Medical image multi-target three-dimensional segmentation method and system

Abstract

The invention discloses a medical image multi-target three-dimensional segmentation method and system, and relates to the technical field of computer platform load balancing, and the method comprises the steps: collecting medical three-dimensional image data, carrying out the preprocessing of the medical three-dimensional image data, and dividing the data into a training set, a verification set and a test set; constructing a deep neural network model through the training set, and optimizing model parameters; and based on the optimized deep neural network model, outputting an optimal segmentation model and an optimal weight through the verification set, and performing performance evaluation through the test set. According to the method, the deep neural network model is constructed, the model parameters are optimized, the model training and reasoning speed is increased, the model robustness is improved, and medical images with different complexities can be effectively dealt with; through a continuous verification evaluation and model weight storage mechanism, the robustness of a final output model in practical application is ensured; through strict evaluation of the test set, the finally output optimal model and weight can provide a more accurate multi-target three-dimensional segmentation result in clinical application.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and system for multi-target three-dimensional segmentation of medical images. Background Art

[0002] Multi-target 3D segmentation of medical images is of great significance in the field of medical image processing. It can provide doctors with accurate anatomical information, which helps in disease diagnosis, treatment planning and postoperative evaluation. For example, the abdominal area contains multiple key organs, such as the liver, spleen, kidneys and pancreas. The morphology and position changes of these organs may reflect different pathological conditions. Therefore, accurate organ segmentation is a key task in medical image analysis.

[0003] Although existing image segmentation methods have made significant progress in abdominal medical image processing, they still face huge challenges when dealing with abdominal multi-organ segmentation involving various complex lesions (such as cirrhosis, pancreatic cancer, and kidney tumors). First, the contrast difference between organs in abdominal images is small. Especially when there are lesions, the grayscale distribution of the lesion area is similar to that of normal tissue, which can easily lead to incorrect segmentation. Secondly, some lesions may cause significant changes in organ morphology. For example, the surface of the liver becomes uneven in cirrhosis, and pancreatic cancer may cause blurred pancreatic boundaries, which increases the difficulty of segmentation. In addition, the mutual compression and overlap between abdominal organs also put forward higher requirements for accurate segmentation. Therefore, developing a multi-target three-dimensional method that can accurately segment medical images is of great significance to improving the accuracy of clinical diagnosis and treatment effects. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: the existing multi-target three-dimensional segmentation method of medical images has low segmentation accuracy, low segmentation efficiency and poor robustness, and how to realize a multi-target three-dimensional method that can accurately segment medical images to improve the accuracy of clinical diagnosis and treatment effect.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a method for multi-target three-dimensional segmentation of medical images, comprising collecting medical three-dimensional image data, preprocessing the medical three-dimensional image data and dividing it into a training set, a validation set and a test set; constructing a deep neural network model through the training set and optimizing the model parameters; based on the optimized deep neural network model, outputting the optimal segmentation model and the optimal weight through the validation set, and performing performance evaluation through the test set.

[0007] As a preferred solution of the medical image multi-target three-dimensional segmentation method described in the present invention, the preprocessing includes cropping the medical three-dimensional image data based on the acquisition of medical three-dimensional image data, removing irrelevant areas of the image, adjusting the image grayscale range through intensity normalization, and using rotation, flipping, scaling and noise addition techniques to perform data enhancement processing on the medical three-dimensional image data to expand the data set.

[0008] As a preferred solution of the medical image multi-target three-dimensional segmentation method described in the present invention, the division into training set, validation set and test set includes dividing the preprocessed medical three-dimensional image data samples into training set, validation set and test set according to the ratio of 2:1:2 and 341:20:90 respectively, the training set is used to train the model, the model is learned through the training set, the validation set is used to adjust the hyperparameters of the model and select the optimal configuration of the model, and the test set is used to finally evaluate the model performance.

[0009] As a preferred solution of the medical image multi-target three-dimensional segmentation method described in the present invention, the construction of the deep neural network model includes using a convolutional neural network with an encoder-decoder architecture as the deep neural network model, and adopting a convolution-Mamba-convolution ladder convolution structure as the backbone, importing the training set data into the deep neural network model, and calculating the first-order moment estimation through the AdamW algorithm, which is expressed as:

[0010] m t =β1m t-1 +(1-β1)g t

[0011] Among them, m t is the first-order moment estimate of the t-th iteration, β1 is the first-order momentum parameter, g t is the gradient; calculate the second-order moment estimate, expressed as:

[0012]

[0013] Among them, v t is the second-order moment estimate of the t-th iteration, β2 is the second-order momentum parameter; calculate the first-order moment estimate correction, expressed as:

[0014]

[0015] in, is the bias correction of the first-order moment estimate of the t-th iteration; calculate the second-order moment estimate correction, expressed as:

[0016]

[0017] in, is the bias correction of the second-order moment estimate of the t-th iteration; the parameter update is calculated and expressed as:

[0018]

[0019] Among them, θ t+1 is the updated parameter, θ t is the parameter before update, η is the learning rate, ε is a constant, λ is the weight decay factor, ηλθ t is the weight decay term.

[0020] As a preferred solution of the multi-target three-dimensional segmentation method of medical images described in the present invention, the optimization model parameters include optimizing the deep neural network model training by combining the multi-classification cross entropy loss function and the Dice loss function, and calculating the total loss function, which is expressed as:

[0021] L seg =αL dice +yL ce

[0022] Among them, L seg is the total loss function, L dice is the Dice loss function, L ce is the multi-classification cross entropy loss function, α is the weight of the Dice loss function in the total loss function, and γ is the weight of the multi-classification cross entropy loss function in the total loss function; calculate the Dice loss function coefficient, expressed as:

[0023]

[0024] Among them, Dice is the coefficient of the Dice loss function, X is the probability map of the true label, Y is the probability map predicted by the model, |X∩Y| is the overlapped part of the probability map of the true label and the probability map predicted by the model; the Dice loss function is calculated as:

[0025] L dice =1-Dice

[0026] Calculate the multi-classification cross entropy loss function, expressed as:

[0027]

[0028] Among them, H is the cross entropy, n is the number of categories, x i is the i-th type of sample, p(x i ) is the true distribution of the i-th class of samples, q(x i ) is the probability that the sample is predicted to be of the i-th category.

[0029] As a preferred solution of the multi-target three-dimensional segmentation method of medical images described in the present invention, the optimal segmentation model and the optimal weight include adjusting the hyperparameters of the optimized deep neural network model through the validation set. During the training process, after each set number of iterations, the model will be evaluated on the validation set, and the average Dice loss function coefficient on the validation set will be calculated. The validation result will be compared with the current optimal model. If the validation result is better than the previous optimal result, the current model weight will be saved. After the entire training is completed, the optimal model weight obtained during the verification process will be output.

[0030] As a preferred solution of the medical image multi-target three-dimensional segmentation method of the present invention, the performance evaluation includes testing the optimal weight output by training on the test set data, using the evaluation index Dice loss function coefficient to evaluate the performance of the optimal model weight, and calculating the final output index value, which is expressed as:

[0031]

[0032] Where Dice' is the coefficient of the Dice loss function of the optimal model weight, X' is the probability map of the true label of the optimal model weight, Y' is the probability map of the model prediction of the optimal model weight, and |X'∩Y'| is the overlapped part of the probability map of the true label of the optimal model weight and the probability map of the model prediction of the optimal model weight.

[0033] Another object of the present invention is to provide a medical image multi-target three-dimensional segmentation system, which can construct a deep neural network model through a training set and optimize model parameters, thereby solving the problem of low segmentation efficiency in the current medical image multi-target three-dimensional segmentation technology.

[0034] As a preferred solution of the medical image multi-target three-dimensional segmentation system described in the present invention, it includes: a data partitioning module, a training optimization module, and a performance evaluation module; the data partitioning module is used to collect medical three-dimensional image data, pre-process the medical three-dimensional image data and divide it into a training set, a validation set and a test set; the training optimization module is used to construct a deep neural network model and optimize model parameters through the training set; the performance evaluation module is used to output the optimal segmentation model and optimal weights through the validation set based on the optimized deep neural network model, and perform performance evaluation through the test set.

[0035] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a step of a multi-objective three-dimensional segmentation method for medical images.

[0036] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for multi-objective three-dimensional segmentation of medical images.

[0037] The beneficial effects of the present invention are as follows: the medical image multi-target three-dimensional segmentation method provided by the present invention eliminates interference from irrelevant areas, enhances data diversity, and improves the segmentation accuracy of the model, especially the fine-grained segmentation of complex organs by cropping, intensity normalizing and data enhancement of medical three-dimensional image data; the training process is efficient by a reasonable data set division method; the computational efficiency is improved by constructing a deep neural network model, using an encoder-decoder architecture and a convolution-Mamba-convolution backbone network structure, the model training and reasoning speed are accelerated, and the segmentation time is shortened; the robustness of the model is optimized by combining a multi-classification cross entropy loss and a Dice loss function, and the model can effectively cope with medical images of different complexities and maintain the good performance of the model on the test set; the robustness of the final output model in practical applications is ensured by a continuous verification evaluation and model weight preservation mechanism; the optimal model and weights finally output by strict evaluation of the test set can provide more accurate multi-target three-dimensional segmentation results in clinical applications, which is helpful to improve the accuracy of clinical diagnosis and the effect of treatment plans. The present invention achieves better results in segmentation accuracy, segmentation efficiency, robustness and clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0039] Figure 1 An overall flow chart of a method for multi-objective three-dimensional segmentation of medical images provided by the first embodiment of the present invention;

[0040] Figure 2 A schematic diagram of modules of a medical image multi-object 3D segmentation system provided in the third embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0042] Example 1, reference Figure 1 , as an embodiment of the present invention, provides a method for multi-objective three-dimensional segmentation of medical images, comprising:

[0043] S1: Collect medical 3D image data, preprocess the medical 3D image data and divide it into a training set, a validation set and a test set.

[0044] Furthermore, preprocessing includes cropping the medical 3D image data based on the acquisition of the medical 3D image data, removing irrelevant areas of the image, adjusting the image grayscale range through intensity normalization, and using rotation, flipping, scaling and noise addition techniques to perform data enhancement processing on the medical 3D image data to expand the data set.

[0045] It should be noted that the division into training set, validation set and test set includes dividing the preprocessed medical three-dimensional image data samples into training set, validation set and test set according to the ratio of 2:1:2 and 341:20:90 respectively, the training set is used to train the model, the model is learned through the training set, the validation set is used to adjust the hyperparameters of the model and select the optimal configuration of the model, and the test set is used to finally evaluate the model performance.

[0046] It should also be noted that the collected medical three-dimensional image data is CT image data. The acquired CT image data is cropped to remove irrelevant areas, improve processing efficiency, perform intensity normalization, ensure the consistency of image data, help model training, implement data enhancement, expand the data set, and improve the generalization ability and robustness of the model. The training set is used to train the model. The model learns through the training set and adjusts its internal parameters to minimize the training error so that the model can better fit the training data. The validation set is used to adjust the model's hyperparameters and select the best configuration of the model. During the training process, the model will be regularly evaluated on the validation set to monitor whether overfitting occurs. The purpose is to help find the optimal model parameters and architecture during the training process. The test set is used to finally evaluate the performance of the model. The purpose is to objectively measure the generalization ability of the model.

[0047] S2: Build a deep neural network model through the training set and optimize the model parameters.

[0048] Furthermore, building a deep neural network model includes using a convolutional neural network with an encoder-decoder architecture as a deep neural network model, and using a convolution-Mamba-convolution staircase convolution structure as the backbone, importing the training set data into the deep neural network model, and calculating the first-order moment estimate through the AdamW algorithm, which is expressed as:

[0049] m t =β1m t-1 +(1-β1)gt

[0050] Among them, m t is the first-order moment estimate of the t-th iteration, β1 is the first-order momentum parameter, g t is the gradient; calculate the second-order moment estimate, expressed as:

[0051]

[0052] Among them, v t is the second-order moment estimate of the t-th iteration, β2 is the second-order momentum parameter; calculate the first-order moment estimate correction, expressed as:

[0053]

[0054] in, is the bias correction of the first-order moment estimate of the t-th iteration; calculate the second-order moment estimate correction, expressed as:

[0055]

[0056] in, is the bias correction of the second-order moment estimate of the t-th iteration; the parameter update is calculated and expressed as:

[0057]

[0058] Among them, θ t+1 is the updated parameter, θ t is the parameter before update, η is the learning rate, ε is a constant, λ is the weight decay factor, ηλθ t is the weight decay term.

[0059] It should be noted that the optimization model parameters include optimizing the deep neural network model training by combining the multi-classification cross entropy loss function and the Dice loss function, and calculating the total loss function, which is expressed as:

[0060] L seg =αL dice +γL ce

[0061] Among them, L seg is the total loss function, L dice is the Dice loss function, L ce is the multi-classification cross entropy loss function, α is the weight of the Dice loss function in the total loss function, and γ is the weight of the multi-classification cross entropy loss function in the total loss function; calculate the Dice loss function coefficient, expressed as:

[0062]

[0063] Among them, Dice is the coefficient of the Dice loss function, X is the probability map of the true label, Y is the probability map predicted by the model, and |X∩Y| is the overlapped part of the probability map of the true label and the probability map predicted by the model; the Dice loss function is calculated as:

[0064] L dice =1-Dice

[0065] Calculate the multi-classification cross entropy loss function, expressed as:

[0066]

[0067] Among them, H is the cross entropy, n is the number of categories, x i is the i-th type of sample, p(x i ) is the true distribution of the i-th class of samples, q(x i ) is the probability that the sample is predicted to be of the i-th category.

[0068] It should also be noted that the model architecture of the deep neural network model is a convolutional neural network with an encoder-decoder architecture. The encoder part consists of a MambaUX hybrid visual coding strategy module and a downsampling layer. The convolution part uses a large convolution kernel (7*7*7), combining the Mamba architecture with the convolution architecture, and adopts a "convolution-Mamba-convolution" step convolution structure as the backbone. It consists of a residual convolution block, a MambaUX hybrid visual coding strategy module and a downsampling layer, combining the advantages of a large convolution architecture in local feature extraction and large receptive field capture, and the Mamba architecture's efficient recognition ability in long-distance dependencies. In addition, at each stage of the decoder, it receives fused features composed of low-level fine-grained features and high-level structural features, which can better Multi-scale feature information is obtained to achieve accurate segmentation of multiple abdominal organs. The training set data is imported into the deep neural network model. The deep neural network model is optimized multiple times using a combination of the AdamW parameter optimization algorithm, the multi-classification cross entropy loss function and the Dice function. AdamW is an improved Adam optimization algorithm. The generalization ability of the model is improved by introducing weight attenuation on the basis of the Adam algorithm. The use of the multi-classification cross entropy loss function and the Dice loss function allows the model to capture local and global features at the same time. The Dice loss function focuses on the matching degree of the overall segmented area, while the multi-classification cross entropy loss function focuses more on the classification results of local pixels, which further helps the model make more accurate predictions at the category boundaries. The two can complement each other to a certain extent.

[0069] S3: Based on the optimized deep neural network model, the optimal segmentation model and optimal weights are output through the validation set, and the performance is evaluated through the test set.

[0070] Furthermore, the optimal segmentation model and the optimal weights include adjusting the hyperparameters of the optimized deep neural network model through the validation set. During the training process, the model will be evaluated on the validation set after each set number of iterations, and the average Dice loss function coefficient on the validation set will be calculated. The validation result will be compared with the current optimal model. If the validation result is better than the previous optimal result, the current model weights will be saved. After the entire training is completed, the optimal model weights obtained during the verification process will be output.

[0071] It should be noted that the performance evaluation includes testing the optimal weights output by training on the test set data, using the evaluation indicator Dice loss function coefficient to evaluate the performance of the optimal model weights, and calculating the final output indicator value, which is expressed as:

[0072]

[0073] Where Dice' is the coefficient of the Dice loss function of the optimal model weight, X' is the probability map of the true label of the optimal model weight, Y' is the probability map of the model prediction of the optimal model weight, and |X'∩Y'| is the overlapped part of the probability map of the true label of the optimal model weight and the probability map of the model prediction of the optimal model weight.

[0074] It should also be noted that the present invention adopts an iterative training method. After each set number of iterations, the model will be evaluated on the validation set, and the average Dice loss function coefficient on the validation set will be calculated. The advantage of this method is that it can timely detect the overfitting or underfitting of the model and improve the model by adjusting the hyperparameters or model structure. For example, if the Dice loss function coefficient of the model on the validation set is high, it means that the generalization ability of the model is poor, and it is necessary to adjust the model structure or increase the training data to improve the robustness of the model. In addition, by comparing the Dice loss function coefficient on the validation set, the model weight with the best performance is saved, avoiding the training The tedious steps of evaluation after completion improves efficiency. Finally, by evaluating the optimal model weights through the test set, the generalization ability of the model can be evaluated more objectively to ensure the effectiveness of the model in practical applications. The Dice loss function coefficient is used as a performance evaluation indicator. The closer its value is to 1, the better the segmentation effect. The advantage of using the Dice loss function coefficient as a performance evaluation indicator is that it can intuitively reflect the segmentation accuracy of the segmentation model and has good robustness to the imbalance of the segmentation results. For example, when segmenting large organs such as the liver and kidneys, the Dice loss function coefficient can accurately reflect the accuracy of the segmentation results.

[0075] Example 2 is an embodiment of the present invention, which provides a multi-target three-dimensional segmentation method for medical images. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0076] First, the model was trained under the Pytorch 2.0.0, CUDA 11.7, and Monai 0.9.0 frameworks. The AdamW optimizer was used in the training process of all networks, and the initial learning rate was set to 0.0001. The training was performed on the NVIDIA RTX A4000 graphics card (16G video memory) based on the NVIDIA Ampere architecture. 68,400 steps of training (400 cycles) were performed for the FLARE2021 dataset. Since FLARE2021 is a competition dataset, the label of the test dataset cannot be obtained. Therefore, the test experiment was performed locally, and the prediction results were uploaded to the MICCAI Grand Challenge official website for indicator evaluation.

[0077] The FLARE2021 dataset is adapted from the abdominal CT dataset AbdomenCT-1K proposed by Ma et al., integrating 50 abdominal CT scans from MSD liver, MSD spleen, MSD pancreas, NIH pancreas, KiTs datasets and Nanjing University, collecting a total of 451 CT scans, each with voxel-level annotations of 4 abdominal organs, including liver, kidney, spleen and pancreas. In the experiment, the 451 CT scans were divided into 341 training cases, 20 validation cases and 90 test cases.

[0078] Intensity clipping was applied in the preprocessing stage to further enhance the contrast of the soft combination (minimum: -175, maximum: 250). Secondly, the intensity of each volume was normalized after clipping and the minimum-maximum normalization (N-N1) / (N 99 -N1) to normalize the intensity value to between 0 and 1, where N P Represents the Pth percentile intensity of the intensity in N. Then, a subvolume of size 96*96*96 is randomly cropped from the foreground and data augmentation is performed, including rotation, intensity shift and scaling (scaling factor: 0.1).

[0079] During the training phase, the training set data is imported into the deep neural network model, and the combination of AdamW parameter optimization algorithm, multi-classification cross entropy loss function and Dice loss function is selected to optimize the deep neural network model multiple times. At the same time, the validation set is used to adjust the hyperparameters of the deep neural network model after multiple optimizations, and finally the optimal segmentation model and the optimal weight are obtained.

[0080] In the final testing phase, test experiments are conducted locally and the prediction results are uploaded to the MICCAI GrandChallenge official website for indicator evaluation.

[0081] The present invention conducts comparative experiments with six mainstream segmentation methods, including a CNN-based method (3DUX-Net), a Transformer-based method (UNETR), three CNN-Transformer hybrid-based methods (SwinUNETR, SwinUNETRv2, and nnFormer), and a Mamba-based method (SegMamba). To ensure a fair comparison, the public implementations of these methods are retrained under the same experimental conditions to generate the best segmentation results.

[0082] Refer to Table 1 for a comparative analysis of the experimental data.

[0083] Table 1 Experimental data comparison table

[0084]

[0085] The quantitative segmentation results of the FLARE2021 dataset are shown in Table 1, which demonstrates the segmentation performance achieved on different organs. In the segmentation of the liver and pancreas, the model of the present invention performed particularly well, reaching Dice similarity coefficient indices of 0.96±0.04 and 0.71±0.18, respectively. As larger organs, the liver and kidneys have clear structures and obvious boundaries, which makes them important objects in the segmentation task. To this end, the model of the present invention adopts a hybrid design combining convolutional neural networks and Mamba architectures. This design can not only capture long-distance dependencies in the image, but also effectively extract local features, thereby accurately segmenting the shape and boundaries of the liver and kidneys. In addition, in the spleen segmentation task, although 3DUXNet showed certain advantages, the model of the present invention still achieved relatively close excellent performance. The appearance and position of the spleen have significant variability between different patients, which adds challenges to the segmentation task. However, the model of the present invention still shows strong adaptability when dealing with changes in the shape and position of the spleen, thanks to its structural design's ability to capture local details. The pancreas is a smaller and more shaped organ. For organs with complex shapes, traditional segmentation methods often find it difficult to achieve satisfactory results. However, the model of the present invention achieves accurate segmentation of the pancreas by introducing a hybrid visual coding strategy module (MambaUX). The MambaUX module adopts a two-stage design, combining the long-distance dependency capture capability of the Mamba architecture and the local feature extraction advantages of the convolutional neural network, which significantly improves the segmentation effect of complex morphological organs such as the pancreas, especially when dealing with small organs with complex morphology and blurred boundaries. This hybrid strategy exhibits excellent performance. In addition, in order to further improve the fine-grained accuracy of the segmentation, skip connections and a dual-level feature fusion module (DLFF) are introduced into the decoder. This design integrates information at different scales, so that the model can still accurately segment the tiny details of the organs when facing complex anatomical structures. In summary, the performance of the present invention on the FLARE2021 dataset shows its comprehensive advantages in the segmentation of large and small organs. Whether it is for large organs with regular morphology or for small organs with complex shapes, the model of the present invention exhibits excellent segmentation capabilities, which are significantly better than other existing methods. Therefore, our invention is creative.

[0086] Example 3, reference Figure 2 , as an embodiment of the present invention, provides a medical image multi-objective three-dimensional segmentation system, including a data partitioning module, a training optimization module, and a performance evaluation module.

[0087] The data division module is used to collect medical 3D image data, pre-process the medical 3D image data and divide it into training set, validation set and test set; the training optimization module is used to build a deep neural network model and optimize model parameters through the training set; the performance evaluation module is used to output the optimal segmentation model and optimal weights through the validation set based on the optimized deep neural network model, and perform performance evaluation through the test set.

[0088] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0089] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0090] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0091] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limited. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

[0092] 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 preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for multi-object 3D segmentation of medical images, characterized in that: include: Collect medical three-dimensional image data, pre-process the medical three-dimensional image data and divide it into a training set, a validation set and a test set; Through the training set, a deep neural network model is constructed and the model parameters are optimized; Based on the optimized deep neural network model, the optimal segmentation model and optimal weights are output through the validation set, and the performance is evaluated through the test set.

2. The method for multi-object 3D segmentation of medical images according to claim 1, characterized in that: The preprocessing includes cropping the medical three-dimensional image data based on the acquisition of medical three-dimensional image data, removing irrelevant areas of the image, adjusting the image grayscale range through intensity normalization, and using rotation, flipping, scaling and noise addition technology to perform data enhancement processing on the medical three-dimensional image data to expand the data set.

3. The method for multi-object 3D segmentation of medical images according to claim 2, characterized in that: The division into training set, validation set and test set includes dividing the preprocessed medical three-dimensional image data samples into training set, validation set and test set according to the ratio of 2:1:2 and 341:20:90 respectively, the training set is used to train the model, the model is learned through the training set, the validation set is used to adjust the hyperparameters of the model and select the optimal configuration of the model, and the test set is used to finally evaluate the model performance.

4. The method for multi-object 3D segmentation of medical images as claimed in claim 3, characterized in that: The construction of the deep neural network model includes using a convolutional neural network with an encoder-decoder architecture as the deep neural network model, and adopting a convolution-Mamba-convolution ladder convolution structure as the backbone, importing the training set data into the deep neural network model, and calculating the first-order moment estimation through the AdamW algorithm, which is expressed as: m t =β1m t-1 +(1-β1)g t Among them, m t is the first-order moment estimate of the t-th iteration, β1 is the first-order momentum parameter, g t is the gradient; Compute the second moment estimate, expressed as: Among them, v t is the second-order moment estimate of the t-th iteration, β2 is the second-order momentum parameter; Calculate the first-order moment estimate correction, expressed as: in, is the bias correction of the first-order moment estimate at the t-th iteration; Calculate the second-order moment estimate correction, expressed as: in, is the bias correction of the second-order moment estimate at the tth iteration; Compute parameter updates, expressed as: Among them, θ t+1 is the updated parameter, θ t is the parameter before update, η is the learning rate, ε is a constant, λ is the weight decay factor, ηλθ t is the weight decay term.

5. The method for multi-object 3D segmentation of medical images according to claim 4, characterized in that: The optimization model parameters include optimizing the deep neural network model training by combining the multi-classification cross entropy loss function and the Dice loss function, and calculating the total loss function, which is expressed as: L seg =λL dice +γL ce Among them, L seg is the total loss function, L dice is the Dice loss function, L ce is the multi-classification cross entropy loss function, α is the weight of the Dice loss function in the total loss function, and γ is the weight of the multi-classification cross entropy loss function in the total loss function; Calculate the Dice loss function coefficient, expressed as: Where Dice is the coefficient of the Dice loss function, X is the probability map of the true label, Y is the probability map predicted by the model, and |X∩Y| is the overlap between the probability map of the true label and the probability map predicted by the model; Calculate the Dice loss function, expressed as: THE dice =1-Dice Calculate the multi-classification cross entropy loss function, expressed as: Among them, H is the cross entropy, n is the number of categories, x i is the i-th type of sample, p(x i ) is the true distribution of the i-th class of samples, q(x i ) is the probability that the sample is predicted to be of the i-th category.

6. The method for multi-object 3D segmentation of medical images according to claim 5, characterized in that: The optimal segmentation model and optimal weights include adjusting the hyperparameters of the optimized deep neural network model through the validation set. During the training process, after each set number of iterations, the model will be evaluated on the validation set, and the average Dice loss function coefficient on the validation set will be calculated. The validation result will be compared with the current optimal model. If the validation result is better than the previous optimal result, the current model weight will be saved. After the entire training is completed, the optimal model weight obtained during the verification process will be output.

7. The method for multi-object 3D segmentation of medical images according to claim 6, characterized in that: The performance evaluation includes testing the optimal weights output by training on the test set data, using the evaluation indicator Dice loss function coefficient to evaluate the performance of the optimal model weights, and calculating the final output indicator value, which is expressed as: Where Dice' is the coefficient of the Dice loss function of the optimal model weight, X' is the probability map of the true label of the optimal model weight, Y' is the probability map of the model prediction of the optimal model weight, and |X'∩Y'| is the overlapped part of the probability map of the true label of the optimal model weight and the probability map of the model prediction of the optimal model weight.

8. A system using the medical image multi-object 3D segmentation method according to any one of claims 1 to 7, characterized in that: Including data partition module, training optimization module, and performance evaluation module; The data division module is used to collect medical three-dimensional image data, pre-process the medical three-dimensional image data and divide it into a training set, a verification set and a test set; The training optimization module is used to construct a deep neural network model and optimize model parameters through a training set; The performance evaluation module is used to output the optimal segmentation model and the optimal weight through the validation set based on the optimized deep neural network model, and to perform performance evaluation through the test set.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the medical image multi-object three-dimensional segmentation method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the medical image multi-object three-dimensional segmentation method described in any one of claims 1 to 7 are implemented.

Images