A 3D medical image segmentation method based on cue-driven mechanism

Abstract

The present invention discloses a three-dimensional medical image segmentation method based on a prompt-driven mechanism, wherein the training process of the image segmentation model adopted by the three-dimensional medical image segmentation method includes: obtaining a data set, and performing preprocessing operations on label samples in the data set, wherein the label samples include a training set and a test set, and the data set includes a medical volume image and a volume segmentation image; improving the three-dimensional adaptation and layering strategy of each module in the visual segmentation basic model, realizing the modification of the image encoder and the prompt encoder in the visual segmentation basic model, and obtaining a training model in which the prompt embedding and the image embedding are aligned; adopting an iterative training strategy of simulated interaction, training the training model based on the training set, and evaluating and saving the model with the best evaluation index based on the test set, and obtaining the image segmentation model. The image segmentation model based on the prompt-driven mechanism of the present invention can improve the overall segmentation accuracy of medical images.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a three-dimensional medical image segmentation method based on a prompt-driven mechanism. Background Art

[0002] Organ segmentation is a task with important research significance and application value in clinical medicine, involving key steps such as lesion detection and surgical planning. Traditional segmentation methods rely on manual drawing by experts. High-quality segmentation tasks require not only rich clinical experience, but also long-term and highly concentrated attention, and are more susceptible to human errors. The emergence of deep learning technology provides an effective solution to this problem. Data-driven deep learning methods can automatically extract image features from data, quickly locate lesion areas, achieve fast and accurate organ segmentation, and improve segmentation efficiency and accuracy. With the continuous advancement of technology, deep learning and multi-organ segmentation technology will provide more powerful support for medical diagnosis.

[0003] At present, although the organ segmentation methods or systems based on supervised learning have been able to approach or even surpass the expert level in certain practical scenarios, such methods have significant limitations. First, the organ segmentation methods based on fully supervised learning are highly dependent on a large amount of accurately labeled data, and the labeling cost is high; while the organ segmentation methods based on semi-supervised learning can use unlabeled data to reduce the demand for labeled data, but the model design is more complex. Secondly, the supervised learning method mainly learns based on the labeled information in the training data, so its generalization ability is greatly affected by the training data. The supervised method cannot cover all possible image changes and anatomical structures. At present, the organ segmentation methods based on supervised learning are designed for specific downstream tasks and lack generalization ability. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a 3D medical image segmentation method based on a prompt-driven mechanism in view of the above-mentioned defects of the prior art. The specific scheme is as follows:

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0006] In a first aspect, the present invention provides an image segmentation model training method based on a prompt-driven mechanism, wherein the method comprises:

[0007] Acquire a data set, and perform a preprocessing operation on labeled samples in the data set, wherein the labeled samples include a training set and a test set, and the data set includes a medical volume image and a volume segmentation image;

[0008] Improve the three-dimensional adaptation and layering strategy of each module in the visual segmentation basic model, modify the image encoder and the prompt encoder in the visual segmentation basic model, and obtain a training model in which the prompt embedding and the image embedding are aligned;

[0009] The image segmentation model is obtained by adopting an iterative training strategy of simulated interaction, training the training model based on the training set, and evaluating and saving the model with the best evaluation index based on the test set.

[0010] In one implementation, the preprocessing operation includes segmentation labels, resampling, and grayscale normalization; wherein the segmentation labels include label division of the volume segmentation image, separating the segmentation labels of each organ and keeping them aligned with the medical volume image;

[0011] The resampling includes adjusting the voxel spacing of the image to the median size of the voxel spacing of the data set using nearest neighbor interpolation, and resampling the medical volume image to a volume image of a preset scale with its corresponding label segmentation as the center;

[0012] The grayscale normalization includes normalizing the medical volume image to [0, 1] using a minimum-maximum normalization method, and adjusting the pixel intensity to [-1, 1] by a linear transformation.

[0013] In one implementation, the improvement of the three-dimensional adaptation and layering strategy of each module in the visual segmentation basic model includes:

[0014] Converting a two-dimensional image encoder in the visual segmentation basic model into a three-dimensional encoder for three-dimensional data;

[0015] A hierarchical strategy is introduced to extract image features at different depths;

[0016] Adjust the cue encoder structure so that the visual segmentation base model can utilize the semantic features corresponding to the cue;

[0017] According to the modifications of the image encoder and the hint encoder, the image decoder is modified to complete the fusion of the hint position information and the semantic information, realize the alignment of the hint embedding and the image embedding and output the corresponding mask.

[0018] In one implementation, the step of introducing a hierarchical strategy to extract image features at different depths includes:

[0019] The image encoding process is divided into four progressive levels, and a visual sampling cue is configured for each level, so that the visual segmentation base model can capture and embed multi-level image features to form multi-level cue feature embedding.

[0020] In one implementation, the workflow of the visual sampling prompter includes:

[0021] Introduce several global tokens for cue feature embedding and apply self-attention technology;

[0022] After adding the updated cue feature embedding to the cue position embedding, self-attention technology is applied to fuse the position information with the semantic information.

[0023] The updated global query is applied to the cross-attention between the image encoding to align the hint embedding and the image embedding.

[0024] In one implementation, the strategy of iterative training using simulated interaction, training the training model based on the training set, includes:

[0025] In each round of training, each training sample is iterated 5 times. During the first prediction, the training sample is fed into the training model and the position coordinates of a positive sample are randomly generated as prompt information. After that, each iteration will generate random multi-point prompts based on the prediction results of the previous iteration and input them into the training model. The prediction results of each iteration are respectively used to calculate the loss function with the corresponding label samples, and the parameters of the training model are updated through gradient backpropagation.

[0026] Among them, except for the first input, the image encoder of the training model does not participate in parameter updating.

[0027] In a second aspect, an embodiment of the present invention further provides a three-dimensional medical image segmentation method based on a prompt-driven mechanism, wherein the three-dimensional medical image segmentation method adopts an image segmentation model trained by the image segmentation model training method based on a prompt-driven mechanism described in any one of the above-mentioned schemes.

[0028] In a third aspect, an embodiment of the present invention further provides an image segmentation model training system based on a prompt-driven mechanism, wherein the system comprises:

[0029] A data preprocessing module, used to obtain a data set and perform preprocessing operations on label samples in the data set, wherein the label samples include a training set and a test set, and the data set includes a medical volume image and a volume segmentation image;

[0030] A model improvement module, used to improve the three-dimensional adaptation and layering strategy of each module in the visual segmentation basic model, to modify the image encoder and the prompt encoder in the visual segmentation basic model, and to obtain a training model in which the prompt embedding and the image embedding are aligned;

[0031] The model training module is used to adopt an iterative training strategy of simulated interaction, train the training model based on the training set, and evaluate and save the model with the best evaluation index based on the test set to obtain the image segmentation model.

[0032] In a fourth aspect, an embodiment of the present invention further provides a terminal, wherein the terminal includes a memory, a processor, and an image segmentation model training program based on a prompt-driven mechanism or a three-dimensional medical image segmentation program based on a prompt-driven mechanism stored in the memory and executable on the processor, wherein when the processor executes the image segmentation model training program based on the prompt-driven mechanism, the processor implements the steps of the image segmentation model training method based on the prompt-driven mechanism as described in any one of the above-mentioned schemes, and when the processor executes the three-dimensional medical image segmentation program based on the prompt-driven mechanism, the processor implements the steps of the three-dimensional medical image segmentation method based on the prompt-driven mechanism as described.

[0033] In a fifth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores an image segmentation model training program based on a prompt-driven mechanism or a three-dimensional medical image segmentation program based on a prompt-driven mechanism, and when the image segmentation model training program based on the prompt-driven mechanism is executed by a processor, the steps of the image segmentation model training method based on the prompt-driven mechanism as described in any of the above schemes are implemented, and when the three-dimensional medical image segmentation program based on the prompt-driven mechanism is executed by a processor, the steps of the three-dimensional medical image segmentation method based on the prompt-driven mechanism as described are implemented.

[0034] Beneficial effects: Compared with the prior art, the present invention provides an image segmentation model training method based on a prompt-driven mechanism, and the training process of the image segmentation model adopted by the three-dimensional medical image segmentation method includes: first, obtaining a data set, and performing a preprocessing operation on the labeled samples in the data set, the labeled samples include a training set and a test set, and the data set includes a medical volume image and a volume segmentation image. Then, the three-dimensional adaptation and layering strategy are improved for each module in the visual segmentation basic model, and the image encoder and the prompt encoder in the visual segmentation basic model are modified to obtain a training model in which the prompt embedding and the image embedding are aligned. Finally, the training model is trained based on the training set using a strategy of iterative training with simulated interaction, and the model with the best evaluation index is evaluated and saved based on the test set to obtain the image segmentation model. The image segmentation model based on the prompt-driven mechanism of the present invention improves the segmentation performance of the prompt-driven model on medical images and realizes accurate segmentation of medical three-dimensional images. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1A flowchart of a specific implementation of the image segmentation model training method based on a prompt-driven mechanism provided in an embodiment of the present invention.

[0036] Figure 2 A framework diagram of a basic model for visual segmentation in an image segmentation model training method based on a prompt-driven mechanism provided in an embodiment of the present invention.

[0037] Figure 3 A schematic diagram of the structure of a visual sampling prompter in an image segmentation model training method based on a prompt-driven mechanism provided in an embodiment of the present invention.

[0038] Figure 4 A principle block diagram of an image segmentation model training system based on a prompt-driven mechanism provided in an embodiment of the present invention.

[0039] Figure 5 A functional block diagram of a terminal provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0041] Based on the defects of the prior art, this embodiment provides an image segmentation model training method based on a prompt-driven mechanism, which can be applied to a terminal, and the terminal can be a product terminal such as a computer. Specifically, Figure 1 As shown in , the image segmentation model training method based on the prompt-driven mechanism of this embodiment includes the following steps:

[0042] Step S100: Acquire a data set, and perform preprocessing operations on labeled samples in the data set, wherein the labeled samples include a training set and a test set, and the data set includes a medical volume image and a volume segmentation image.

[0043] Specifically, this embodiment can use five public data sets, such as AbdomenCT-1K, BTCV-Abdomen, BTCV-Cervi x, FLARE22, and Ki PA22. The data set involves 17 human organs, including 561 medical volume images and 3184 volume segmentation images. This embodiment divides the data set into 2894 training sets and 290 test sets to evaluate the segmentation performance. Further, this embodiment performs a preprocessing operation on the data set, and the preprocessing operation includes segmentation labeling, resampling, and grayscale normalization. Specifically, since the prompt-driven image segmentation model does not segment according to the category of the target, but only segments the target according to the prompt input, the segmentation label includes labeling the volume segmentation image, separating the segmentation labels of each organ and keeping them aligned with the medical volume image. After this step of operation, there are 3184 volume segmentation images for the training task of this embodiment, of which 2894 and 290 slices are used for organ segmentation training and testing, respectively. Since medical volume images occupy a large amount of memory, it is necessary to resample the medical volume images. The resampling includes using nearest neighbor interpolation to adjust the voxel interval of the image to the median size of the voxel interval of the data set, and resampling the medical volume image to a volume image of a preset scale with its corresponding label segmentation as the center. Since different medical volume images may have great differences in grayscale distribution, the non-centralized grayscale distribution of the image will directly affect the subsequent feature extraction and recognition. Therefore, the medical volume image needs to be grayscale normalized. The grayscale normalization includes using the minimum-maximum normalization method to normalize the medical volume image to [0,1], and adjusting the pixel intensity to [-1,1] through linear transformation.

[0044] Step S200, improve the three-dimensional adaptation and layering strategy of each module in the visual segmentation basic model, modify the image encoder and prompt encoder in the visual segmentation basic model, and obtain a training model in which prompt embedding and image embedding are aligned.

[0045] The model of the present invention adopts SAM (Segment Anything Model) as the basic model of visual segmentation, and improves its modules with three-dimensional adaptation and layering strategies. Specifically, it includes: converting the two-dimensional image encoder in the basic model of visual segmentation into a three-dimensional encoder for three-dimensional data, and introducing a layering strategy to extract image features at different depths. Then, the prompt encoder structure is adjusted so that the basic model of visual segmentation can utilize the semantic features corresponding to the prompt. Next, according to the modifications of the image encoder and the prompt encoder, the image decoder is modified, and the image decoder is modified so that it can complete the fusion of the prompt's position information and semantic information, realize the alignment of the prompt embedding and the image embedding, and output the corresponding mask. The overall framework diagram of the basic model of visual segmentation is shown in the figure. Figure 2 shown.

[0046] In this embodiment, the purpose of adopting a multi-level sampling strategy is to capture and embed multi-level image features, thereby improving the richness and accuracy of image representation. This embodiment divides the image encoding process into four progressive levels, and configures a visual sampling cue for each level, so that the visual segmentation basic model can capture and embed multi-level image features to form a multi-level cue feature embedding, thereby comprehensively representing the depth and breadth of image information. This embodiment uses an image decoder to achieve efficient alignment between multi-layer image features and corresponding cue feature embeddings, ensuring accurate transmission and mapping of image semantic features and cue information. Subsequently, the shallow and deep image features are spliced ​​and fused, and the fused features are sent to the progressive convolutional layer, and the corresponding mask is output with the help of upsampling technology to a suitable size.

[0047] In this embodiment, the hint encoder performs interpolation extraction of image features based on the coordinates of a series of hint points to complete image sampling, and generates corresponding hint feature embedding and hint position embedding. Through interpolation extraction, the consistency of hint feature embedding and image feature embedding in semantic information is ensured. Subsequently, this embodiment uses transformer self-attention and cross-attention to complete the alignment of hint embedding and image embedding. Specifically, the structure diagram of the visual sampling hinter is as follows: Figure 3 As shown in , the workflow includes: first, introducing several global tokens for the hint feature embedding and applying self-attention technology. The global token can learn feature point embedding with high similarity semantic information, which can alleviate the large number of false positives and false negatives caused by unclear edges. Then, the updated hint feature embedding is added to the hint position embedding and the self-attention technology is applied to fuse the position information with the semantic information. The fusion of position information can maintain the model's sensitivity to interactive information. Finally, the updated global query is applied to the cross-attention between the image encoding to complete the alignment of the hint embedding and the image embedding.

[0048] Step S300: adopting an iterative training strategy of simulated interaction, training the training model based on the training set, and evaluating and saving the model with the best evaluation index based on the test set to obtain the image segmentation model.

[0049] In each round of training, each training sample is iterated 5 times. During the first prediction, the training sample is sent to the training model and the position coordinates of a positive sample are randomly generated as prompt information. Then, each iteration will generate random multi-point prompts according to the prediction results of the previous iteration and input them into the training model. The prediction results of each iteration are respectively used to calculate the loss function with the corresponding label samples, and the parameters of the training model are updated through gradient backpropagation. Among them, except for the first input, the image encoder of the training model does not participate in parameter update.

[0050] In the training process of the image segmentation model, this embodiment evaluates and saves the model with the best evaluation index in the validation set. During the training process, EPOCH (indicates the number of updates when all training data in learning have been used once) is used as the basic cycle unit, a total of 100 EPOCHs are trained, BATCH SIZE (batch size) is 4, the optimizer is AdamW, a dynamic learning rate is used, the initial learning rate is 0.0008, the loss function is the weighted sum of the cross entropy loss and the Dice loss, and the evaluation index is the Dice coefficient. The Dice coefficient can be used to measure the segmentation performance, thereby evaluating the segmentation performance of the image segmentation model for medical images.

[0051] During the testing phase of the image segmentation model, this embodiment loads the trained weights, performs five iterative predictions on each sample, saves each prediction result in NI FT I format, and calculates the Di ce index to accurately evaluate the segmentation performance.

[0052] It can be seen that this embodiment reconstructs the SAM algorithm for three-dimensional images to fully adapt to the high-dimensional characteristics and complexity of three-dimensional image data, while ensuring the efficiency and stability of the algorithm when processing three-dimensional data, and can more efficiently use the prompt-driven model to segment three-dimensional data, and can more accurately capture the three-dimensional morphology and position information of organs or lesions in medical images, providing doctors with more intuitive and comprehensive diagnostic basis. This embodiment proposes a novel prompt feature generation method, that is, using visual sampling for interpolation extraction of image features and the fusion of prompt position information to generate prompt feature coding, thereby enhancing the model's ability to understand image content and segmentation accuracy. The strategy of generating prompt embedding with position coding technology achieves excellent target segmentation effects on images with clear boundaries, while medical image segmentation targets often have complex structures and fuzzy boundaries, which directly affects the performance of position coding technology in prompt-driven segmentation of medical images. Visual use of prompts can use the deep semantic features of the image to construct prompt features, and the fusion of position information can maintain the model's sensitivity to interactive information, effectively guiding the model's segmentation process in areas with fuzzy boundaries or complex structures. This prompting method enables the model to ignore imprecise or ambiguous prompt inputs to a certain extent, enhancing the robustness and intelligent adaptability of the algorithm. In order to further mine the multi-level information of image features and improve the accuracy of segmentation results, this embodiment introduces a hierarchical model architecture to implement a multi-level sampling strategy. This architecture promotes the effective integration of cross-level features by combining shallow and deep image features, and also significantly improves the richness and accuracy of image representation, thereby providing more accurate and reliable guidance for the segmentation process.

[0053] Furthermore, this embodiment proposes a method in which the visual sampling prompt strategy is independent of the network model, and this embodiment can be applied to the prompt-driven model for three-dimensional images. In fact, any network used for image segmentation tasks can adopt the prompt method proposed by the present invention. The application scenario proposed in this embodiment is a three-dimensional volume image such as a CT image or MRI, and it can also be easily extended to other three-dimensional images.

[0054] Based on this, the present invention also provides a three-dimensional medical image segmentation method based on a prompt-driven mechanism. The image segmentation method adopts the image segmentation model trained by the image segmentation model training method based on the prompt-driven mechanism described in the above embodiment. Based on the image segmentation model trained in the above embodiment, in the organ segmentation task of three-dimensional medical images, this embodiment uses 5 training data sets for training. Under the condition of one prompt, the DSC score of the present invention reaches 85.08%, which is 6.39% higher than the current prompt segmentation algorithm for medical images and 59.27% ​​higher than the prompt segmentation algorithm in the general field. This embodiment can accurately segment the target organ with a small amount of prompts in the three-dimensional image, which solves the limitation that a large number of manual prompts are required when the current prompt-driven model is applied. The algorithm keeps the error within 5% in the face of abnormal and non-standard prompt input, which proves that the algorithm can ignore inaccurate or ambiguous prompt input to a certain extent, and has strong robustness and intelligent adaptability. In terms of zero-shot capability, the DSC score of the algorithm reaches 68.05%, which shows that the invention overcomes the limitation of poor generalization ability of customized segmentation algorithms to a certain extent, allowing the model to segment image patterns and anatomical structures that are not in the data set to a certain extent. These results indicate that the image segmentation model based on the cue-driven mechanism allows doctors to segment the target area on the 3D image by inputting cue points, thereby improving the accuracy of segmentation.

[0055] Based on the above embodiments, the present invention also provides an image segmentation model training system based on a prompt-driven mechanism, such as Figure 4 As shown in, the system includes: a data preprocessing module 10, a model improvement module 20 and a model training module 30. Specifically, the data preprocessing module is used to obtain a data set and perform preprocessing operations on the labeled samples in the data set, the labeled samples include a training set and a test set, and the data set includes a medical volume image and a volume segmentation image. The model improvement module 20 is used to improve the three-dimensional adaptation and layering strategy of each module in the visual segmentation basic model, to modify the image encoder and the prompt encoder in the visual segmentation basic model, and obtain a training model in which the prompt embedding and the image embedding are aligned. The model training module 30 is used to adopt an iterative training strategy of simulated interaction, train the training model based on the training set, and evaluate and save the model with the best evaluation index based on the test set to obtain the image segmentation model.

[0056] The working principles of each module in the image segmentation model training system based on the prompt-driven mechanism of this embodiment are the same as the principles of each step in the above method embodiment, and will not be repeated here.

[0057] Based on the above embodiment, the present invention further provides a terminal, the principle block diagram of the terminal can be as follows: Figure 5 The terminal may include one or more processors 100 ( Figure 5 Only one is shown), a memory 101 and a computer program 102 stored in the memory 101 and executable on one or more processors 100, for example, an image segmentation model training program based on a prompt-driven mechanism. When one or more processors 100 execute the computer program 102, the various steps in the embodiment of the image segmentation model training method based on the prompt-driven mechanism can be implemented. Alternatively, when one or more processors 100 execute the computer program 102, the functions of the various modules / units in the embodiment of the image segmentation model training method based on the prompt-driven mechanism can be implemented, which is not limited here.

[0058] In one embodiment, the processor 100 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0059] In one embodiment, the memory 101 may be an internal storage unit of an electronic device, such as a hard disk or memory of the electronic device. The memory 101 may also be an external storage device of the electronic device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device. Further, the memory 101 may also include both an internal storage unit of the electronic device and an external storage device. The memory 101 is used to store computer programs and other programs and data required by the terminal. The memory 101 may also be used to temporarily store data that has been output or is to be output.

[0060] Those skilled in the art will understand that Figure 5The principle block diagram shown in the figure is only a block diagram of a partial structure related to the scheme of the present invention, and does not constitute a limitation on the terminal to which the scheme of the present invention is applied. The specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0061] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, operating database or other media used in the embodiments provided by the present invention may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training an image segmentation model based on a prompt-driven mechanism, characterized in that: The method comprises: Acquire a data set, and perform a preprocessing operation on labeled samples in the data set, wherein the labeled samples include a training set and a test set, and the data set includes a medical volume image and a volume segmentation image; Improve the three-dimensional adaptation and layering strategy of each module in the visual segmentation basic model, modify the image encoder and the prompt encoder in the visual segmentation basic model, and obtain a training model in which the prompt embedding and the image embedding are aligned; Adopting a simulated interactive iterative training strategy, training the training model based on the training set, and evaluating and saving the model with the best evaluation index based on the test set to obtain the image segmentation model; The improvement of the three-dimensional adaptation and layering strategy of each module in the visual segmentation basic model includes: Converting a two-dimensional image encoder in the visual segmentation basic model into a three-dimensional encoder for three-dimensional data; A hierarchical strategy is introduced to extract image features at different depths; Adjust the cue encoder structure so that the visual segmentation base model can utilize the semantic features corresponding to the cue; According to the modifications of the image encoder and the hint encoder, the image decoder is modified to complete the fusion of the hint position information and the semantic information, realize the alignment of the hint embedding and the image embedding and output the corresponding mask; The hierarchical strategy is introduced to extract image features at different depths, including: Dividing the image encoding process into four progressive levels, and configuring a visual sampling cue for each level, so that the visual segmentation base model can capture and embed multi-level image features to form a multi-level cue feature embedding; The workflow of the visual sampling prompter includes: Introduce several global tokens for cue feature embedding and apply self-attention technology; After adding the updated cue feature embedding to the cue position embedding, self-attention technology is applied to fuse the position information with the semantic information. The updated global query is applied to the cross-attention between the image encoding to align the hint embedding and the image embedding.

2. The image segmentation model training method based on the prompt-driven mechanism according to claim 1, characterized in that: The preprocessing operation includes segmentation label, resampling and grayscale normalization; wherein, The segmentation label includes labeling the volume segmentation image, separating the segmentation label of each organ and keeping it aligned with the medical volume image; The resampling includes adjusting the voxel spacing of the image to the median size of the voxel spacing of the data set using nearest neighbor interpolation, and resampling the medical volume image to a volume image of a preset scale with its corresponding label segmentation as the center; The grayscale normalization includes normalizing the medical volume image to [0, 1] using a minimum-maximum normalization method, and adjusting the pixel intensity to [-1, 1] by a linear transformation.

3. The image segmentation model training method based on the prompt-driven mechanism according to claim 1, characterized in that: The strategy of iterative training using simulated interaction, training the training model based on the training set, includes: In each round of training, each training sample is iterated 5 times. During the first prediction, the training sample is fed into the training model and the position coordinates of a positive sample are randomly generated as prompt information. After that, each iteration will generate random multi-point prompts based on the prediction results of the previous iteration and input them into the training model. The prediction results of each iteration are respectively used to calculate the loss function with the corresponding label samples, and the parameters of the training model are updated through gradient backpropagation. Among them, except for the first input, the image encoder of the training model does not participate in parameter updating.

4. A three-dimensional medical image segmentation method based on a prompt-driven mechanism, characterized in that: The three-dimensional medical image segmentation method adopts the image segmentation model trained by the image segmentation model training method based on the prompt-driven mechanism described in any one of claims 1 to 3.

5. A prompt-driven image segmentation model training system, characterized in that: The system is used to implement the steps of the image segmentation model training method based on the prompt-driven mechanism according to any one of claims 1 to 3, and the system comprises: A data preprocessing module, used to obtain a data set and perform preprocessing operations on label samples in the data set, wherein the label samples include a training set and a test set, and the data set includes a medical volume image and a volume segmentation image; A model improvement module, used to improve the three-dimensional adaptation and layering strategy of each module in the visual segmentation basic model, to modify the image encoder and the prompt encoder in the visual segmentation basic model, and to obtain a training model in which the prompt embedding and the image embedding are aligned; The model training module is used to adopt an iterative training strategy of simulated interaction, train the training model based on the training set, and evaluate and save the model with the best evaluation index based on the test set to obtain the image segmentation model.

6. A terminal, characterized in that: The terminal includes a memory, a processor, and an image segmentation model training program based on a prompt-driven mechanism or a three-dimensional medical image segmentation program based on a prompt-driven mechanism stored in the memory and executable on the processor. When the processor executes the image segmentation model training program based on the prompt-driven mechanism, the processor implements the steps of the image segmentation model training method based on the prompt-driven mechanism as described in any one of claims 1 to 3. When the processor executes the three-dimensional medical image segmentation program based on the prompt-driven mechanism, the processor implements the steps of the three-dimensional medical image segmentation method based on the prompt-driven mechanism as described in claim 4.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an image segmentation model training program based on a prompt-driven mechanism or a three-dimensional medical image segmentation program based on a prompt-driven mechanism. When the image segmentation model training program based on the prompt-driven mechanism is executed by the processor, the steps of the image segmentation model training method based on the prompt-driven mechanism as described in any one of claims 1 to 3 are implemented. When the three-dimensional medical image segmentation program based on the prompt-driven mechanism is executed by the processor, the steps of the three-dimensional medical image segmentation method based on the prompt-driven mechanism as described in claim 4 are implemented.

Images