scale ultrasound image lesion site according to claim 3, characterized in that: The step S2 is specifically:

5. The method for universal positioning segmentation of gray-scale ultrasound image lesion site according to claim 3, characterized in that: The step S3 is specifically: S31: obtaining true positive TP1, false positive FP1 and false negative FN1 by comparing the prediction output FM1 of the first output branch with the artificial labeled image A1; calculating the Dice index value D1 of the first output branch as: D1 = 2*TP1 / (2*TP1+FP1+FN1); S32: scaling FM2 by a ratio of 1 / EL, and replacing the corresponding part of FM1 with the scaled FM2 based on the BB coordinates obtained in step S25 to obtain a new FM2; S33: calculating the model prediction result of the fusion of the outputs of the two branches as: FMF = D1*FM1+D2*FM2. The dual-input dual-output deep learning segmentation model, the adaptive target function module focusing on the size of the lesion site, the two-stage lesion site perception algorithm module and the weight addition algorithm module for fusing the outputs of the two branches; The dual-input dual-output deep learning segmentation network and the adaptive target function module focusing on the size of the lesion site are used to efficiently extract features related to the lesion site from the ultrasound image; 6. A system for the general-purpose localization segmentation method of lesions in grayscale ultrasound images according to any one of claims 1 to 5, characterized in that: The two-stage lesion site perception algorithm module is used to realize the gradual detection and positioning of the lesion site through multiple iterations; The weight addition algorithm module for fusing the outputs of the two branches is used to provide the final prediction result of the network.

7. The system of claim 6, wherein: The dual-input dual-output deep learning segmentation model adopts a convolutional neural network as the backbone, and includes a first input branch, a second input branch, a first encoder, a second encoder, a decoder, a first output branch and a second output branch; The first input branch is connected to the first encoder; The second input branch is connected to the second encoder; The outputs of the first encoder and the second encoder are merged and then connected to the decoder; The first encoder and the second encoder are connected to the decoder through a skip connection respectively; The decoder is connected to the first output branch and the second output branch respectively; The first input branch is used to receive the original ultrasound image as input; The second input branch is used to receive the intercepted partial ultrasound image including the lesion site as input; The first encoder and the second encoder are respectively used to perform layer-by-layer encoding on the images output by the first input branch and the second input branch and obtain two feature maps with the same size and number of channels, and then splice the two feature maps in the channel dimension and input them into the decoder; The decoder is used to accept the feature maps from the first encoder and the second encoder at different convolution layers through a skip connection, and splice the feature maps with the internally decoded feature maps in the channel dimension to generate two feature maps with the same size as the input image; ​ ​ The first output branch is configured to output a probability map of a segmentation prediction result of the original ultrasound image in a feature map output by the decoder; The second output branch is configured to output a probability map of a segmentation prediction result of the intercepted partial ultrasound image including the lesion site in a feature map output by the decoder.

8. The system of claim 6, wherein: The adaptive objective function module focusing on the size of the lesion site includes two weighted binary cross-entropy loss functions and a standard binary cross-entropy loss function; the weighted binary cross-entropy loss functions are used to evaluate the error of the dual-branch output of the dual-input dual-output deep learning segmentation network and the corresponding artificial lesion annotation; the standard binary cross-entropy loss function is used to evaluate the similarity between the predictions of the dual-branch output for the lesion site.

9. A computer storage medium, characterized in that: The computer program stored therein can be executed by a computer processor, and the computer program performs the general positioning segmentation method for the lesion site of the grayscale ultrasound image as claimed in any one of claims 1 to 5.

Images