Interventional medical instrument image segmentation method and system

Abstract

The application provides an interventional medical instrument image segmentation method and system, comprising the following steps: acquiring DSA image data from an interventional surgery case database system and dividing the DSA image data into a training set and a test set; building a Unet model, converting an input picture into a semantic feature map through an Encoder module, converting the semantic feature map into a binary classification result through a Decoder module, completing resolution recovery, and obtaining a final segmentation result map; defining a weight function during the training process; defining a new energy function in combination with cross entropy and softmax; and fusing and mapping the segmentation result of the model to a DSA original map and highlighting the DSA original map. The application greatly improves the precision and stability of surgical operation by means of computer and robot technology, and can effectively reduce the harm of radiation to interventional doctors and the probability of intraoperative accidents.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to an interventional medical instrument image segmentation method and system, and more particularly to an interventional medical instrument image segmentation method and device, an electronic device and a storage medium. BACKGROUND

[0002] With the improvement of people's living standards and the aggravation of aging, as well as the youth of staying up late, cardiovascular and cerebrovascular diseases have become one of the main causes threatening people's health and safety. Because interventional surgery has the advantages of precision, rapidity, and small harm to patients, it has become the main means of treating cardiovascular and cerebrovascular diseases. Based on the current medical environment, all vascular interventional surgeries are completed under DSA images, which results in long-term exposure of doctors and patients to X-rays.

[0003] The invention patent with publication number CN114820473A discloses a medical image segmentation method, which extracts a region of interest from a picture, calculates a scaling matrix according to the damage area, scales the region of interest, segments the scaled image, and uses an uncertainty-guided method to supervise the segmentation and uses the damage area perception technology to process small sample segmentation in the sample. This method is only limited to CT medical images and cannot be used for DSA influence, and this method is for damage area segmentation and cannot segment interventional medical instruments.

[0004] The invention patent with publication number CN112712527A discloses a medical image segmentation method, which uses a DR-Unet104 model, standardization and data augmentation preprocessing methods to improve the model generalization ability, and combines the advantages of the basic Unet model and the residual connection module to segment medical images. This method expands the model parameters, so that the model training and inference need strong GPU hardware support, prolongs the model training time, and also slows down the inference result, which cannot be used in real-time surgery and cannot provide real-time information for doctors to judge.

[0005] Therefore, a new technical solution is needed to improve the above technical problems. SUMMARY

[0006] In view of the defects in the prior art, the purpose of the present application is to provide an interventional medical instrument image segmentation method and system.

[0007] According to the interventional medical instrument image segmentation method provided by the present application, the method comprises the following steps:

[0008] Step S1: Obtain the DSA image dataset from the interventional surgery case database system, limit the format and pixel size of the picture, define the standard tool, limit the class and boundary condition of the annotation; divide the dataset into training set and test set;

[0009] Step S2: build a training model by imitating the Unet neural network, convert the input picture into a semantic feature map through the Encoder module, convert the semantic feature map into a binary classification result through the repeated execution of the Decoder module and the last convolution dimension reduction processing, complete the resolution recovery, and obtain the final segmentation result map;

[0010] Step S3: define the weight function during the training process:

[0011]

[0012] Wherein: w m is the weight value of balancing the frequency of different interventional medical instrument categories such as background, catheter and balloon;

[0013] d a is the position to the nearest instrument edge;

[0014] d b is the position to the second nearest instrument edge;

[0015] w0 and σ are set to 9.8 and 4.9 through multiple experience calibration;

[0016] Step S4: during the training process, define the loss function combined with softmax to represent the difference between the model prediction result and the actual result;

[0017] The softmax function is as follows:

[0018]

[0019] Wherein: a k (x) is the activation value of a point at pixel position k in the feature channel;

[0020] K is the number of interventional medical instrument categories, wherein k = 1, …, K;

[0021] From k = 1 to the maximum number of categories K, when k has the maximum activation value, p k (x) is approximately equal to the maximum probability 1, and when k is other values, p k (x) is approximately equal to the minimum probability 0;

[0022] Step S5: during the DSA image fusion enhancement process, map the model segmentation result to the DSA original picture.

[0023] Preferably, the boundary condition of the data set labeling in the step S1 is that no more than 5%-10% of pixels.

[0024] Preferably, the step S1 obtains a DSA image data set from an interventional surgery case database system, outputs the DSA image in a standard pixel size, and all are drafted as a single-channel grayscale threshold picture data set; a labeling tool is used to label the data set, and a sharpening process is performed on medical instruments with fuzzy edges and small target areas; the data set after sharpening is smoothed through Gaussian filtering, and finally divided into a training set and a test set.

[0025] Preferably, the step S2 establishes an Encoder-Decoder structure of an interventional medical instrument image segmentation.

[0026] In the structure of the Encoder, the size of the input layer is defined as the size of the training set, a down-sampling module composed of two convolutional layers and a pooling layer is used to perform a down-sampling operation to obtain semantic features of the interventional medical instrument; the Encoder structure converts the resolution of the test set image from high to low in an encoding form.

[0027] In the structure of the Decoder, an up-sampling module composed of an inverse convolution and a feature concatenation layer plus two convolutional layers is used to perform an up-sampling operation corresponding to the Encoder, and finally a convolutional layer is used for dimension reduction processing to convert the feature map into a binary classification result, obtaining the final segmentation result image; the Decoder structure converts the resolution of the test set image from low to high in a decoding form.

[0028] Preferably, the step S5 verifies the accuracy of the trained and converged model on the test set; and the configuration of the segmentation result display module is that the original image of the current frame and the target object highlighted in the segmentation result are superimposed, different target objects are displayed with different labels, and the obtained segmentation result image is compared with the test set image.

[0029] The application also provides an interventional medical instrument image segmentation system, which comprises the following modules:

[0030] Module M1: obtaining a DSA image data set from an interventional surgery case database system, limiting the format and pixel size of the picture, defining a standard tool, limiting the category and boundary condition of the labeling, and dividing the data set into a training set and a test set;

[0031] Module M2: building a training model by imitating a Unet neural network, converting an input picture into a semantic feature map through an Encoder module, converting the semantic feature map into a binary classification result through repeated execution of a Decoder module and final convolutional dimension reduction processing, completing resolution recovery, and obtaining a final segmentation result image;

[0032] Module M3: Define the weight function in the training process:

[0033]

[0034] wherein: w m is the weight value of balancing the frequency of different categories of interventional medical devices such as background, catheter, balloon, etc.;

[0035] d a is the position to the nearest device edge;

[0036] d b is the position to the second nearest device edge;

[0037] w0 and σ are set to 9.8 and 4.9 through multiple empirical calibrations;

[0038] Module M4: In the training process, the loss function is defined in combination with softmax to represent the difference between the model prediction result and the actual result.

[0039] The softmax function is as follows:

[0040]

[0041] wherein: a k (x) is the activation value of the pixel position of point a at feature channel k;

[0042] K is the number of categories of interventional medical devices, wherein k = 1, …, K;

[0043] From k = 1 to the maximum number of categories K, when k has the maximum activation value, p k (x) is approximately the maximum probability 1, and when k is other values, p k (x) is approximately the minimum probability 0;

[0044] Module M5: In the DSA image fusion enhancement process, the model segmentation result is fused and mapped to the DSA original image.

[0045] Preferably, the boundary condition of the data set annotation in the module M1 is that the pixel ratio is not more than 5%-10%.

[0046] Preferably, the module M1 obtains the DSA image data set from the interventional surgery case database system, outputs the DSA image in a standard pixel size, and all are designed as single-channel grayscale threshold picture data sets; the data set is labeled using a labeling tool, and the medical devices with blurred edges and small target areas are sharpened; the sharpened data set is smoothed by Gaussian filtering, and finally divided into a training set and a test set.

[0047] Preferably, the module M2 establishes an Encoder-Decoder structure of interventional medical instrument image segmentation;

[0048] In the structure of the Encoder, the size of the input layer is defined as the size of the training set, a down-sampling module composed of two convolutional layers and a pooling layer performs a down-sampling operation to obtain semantic features of the interventional medical instrument; the Encoder structure converts the resolution of the test set image from high to low in an encoding manner.

[0049] In the structure of the Decoder, an up-sampling module composed of an inverse convolutional and feature concatenation layer and two convolutional layers performs an up-sampling operation corresponding to the Encoder, and finally a convolutional layer is used for dimension reduction processing to convert the feature map into a binary classification result, thereby obtaining the final segmentation result image; the Decoder structure converts the resolution of the test set image from low to high in a decoding manner.

[0050] Preferably, the module M5 verifies the accuracy of the trained convergence model on the test set; and the segmentation result display module is configured to superimpose the original image of the current frame and the highlight display of the target object in the segmentation result, different target objects are displayed with different labels, and the obtained segmentation result image is compared with the test set image.

[0051] Compared with the prior art, the present application has the following beneficial effects:

[0052] 1. The present application builds a training model through the Unet neural network to realize image segmentation of interventional medical instruments; the mean IOU value is introduced as a standard for evaluating the model, and the efficiency of the present application for interventional medical instrument image segmentation is verified through verification on the DSA image data set.

[0053] 2. The present application can enable doctors to monitor the image segmentation of interventional medical instruments in the automatic operation process of the robot in real time through remote supervision, which can avoid the exposure of doctors to X-rays.

[0054] 3. The present application can accelerate the entire interventional operation process, reduce the extension of the operation time caused by the tension and fatigue of the doctors, and indirectly reduce the time for which the patient is exposed to X-rays during the interventional operation. BRIEF DESCRIPTION OF DRAWINGS

[0055] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments, with reference to the accompanying drawings:

[0056] Figure 1 The original image of the DSA data set of the present application is shown in Figure 1.

[0057] Figure 2The test set of the present application is annotated image;

[0058] Figure 3 The model segmentation result image of the present application;

[0059] Figure 4 The DSA image enhanced display image;

[0060] Figure 5 The interventional medical instrument image segmentation method flowchart of the present application; DETAILED DESCRIPTION

[0061] The present application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be pointed out that for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made. These all belong to the protection scope of the present application.

[0062] Example 1:

[0063] Reference Figure 5 According to the interventional medical instrument image segmentation method provided by the present application, the method comprises the following steps:

[0064] Step S1: Obtain the DSA image data set from the interventional surgery case database system, limit the format and pixel size of the picture, define the standard tool, limit the class and boundary condition of the annotation, divide the data set into training set and test set, the boundary condition of the data set annotation is that the pixel ratio is not more than 5%-10%, obtain the DSA image data set from the interventional surgery case database system, output the DSA image in standard pixel size, all of which are drafted as single-channel gray threshold picture data set, use the annotation tool to annotate the data set, sharpen the medical instrument with fuzzy edge and small target area, and finally divide the sharpened data set into training set and test set through Gaussian filtering for smoothing processing.

[0065] Step S2: build a training model by imitating Unet neural network, convert the input picture into semantic feature map through the Encoder module, convert the semantic feature map into binary classification result through repeated execution of the Decoder module and final convolution dimension reduction processing, complete the resolution recovery, and obtain the final segmentation result image; establish the Encoder-Decoder structure of the interventional medical instrument image segmentation;

[0066] In the structure of the Encoder, the size of the input layer is defined as the size of the training set, a down-sampling module composed of two convolutional layers and a pooling layer performs down-sampling operation to obtain the semantic features of the interventional medical instrument; the Encoder structure converts the resolution of the test set image from high to low in an encoding form.

[0067] In the structure of the Decoder, an up-sampling module composed of an inverse convolution and a feature concatenation layer plus two convolutional layers performs up-sampling operation, corresponding to the Encoder, and finally a convolutional layer is used for dimension reduction processing to convert the feature map into a binary classification result, obtaining the final segmentation result map; the Decoder structure converts the resolution of the test set image from low to high in a decoding form.

[0068] Step S3: Define the weight function in the training process:

[0069]

[0070] wherein: w m is the weight value balancing the frequency of different interventional medical instrument categories such as background, catheter and balloon;

[0071] d a is the position to the nearest instrument edge;

[0072] d b is the position to the second nearest instrument edge;

[0073] w0 and σ are set to 9.8 and 4.9 through multiple empirical calibrations;

[0074] Step S4: In the training process, the loss function is defined in combination with softmax to represent the difference between the model prediction result and the actual result;

[0075] The softmax function is as follows:

[0076]

[0077] wherein: a k (x) is the activation value of the pixel position at point a in feature channel k;

[0078] K is the number of interventional medical instrument categories, wherein k = 1, …, K;

[0079] From k = 1 to the maximum category number K, when k has the maximum activation value, p k (x) is approximately equal to the maximum probability 1, and when k is other values, p k (x) is approximately equal to the minimum probability 0;

[0080] Step S5: In the DSA image fusion enhancement process, the model segmentation result is fused and mapped to the DSA original image, and is highlighted; based on the trained converged model, the accuracy thereof is verified on the test set; and the segmentation result display module is configured as: the superposition of the original image of the current frame and the target object highlighted in the segmentation result, different target objects are displayed with different labels, and the obtained segmentation result image is compared with the test set image.

[0081] Example 2:

[0082] Embodiment 2 is a preferred example of Embodiment 1, which more specifically illustrates the present application.

[0083] The present application also provides an interventional medical instrument image segmentation system, which comprises the following modules:

[0084] Module M1: Obtain the DSA image data set from the interventional surgery case database system, limit the format and pixel size of the picture, define the standard tool, limit the labeled category and boundary condition; divide the data set into a training set and a test set; the boundary condition of the data set labeling is that the pixel ratio is not more than 5%-10%; obtain the DSA image data set from the interventional surgery case database system, output the DSA image in a standard pixel size, and all are drafted as a single-channel grayscale threshold picture data set; use the labeling tool to label the data set, and perform sharpening processing on the medical instrument with blurred edges and small target area; the sharpened data set is smoothed by Gaussian filtering, and finally divided into a training set and a test set.

[0085] Module M2: build a training model by imitating the Unet neural network, convert the input picture into a semantic feature map through the Encoder module, convert the semantic feature map into a binary classification result through repeated execution of the Decoder module and final convolution dimension reduction processing, complete the resolution recovery, obtain the final segmentation result image, and establish the Encoder-Decoder structure of the interventional medical instrument image segmentation;

[0086] In the structure of the Encoder, the size of the input layer is defined as the size of the above-mentioned training set, a down-sampling module composed of two convolution layers and one pooling layer is formed, and a down-sampling operation is performed to obtain the semantic features of the interventional medical instrument; the Encoder structure converts the resolution of the test set image into high edge and low resolution in an encoding form.

[0087] In the structure of Decoder, an up-sampling module is composed of an anti-convolution and feature splicing layer and two convolution layers. After the up-sampling operation, the corresponding Encoder is passed through the convolution layer dimension reduction processing, the feature map is converted into the result of binary classification, and the final segmentation result map is obtained. The resolution of the test set image is changed from low to high in the decoding form of Decoder structure.

[0088] Module M3: define the weight function during the training process:

[0089]

[0090] wherein: w m is the weight value of balancing the frequency of different categories of interventional medical instruments such as background, catheter and balloon;

[0091] d a is the position to the nearest instrument edge;

[0092] d b is the position to the second nearest instrument edge;

[0093] w0 and σ are set to 9.8 and 4.9 through multiple experience calibration;

[0094] Module M4: during the training process, the loss function is defined by combining softmax to represent the difference between the model prediction result and the actual result;

[0095] The softmax function is as follows:

[0096]

[0097] wherein: a k (x) is the activation value of a point at pixel position k in the feature channel;

[0098] K is the number of categories of interventional medical instruments, wherein k = 1, …, K;

[0099] From k = 1 to the maximum number of categories K, when k has the maximum activation value, p k (x) is approximately equal to the maximum probability 1, and when k is other values, p k (x) is approximately equal to the minimum probability 0;

[0100] Module M5: during the DSA image fusion enhancement process, the segmentation result of the model is fused and mapped to the DSA original image; based on the trained converged model, the accuracy thereof is verified on the test set; and the segmentation result display module is configured as: the superposition of the original image of the current frame and the highlight display of the target object of the segmentation result, different target objects are displayed with different labels, and the obtained segmentation result map is compared with the test set image.

[0101] Example 3:

[0102] Embodiment 3 is a preferred example of Embodiment 1 to more specifically illustrate the present application.

[0103] The present application relates to the field of medical image processing, and belongs to the field of medical image segmentation. Specifically, through a model trained based on Unet, the features of interventional medical devices such as guide wires, catheters, balloons and stents annotated by humans are learned, the basic features of the medical devices are stored in the weight values connected between the neurons of the model, so as to complete the training and obtain the final model for prediction, thereby completing the segmentation of the guide wires, catheters, balloons and other medical devices under the current real-time DSA image. After the current segmentation result is enhanced in real time in the DSA image, it is displayed to the doctor, which can help the doctor or computer or surgical robot to judge the pushing position of the guide wire, catheter, balloon and stent and other medical devices in the interventional operation in real time, feed back whether the pushing of the device is smooth, and assist the doctor or computer or surgical robot to make more accurate judgments on the travel distance and planned path of the device.

[0104] In view of the current surgical defects, the present application develops a vascular interventional automatic surgical navigation system. First, the doctor can monitor the automatic surgical process of the robot in real time through remote supervision, which can avoid the doctor from being irradiated by X-rays. Second, the process of the entire interventional surgery can be accelerated, which reduces the extension of the operation time of the doctor due to nervousness and fatigue, and indirectly reduces the time of the patient in the interventional surgery irradiated by X-rays. The current method for realizing the navigation system mainly consists of two parts of interventional medical device image segmentation and DSA image enhancement. Through the segmentation of the interventional medical devices such as guide wires, catheters, balloons and stents, it can be judged whether the current wire is pushed smoothly, and through the real-time display of the small parts such as the head of the guide wire and catheter, the doctor can evaluate the position of the current catheter and control the entire automatic surgical process. The significance of DSA image enhancement is to highlight the position of the interventional device under the current DSA image, so that the doctor can have an intuitive effect on the position of the entire guide wire and catheter, greatly improving the precision and stability of the operation, and effectively reducing the harm of radiation to the interventional doctor and the probability of intraoperative accidents.

[0105] The present application proposes an interventional medical device image segmentation method. The present application mainly includes the following five processes:

[0106] (1) Data preparation:

[0107] The DSA image data set was obtained from the interventional surgery case database system, and the DSA image was output in standard pixel size, all of which were prepared as single-channel grayscale threshold picture data sets; the data set was labeled using a labeling tool, and the medical instruments with blurred edges and small target areas were sharpened; the sharpened data set was smoothed by Gaussian filtering, and finally divided into training set and test set.(2) Model structure:

[0108] The training model was built by imitating the Unet neural network, and the Encoder-Decoder structure of interventional medical instrument image segmentation was established.

[0109] In the structure of Encoder, the size of the input layer is defined as the size of the training set mentioned above, which consists of two convolutional layers and a pooling layer to form a down-sampling module to perform down-sampling operation to obtain the semantic features of interventional medical instruments; the Encoder structure encodes the resolution of the test set image from high to low.

[0110] In the structure of Decoder, an up-sampling module is composed of an inverse convolution and a feature concatenation layer plus two convolutional layers, which is processed by up-sampling operation corresponding to Encoder, and finally processed by convolutional layer to reduce dimension, which converts the feature map to binary classification result to obtain the final segmentation result image; the Decoder structure decodes the resolution of the test set image from low to high.

[0111] (3) Model training:

[0112] The random gradient descent method is used to train the error reduction and weight change during training. In order to make some interventional instruments more important, we introduce w(x) to calculate a weight value for each labeled image, so that the model pays more attention to the small segmentation boundary between interventional medical instruments and background. The formula of weight value is as follows:

[0113]

[0114] Where: w m is the weight value of balancing the frequency of different interventional medical instrument categories such as background, catheter and balloon;

[0115] d a is the position to the nearest instrument edge;

[0116] d b is the position to the second nearest instrument edge;

[0117] w0 and σ are set to 9.8 and 4.9 through multiple experience calibration;

[0118] (4) Loss function of the model:

[0119] In the last layer of the model, the loss function is defined by combining softmax to represent the difference between the model prediction result and the actual category.

[0120] The softmax function is as follows:

[0121]

[0122] wherein: a k (x) is the activation value of the pixel position of point a at feature channel k;

[0123] K is the number of categories of interventional medical instruments, wherein k = 1, …, K;

[0124] From k = 1 to the maximum number of categories K, when there is the maximum activation value at k, p k (x) is approximately the maximum probability 1, and when k is other values, p k (x) is approximately the minimum probability 0;

[0125] (5) DSA image fusion enhancement:

[0126] Based on the trained and converged model, the accuracy is verified on the test set. The comparison between the obtained model segmentation inference graph and the test set image is shown in Figure 1 , Figure 2 and Figure 3 .

[0127] After obtaining the final segmentation result graph by the model, the configuration of the segmentation result display module is set as: the superposition of the original picture of the current frame of DSA image and the highlight display of the target object of the segmentation result, different target objects are displayed with different labels, the segmentation result graph is fused and enhanced on the DSA, and the effect is shown in Figure 4 .

[0128] An interventional medical instrument image segmentation device, the image segmentation device comprises:

[0129] A data acquisition module configured to acquire a DSA image data set from an interventional surgery case database system;

[0130] A data labeling module configured to divide the DSA image data set into a training set and a test set based on the DSA image data set, wherein the training set image comprises at least one image-segmentable medical instrument;

[0131] A segmentation module configured to segment the original DSA image to obtain a segmentation result of the at least one image-segmentable medical instrument.

[0132] An electronic device for image segmentation of an interventional medical instrument, the image segmentation device comprising a processor and a memory, the memory for storing program data, the processor for executing the program data to implement the steps and methods in the above.

[0133] A computer readable storage medium for storing program data, the program data being executable to implement the steps and methods in the above.

[0134] The present application provides an image segmentation method and system for an interventional medical instrument. The method builds a training model through an Unet neural network to realize segmentation of the interventional medical instrument. The mean IOU value is introduced as a standard for evaluating the model through verification on a data set, and the efficiency of the present application for image segmentation of the interventional medical instrument is verified.

[0135] Those skilled in the art can understand the present embodiment as a more specific description of embodiment 1 and embodiment 2.

[0136] Those skilled in the art know that, in addition to implementing the system provided by the present application and each device, module and unit thereof in a pure computer readable program code manner, the system provided by the present application and each device, module and unit thereof can also be implemented in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps to achieve the same functions. Therefore, the system provided by the present application and each device, module and unit thereof can be considered as a hardware component, and the devices, modules and units included therein for achieving various functions can also be considered as structures within the hardware component; the devices, modules and units for achieving various functions can also be considered as both software modules for implementing methods and structures within hardware components.

[0137] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the present application. In the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. An interventional medical instrument image segmentation method, characterized by, The method comprises the following steps: Step S1: obtaining a DSA image data set from an interventional surgery case database system, limiting the format and pixel size of the picture, defining a standard tool, limiting the category and boundary condition of the annotation, and dividing the data set into a training set and a test set; Step S2: building a training model by imitating a Unet neural network, converting the input picture into a semantic feature map through an Encoder module, converting the semantic feature map into a binary classification result through repeated execution of a Decoder module and final convolution dimension reduction processing, completing resolution recovery, and obtaining a final segmentation result map; Step S3: defining a weight function during the training process: wherein: w m is a weight value balancing the frequencies of different interventional medical instrument categories, such as background, catheter, balloon, etc. d a is representative of the position to the nearest instrument edge; d b the position of the instrument edge to be changed to the second proximity w0 and sigma are set to 9.8 and 4.9 through multiple empirical calibrations; Step S4: during the training process, a loss function is defined in combination with softmax to represent the difference between the model prediction result and the actual result; The softmax function is as follows: where: a k (x) is the activation value for a point a at pixel position in feature channel k; K is the number of types of interventional medical instruments, wherein k = 1, …, K; From k = 1 to maximum number of categories K, when k has the maximum activation value, p k (x) is approximated to the maximum probability 1, when k has other values, p k (x) is approximated to the minimum probability 0; Step S5: during the DSA image fusion enhancement process, the segmentation result of the model is fused and mapped onto the DSA original picture.

2. The interventional medical instrument image segmentation method of claim 1, wherein, The boundary condition of the data set annotation in the step S1 is that the pixel proportion is not more than 5%-10%.

3. The interventional medical instrument image segmentation method of claim 1, wherein, In the step S1, the DSA image data set is obtained from the interventional surgery case database system, the DSA image is output in a standard pixel size, and all are drafted as single-channel grayscale threshold picture data sets; the data set is annotated by using an annotation tool, and the medical instruments with fuzzy edges and small target areas are sharpened; the sharpened data set is smoothed through Gaussian filtering, and finally divided into a training set and a test set.

4. The interventional medical instrument image segmentation method of claim 1, wherein, The step S2 establishes an Encoder-Decoder structure of the interventional medical instrument image segmentation; In the structure of the Encoder, the size of the input layer is defined as the size of the above-mentioned training set, a lower sampling module is composed of two convolution layers and a pooling layer, a lower sampling operation is performed to obtain the semantic features of the interventional medical instrument, the Encoder structure converts the resolution of the test set image from high to low in an encoding form; In the structure of the Decoder, an upper sampling module is composed of an inverse convolution and a feature splicing layer plus two convolution layers, an upper sampling operation is performed, corresponding to the Encoder, and finally a convolution layer dimension reduction processing is performed to convert the feature map into a binary classification result, and a final segmentation result map is obtained; the Decoder structure converts the resolution of the test set image from low to high in a decoding form.

5. The interventional medical instrument image segmentation method of claim 1, wherein, The step S5 verifies the accuracy of the converged model on the test set based on the trained model, and the configuration of the segmentation result display module is that the original picture of the current frame and the target object highlight display of the segmentation result are superimposed, different target objects are displayed with different labels, and the obtained segmentation result map is compared with the test set image.

6. An interventional medical instrument image segmentation system, characterized by, The system comprises the following modules: Module M1: obtaining a DSA image data set from an interventional surgery case database system, limiting the format and pixel size of the picture, defining a standard tool, limiting the category and boundary condition of the annotation, and dividing the data set into a training set and a test set; Module M2: a training model is built by imitating the Unet neural network, the input picture is converted into a semantic feature map through the Encoder module, the semantic feature map is converted into a binary classification result through the repeated execution of the Decoder module and the final convolution dimension reduction processing, the resolution is restored, and the final segmentation result map is obtained; Module M3: define the weight function during training: where: w m is the weight value balancing the frequencies of different interventional medical instrument categories, such as background, catheter, balloon, etc. d a is representative of the position to the nearest instrument edge; d b the position of the instrument edge to be changed to the second proximity w0 and σ are set to 9.8 and 4.9 through multiple empirical calibrations; Module M4: during training, define the loss function in combination with softmax to represent the difference between the model prediction result and the actual result; The Softmax function is as follows: where: a k (x) is the activation value for a point a at pixel position in feature channel k; K is the number of interventional medical devices, where k = 1, …, K; From k = 1 to maximum number of categories K, when k has the maximum activation value, p k (x) is approximated to the maximum probability 1, when k has other values, p k (x) is approximated to the minimum probability 0; Module M5: during the DSA image fusion enhancement process, the segmentation result of the model is fused and mapped to the DSA original image.

7. The interventional medical instrument image segmentation system of claim 6, wherein, The boundary condition of the data set annotation in the module M1 is that the pixel ratio is not more than 5%-10%.

8. The interventional medical instrument image segmentation system of claim 6, wherein, The module M1 obtains the DSA image data set from the interventional surgery case database system, outputs the DSA image in a standard pixel size, and all are designed as single-channel grayscale threshold picture data sets; the data set is labeled using a labeling tool, and the medical devices with fuzzy edges and small target areas are sharpened; the sharpened data set is smoothed by Gaussian filtering, and finally divided into a training set and a test set.

9. The interventional medical instrument image segmentation system of claim 6, wherein, The module M2 establishes an Encoder-Decoder structure for interventional medical device image segmentation. In the structure of the Encoder, the size of the input layer is defined as the size of the training set, a down-sampling module is composed of two convolution layers and a pooling layer, and a down-sampling operation is performed to obtain the semantic features of the interventional medical device; the Encoder structure converts the resolution of the test set image from high to low in an encoding manner. In the structure of the Decoder, an up-sampling module is composed of an inverse convolution and a feature concatenation layer plus two convolution layers, which corresponds to the Encoder after the up-sampling operation, and finally the feature map is converted into a binary classification result through the convolution layer dimension reduction processing to obtain the final segmentation result map; the Decoder structure converts the resolution of the test set image from low to high in a decoding manner.

10. The interventional medical instrument image segmentation system of claim 6, wherein, The module M5 verifies the accuracy of the trained and converged model on the test set; and the segmentation result display module is configured to superimpose the original image of the current frame and the highlight display of the target object of the segmentation result, different target objects are displayed with different labels, and the obtained segmentation result map is compared with the test set image.

Images