A three-dimensional medical image segmentation method based on timing information enhancement

Abstract

The application discloses a three-dimensional medical image segmentation method based on time sequence information enhancement, and belongs to the medical image segmentation field.The method proposes a Circle Transformer module for extraction and fusion of time sequence information, uses time sequence input to improve the training effect of a deep learning model, and can effectively eliminate the interference of similar features and fuzzy images.In the training stage, an input is a time sequence sample, the training effect of a time sequence information extraction and enhancement model is improved, and the model is no longer dependent on time sequence by simultaneously constraining time sequence information and combining segmentation results before and after the model.Compared with a training method of inputting a single sample, the precision of a segmentation model based on a coding and decoding structure can be improved without cost.In the application stage, only a single three-dimensional image needs to be input, and a sequence does not need to be used as the input, so that the application mode is more flexible.

Description

Technical Field

[0001] This invention belongs to the field of medical image segmentation, and more specifically, relates to a three-dimensional medical image segmentation method based on temporal information enhancement. Background Technology

[0002] Automatic image segmentation technology has wide applications in the medical field and plays a crucial role in the development of computer-aided diagnostic and treatment systems such as clinical diagnosis and surgical navigation. Taking biopsy as an example, it is the gold standard for diagnosing prostate cancer. Accurate biopsy navigation technology can effectively improve the detection rate of biopsy procedures and reduce trauma to patients. Real-time 3D ultrasound has attracted much attention in biopsy navigation; automatic segmentation of the biopsy needle in 3D ultrasound images is a key technology for achieving intraoperative navigation during biopsy procedures.

[0003] Currently, many deep learning models have been proposed for image segmentation tasks, among which Convolutional Neural Networks (CNNs) and Transformers have become mainstream methods for image segmentation. In recent years, the combination of CNNs and Transformers has further improved the accuracy of segmentation networks, but Transformer modules require a large amount of computational and memory resources. Therefore, the combination of CNNs and Transformers mostly relies on encoder-decoder structures to eliminate redundant information through encoding and reduce the complexity of the model.

[0004] With the development of image processing technology, many researchers have found that utilizing temporal information is one of the main directions for improving the accuracy of medical image segmentation. Taking biopsy needles as an example, as they move, temporal information can provide a reference for relative motion and shape, thus greatly reducing the difficulty of detection. Due to the global feature correlation learning capability of the Transformer, it is very suitable for processing multi-frame images and has been widely used in video segmentation tasks. However, existing research mainly focuses on two-dimensional natural image sequences and cannot achieve good results for real-time needle detection based on three-dimensional ultrasound.

[0005] High data annotation costs are another major problem facing medical image segmentation. Therefore, many researchers have focused on exploring semi-supervised segmentation methods, which annotate only a small number of images in the dataset to achieve segmentation accuracy close to that of annotated data. Currently popular semi-supervised segmentation strategies include consistency learning, which encourages the model to produce similar outputs when samples or model parameters are slightly disturbed. This forces similar samples to have more similar output features, while output features of different categories differ more, thus indirectly improving model performance by utilizing unlabeled samples. Existing consistency learning schemes mainly achieve this by building parallel networks or transforming the input data. The former requires additional computational resources, while the latter's performance is not ideal. Summary of the Invention

[0006] To address the aforementioned shortcomings or improvement needs of existing technologies, this invention provides a 3D medical image segmentation method based on temporal information enhancement. This method employs a Circle Transformer to extract motion information of targets in a 3D image sequence for training, which improves the segmentation accuracy of target regions in 3D medical images. This method is applicable to all segmentation models based on encoder-decoder structures. By simultaneously constraining temporal information and combining the segmentation results before and after, it eliminates the model's dependence on the temporal module, improving segmentation accuracy at no cost. During application, only a single frame of 3D image is required as input, eliminating the need for a sequence as input, making the application mode more flexible. For unlabeled data, this method calculates consistency loss based on temporal information combined with the output probability maps before and after, further improving model performance without requiring additional memory.

[0007] To achieve the above objectives, according to a first aspect of the present invention, a three-dimensional medical image segmentation method based on temporal information enhancement is provided, comprising:

[0008] Training phase:

[0009] Semi-supervised training of the segmentation model was performed using the training set and the Circle Transformer module.

[0010] The training set includes three-dimensional medical image sequences containing target regions, wherein the gold standard for target region segmentation of some of the three-dimensional medical image sequences is known and is labeled data, and the other part is unlabeled data;

[0011] The segmentation model includes an encoder and a decoder; wherein, the encoder is used to encode each image in the three-dimensional medical image sequence to obtain encoded features; the Circle Transformer module is used to process each image in the three-dimensional medical image sequence as a target image to obtain its pre-decoding features, wherein the processing is: converting the encoded features f of the target image... s As K and V, the encoding features f of other images in a three-dimensional medical image sequence d f is obtained by performing self-attention calculation on Q. m ; will f m As K and V, f s Q is used to perform self-attention calculation again to obtain f. a , for f a The pre-decoding features f of the target image are obtained by performing layer normalization and feedforward neural network calculation. o The decoder is used to process the f values ​​of each image respectively. o Decode the images to obtain the segmentation results for each image;

[0012] For labeled data, the training objective is to minimize the difference between the segmentation result of each image and its label; for unlabeled data, the training objective is to minimize the difference between the segmentation result of each image and the segmentation result obtained by inputting the encoded features of each image into the decoder.

[0013] Application phase:

[0014] The three-dimensional medical image to be segmented is input into the trained segmentation model to obtain the segmentation result.

[0015] According to a second aspect of the present invention, a three-dimensional medical image segmentation system based on temporal information enhancement is provided, comprising: a computer-readable storage medium and a processor;

[0016] The computer-readable storage medium is used to store executable instructions;

[0017] The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in the first aspect.

[0018] According to a third aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to perform the method as described in the first aspect.

[0019] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0020] 1. The method provided by this invention proposes a Circle Transformer module for the extraction and fusion of temporal information. It utilizes temporal input to improve the training effect of deep learning models and can effectively eliminate interference from similar features and blurred images. In the training phase, the input is a time-series sample, and the model training effect is enhanced by extracting temporal information. Furthermore, by simultaneously constraining the temporal information and combining the segmentation results before and after, the model's dependence on temporal sequence is eliminated. Compared with training methods that use a single input sample, this method can improve the accuracy of segmentation models based on encoding and decoding structures at no cost. In the application phase, only a single frame of 3D image needs to be input, without using a sequence as input, making the application mode more flexible.

[0021] 2. This invention proposes using labels to simultaneously constrain the segmentation results before and after combining temporal information, which can improve the training effect of the segmentation model with the encoding and decoding structure without increasing additional costs. Furthermore, a new consistency loss is proposed based on the combination of temporal information and the segmentation results before and after combining temporal information to improve the semi-supervised training effect of the model.

[0022] In summary, this invention combines deep learning with semi-supervised training of a segmentation model based on temporal information enhancement to train a 3D medical image segmentation model. During the semi-supervised training of the temporal information enhancement-based segmentation model, a Circle Transformer module is constructed to extract temporal motion information and optimize the training process. Simultaneously, it proposes to simultaneously supervise the segmentation results before and after combining temporal information, eliminating the model's dependence on temporal information and thus improving the training effect of the segmentation model at no cost. Furthermore, a new consistency loss is proposed based on the segmentation results before and after combining temporal information to constrain unlabeled data, thereby achieving semi-supervised training, which is beneficial for improving model performance and requires no additional memory. Attached Figure Description

[0023] Figure 1 This is a schematic diagram illustrating the training process of the segmentation model provided in an embodiment of the present invention.

[0024] Figure 2 The images shown are visual representations of the segmented images corresponding to Embodiment 1 and Comparative Examples 1-5 of the present invention. Figure 2 In the diagram, (a) to (h) correspond to the following images in sequence: original image, label, visual effect image after processing in Example 1, visual effect image after processing in Comparative Example 1, visual effect image after processing in Comparative Example 2, visual effect image after processing in Comparative Example 3, visual effect image after processing in Comparative Example 4, and visual effect image after processing in Comparative Example 5. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0026] This invention provides a three-dimensional medical image segmentation method based on temporal information enhancement. This method uses a Circle Transformer module to extract temporal information to enhance the training effect of the segmentation model. The trained model can segment target regions in medical images, such as... Figure 1 As shown, it includes:

[0027] Training phase:

[0028] Semi-supervised training of the segmentation model was performed using the training set and the Circle Transformer module.

[0029] The training set includes three-dimensional medical image sequences containing target regions, wherein the gold standard for target region segmentation of some of the three-dimensional medical image sequences is known and is labeled data, and the gold standard for target region segmentation of other three-dimensional medical image sequences is unknown and is unlabeled data;

[0030] The segmentation model includes an encoder and a decoder; wherein, the encoder is used to encode each image in the three-dimensional medical image sequence to obtain encoded features; the Circle Transformer module is used to process each image in the three-dimensional medical image sequence as a target image to obtain its pre-decoding features, wherein the processing is: converting the encoded features f of the target image... s As K and V, the encoding features f of other images in a three-dimensional medical image sequence d f is obtained by performing self-attention calculation on Q. m ; will f m As K and V, f s Q is used to perform self-attention calculation again to obtain f. a , for f a The pre-decoding features f of the target image are obtained by performing layer normalization with residual connections and feedforward neural network calculation. o The decoder is used to process the f values ​​of each image respectively. o Decode the images to obtain the segmentation results for each image;

[0031] For labeled data, the training objective is to minimize the difference between the segmentation result of each image and its label; for unlabeled data, the training objective is to minimize the difference between the segmentation result of each image and the segmentation result obtained by inputting the encoded features of each image into the decoder.

[0032] Application phase:

[0033] The three-dimensional medical image to be segmented is input into the trained segmentation model to obtain the segmentation result.

[0034] Specifically, during the training phase:

[0035] (1) Collect three-dimensional medical image sequences containing target regions and known partial sample target region segmentation gold standard, crop the original image sequences according to the region of interest and perform pixel value normalization to obtain training set samples;

[0036] (2) Select any mainstream deep learning segmentation model with any encoding and decoding structure, split it into an encoder and a decoder, and construct a semi-supervised training framework for the segmentation model based on temporal information enhancement. The framework first obtains the encoded features of all images in a sequence through an encoder with shared weights. Then, these encoded features are combined with temporal information through the CircleTransformer module to generate the pre-decoding features of each frame. Finally, the pre-decoding features are input into the decoder with shared weights to obtain the segmentation results of each frame. During training, for labeled data, the labels will supervise the segmentation results obtained from the encoded features and pre-decoding features through the segmentation loss. For unlabeled data, the segmentation results obtained from the encoded features and pre-decoding features will be kept consistent through the consistency loss.

[0037] Using the training set samples obtained in step (1), a semi-supervised training framework for a segmentation model based on temporal information enhancement is constructed and trained using the above method, so that the trained model can segment the target region contained in the input image.

[0038] In the application phase:

[0039] For the original 3D medical image to be segmented, the original image is cropped according to the region of interest, and the pixel values ​​of the cropped image are normalized to obtain the sample to be segmented.

[0040] The sample to be segmented obtained in step (2) is input as the input image into the trained segmentation model obtained in step (3), and the segmentation result of the target region in the sample to be segmented can be output.

[0041] The semi-supervised training framework for the segmentation model based on temporal information enhancement constructed in step (3) specifically includes:

[0042] The encoder and decoder are obtained by splitting the mainstream deep learning segmentation model of the encoding and decoding structure;

[0043] The Circle Transformer module is used for the extraction and fusion of temporal information. It consists of a self-attention calculation module, a feedforward neural network (FFN), and a layer normalization module. The self-attention calculation module performs two self-attention calculations. The FFN includes two fully connected layers to enhance the model's non-linear expressive power. The two inputs of the Circle Transformer module are represented as static features and dynamic features, respectively. During the training phase, when the Circle Transformer module processes a frame of a 3D medical image sequence, the encoded features of that frame are denoted as static features, and the encoded features of the remaining frames in the 3D medical image sequence are denoted as dynamic features. That is, different inputs are used to generate the pre-decoding features for different frames. Let F...e ={F e 1 ,F e 2 ,...,F e n}and Let represent the encoded features and undecoded features of n images, respectively. The generation of the undecoded features can be represented as:

[0044]

[0045] Regarding the structure of the Circle Transformer module, it generates pre-decoding features by fusing two sets of features through two self-attention computations that exchange query objects. Specifically, in the first self-attention computation, static features are treated as Key and Value, and dynamic features as Query. A weighted sum of the static features produces an intermediate variable of the same size as the dynamic features. In the second self-attention computation, this intermediate variable is treated as Key and Value, and the static features as Query. The result of the second self-attention computation is processed by layer normalization with residual connections and a feedforward neural network to obtain the pre-decoding features, which is the final output of the Circle Transformer module. The final output of the Circle Transformer module is a weighted sum of the static features from the two computations. Through this computation, the Circle Transformer module achieves weighted updates of static features based on the correlation between dynamic and static features, guided by the dynamic features, while maintaining the feature size.

[0046] The method provided by this invention utilizes the Circle Transformer module to introduce the motion information of the target between adjacent frames and the current frame image. This allows the model to focus on the shape and position features of the needle during training, resulting in higher segmentation accuracy and better robustness in complex environments after training. When calculating the features before decoding a certain frame image, let the static features be denoted as f. s The dynamic feature is f d The intermediate variable for self-attention computation is f. m The output of self-attention computation is f a The output of the Circle Transformer module is f o LN(·) and FFN(·) represent the functions corresponding to the names, and the calculation process is as follows:

[0047]

[0048] f o =LN(FFN(LN(f a )+f a))+LN(f a )+f a (2)

[0049] Segmentation loss (L) seg ), used to supervise the segmentation results obtained from encoded features and pre-decoding features in labeled samples, using Dice loss (L Dice ) and cross-entropy loss (L CE Composed of ) y and Let i represent the label value and predicted value of pixel i, respectively, and M be the total number of pixels in the sample. The segmentation loss is defined as:

[0050]

[0051] Consistency loss (L) cons This is used to ensure that the segmentation results obtained from encoded features and features before decoding are consistent for unlabeled data. and Let the predicted values ​​of pixel i in the segmentation results obtained from the encoded features and the undecoded features be defined as follows:

[0052]

[0053] Understandably, the segmentation loss and consistency loss can also be constructed with reference to relevant existing technologies.

[0054] As a further preferred embodiment of the present invention, before performing semi-supervised training on the segmentation model, the method further includes: cropping each image in the training set according to the region of interest, and then performing pixel value normalization processing.

[0055] Before inputting the three-dimensional medical image to be segmented into the trained segmentation model, the method further includes: cropping the three-dimensional medical image to be segmented according to the region of interest, and then performing pixel value normalization processing.

[0056] Specifically, by cropping the training set samples and rotating, translating, and flipping the cropped images, and then normalizing the pixel values, the computational complexity can be reduced and the diversity of training data can be increased.

[0057] Similarly, for the original 3D medical image to be segmented, the original image is cropped according to the region of interest, and the pixel values ​​of the cropped image are normalized to obtain the sample to be segmented.

[0058] That is, the semi-supervised training framework for the segmentation model based on temporal information enhancement described above can be used to train an image segmentation model with an encoding and decoding structure. After training, the model takes the sample to be segmented obtained by the image preprocessing module as input and outputs the segmentation result of the target region in the sample to be segmented. The training is performed using training set samples and the segmentation gold standard of some samples as labels. The training set samples are obtained by collecting three-dimensional medical image sequences containing target regions and knowing the segmentation gold standard of some sample target regions, cropping the original image sequences according to the region of interest and normalizing the pixel values.

[0059] Taking a medical image sequence with a total of n frames as an example, Figure 1 The diagram shown is an overall structural schematic of the deep learning training framework provided in this embodiment of the invention. Its calculation process is applied to the training of a 3D medical image segmentation model. After training, the model inputs an image to be segmented and obtains accurate biopsy needle region segmentation results. Specific steps may include:

[0060] (1) The N frames of images are respectively encoded by the encoder to generate coded features, denoted as F. e ={F e 1 ,F e 2 ,...,F e n}

[0061] (2) For the i-th frame image, F e i Considered a static feature, except for F e i The features outside are considered dynamic features, and pre-decoding features are generated through the Circle Transformer module, denoted as...

[0062] (3)F e and F d The segmentation results of the target region are generated by the decoder respectively. For labeled data, the segmentation loss is supervised according to the corresponding label. For unlabeled data, the segmentation results of the corresponding frames in the two are constrained by the consistency loss.

[0063] Example 1

[0064] Based on the above method, the actual 3D medical image segmentation model is trained, which includes the following steps:

[0065] (1) The dataset consists of ultrasound-guided kidney biopsies performed on ten beagle dogs. We collected 10 videos, including 443 three-dimensional volumetric data points. Each sequence was extracted to include 3 frames, resulting in a total of 423 sequences. Of these, 393 sequences were used for training and 30 sequences were used for testing. During training, the volumetric data in each sequence were subjected to online random data augmentation with the same parameters, including translation, scaling, flipping, and rotation.

[0066] (2) Using the above training set, the SLM-SA model proposed in the literature (Short-term and long-term memory self-attention network for segmentation of tumours in 3D medical images. CAAI Transactions on Intelligence Technology, 1-14, 2023) is trained using the training framework proposed in this invention. The optimal model parameters are loaded, applied to the test set, and the segmentation results of the biopsy needle region in the image are output. Finally, the segmentation results are quantitatively evaluated.

[0067] Furthermore, to verify the method of the present invention, the following comparative examples were designed (each comparative example uses the same dataset):

[0068] Comparative Example 1

[0069] The TriANet proposed in the literature (Triple attention network for videosegmentation. Neurocomputing 417, 202-211, 2020) is used to perform the task of segmenting the biopsy needle region in a three-dimensional medical image sequence. The same dataset, learning rate, number of iterations and optimizer parameters as the method of this invention are used for training.

[0070] Comparative Example 2

[0071] The VisTR method proposed in the literature (End-to-end video instance segmentation with transformers. In: Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition, 8737-8746, 2021) is used to perform the segmentation task of biopsy needle region in three-dimensional medical image sequence. The method is trained using the same dataset, learning rate, number of iterations and optimizer parameters as the method of this invention.

[0072] Comparative Example 3

[0073] The IFC proposed in the literature (Video instance segmentation using inter-frame communication transformers. In: Proceedings of the 35th Conference on Neural Information Processing Systems, 2021) is used to perform the task of segmenting the biopsy needle region in a three-dimensional medical image sequence. The same dataset, learning rate, number of iterations, and optimizer parameters as the method of this invention are used for training.

[0074] Comparative Example 4

[0075] The AOT proposed in the literature (Associating objects with transformers for video objects segmentation. In: Proceedings of the 35th Conference on Neural Information Processing Systems, 2021) is used to perform the task of segmenting the biopsy needle region in a three-dimensional medical image sequence. The same dataset, learning rate, number of iterations, and optimizer parameters as the method of this invention are used for training.

[0076] Comparative Example 5

[0077] The DeAOT method proposed in the literature (Decoupling features in hierarchical propagation for videoobject segmentation. In: Proceedings of the 35th Conference on Neural Information Processing Systems, 2022) is used to perform the segmentation task of biopsy needle region in three-dimensional medical image sequence. It is trained using the same dataset, learning rate, number of iterations, and optimizer parameters as the method of this invention.

[0078] Results analysis:

[0079] To demonstrate the advantages of the present invention, the segmentation effect of Example 1 was compared with that of Comparative Examples 1-5. Quantitative comparisons were made using the Dice coefficient and the needle tip positioning error (E). tip Needle length error (E) len ) and needle angle error (E) ang The DSC is defined as follows:

[0080]

[0081] Where TP, FP, and FN represent the number of true positives, false positives, and false negatives in the predicted pixel classification results compared to the gold standard, respectively. For E tip E len and E ang First, the needle tip position, length, and angle are extracted from the segmentation results and labels using a linear fitting algorithm, and then the mean square error (MSE) of each parameter is calculated.

[0082] Table 1 lists four quantitative evaluation indicators, as well as operating speed and parameter quantity, for the segmentation results of Example 1 and Comparative Examples 1-5. It can be seen from the table that, compared with Comparative Examples 1-5, Example 1 shows better performance in DSC, E... tip E len and Significant improvements were achieved in the four quantitative evaluation metrics. Furthermore, Example 1 involved only 0.87M parameters and achieved a computation speed of 55 frames per second (FPS), enabling faster and more accurate real-time segmentation of biopsy needles in three-dimensional ultrasound images.

[0083] Table 1. Segmentation accuracy, speed, and parameter count of each segmentation method on the selected dataset.

[0084]

[0085] To more intuitively demonstrate the superiority of the present invention, we provide visual effect diagrams of the segmented images corresponding to Embodiment 1 and Comparative Examples 1-5. For example... Figure 2 As shown in (a) to (h), (a) and (b) present the original medical images and their corresponding ground truth labels. (c)-(h) show the segmentation results obtained by the Ours, TriANet, VisTR, IFC, AOT, and DeAOT models, respectively, where the white, light gray, and dark gray areas represent the true positive, false positive, and false negative regions in the biopsy needle detection results, respectively. Clearly, our method detects biopsy needle length, tip position, and needle angle in good agreement with their gold standards.

[0086] The above embodiments, taking needle detection as an example, fully illustrate that the present invention is beneficial to improving the detection accuracy of biopsy needles in three-dimensional medical images. The method and system are based on deep learning and can be used to train segmentation models based on encoding and decoding structures, especially for segmentation tasks of biopsy needle regions in three-dimensional medical images.

[0087] The above embodiments are merely examples. In addition to biopsy needles, the method and system of the present invention are also applicable to the segmentation of three-dimensional medical images of other target areas such as surgical instruments and organs that move over time.

[0088] This invention provides a three-dimensional medical image segmentation system based on temporal information enhancement, comprising: a computer-readable storage medium and a processor;

[0089] The computer-readable storage medium is used to store executable instructions;

[0090] The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in any of the above embodiments.

[0091] This invention provides a computer-readable storage medium storing computer instructions that cause a processor to perform the method described in any of the above embodiments.

[0092] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A three-dimensional medical image segmentation method based on temporal information enhancement, characterized in that, include: Training phase: Semi-supervised training of the segmentation model was performed using the training set and the Circle Transformer module. The training set includes three-dimensional medical image sequences containing target regions, wherein the gold standard for target region segmentation of some of the three-dimensional medical image sequences is known and is labeled data, and the other part is unlabeled data; The segmentation model includes an encoder and a decoder; wherein, the encoder is used to encode each image in the three-dimensional medical image sequence to obtain encoded features; the Circle Transformer module is used to process each image in the three-dimensional medical image sequence as a target image to obtain its pre-decoding features, wherein the processing is: converting the encoded features f of the target image... s As K and V, the encoding features f of other images in a three-dimensional medical image sequence d f is obtained by performing self-attention calculation on Q. m ; will f m As K and V, f s Q is used to perform self-attention calculation again to obtain f. a , for f a The pre-decoding features f of the target image are obtained by performing layer normalization and feedforward neural network calculation. o The decoder is used to process the f values ​​of each image respectively. o Decode the images to obtain the segmentation results for each image; For labeled data, the training objective is to minimize the difference between the segmentation result of each image and its label. For unlabeled data, the training objective is to minimize the difference between the segmentation result of each image and the segmentation result obtained by inputting the encoded features of each image into the decoder. Application phase: The three-dimensional medical image to be segmented is input into the trained segmentation model to obtain the segmentation result.

2. The method as described in claim 1, characterized in that, f o =LN(FFN(LN(f a )+f a ))+LN(f a )+f a Where LN(·) and FFN(·) represent the functions corresponding to the names, f s ′、f m ′ are respectively f s f m The transpose of , where C is the length of the feature vector defined in the segmentation model.

3. The method as described in claim 1 or 2, characterized in that, Before performing semi-supervised training on the segmentation model, the method further includes: cropping each image in the training set according to the region of interest, and then performing pixel value normalization processing. Before inputting the three-dimensional medical image to be segmented into the trained segmentation model, the method further includes: cropping the three-dimensional medical image to be segmented according to the region of interest, and then performing pixel value normalization processing.

4. The method as described in claim 1 or 2, characterized in that, The loss during the training phase includes the segmentation loss L of the labeled data. seg and consistency loss of unlabeled data L cons .

5. The method as described in claim 4, characterized in that, Among them, L Dice For Dice's loss, L CE For cross-entropy loss, y and Let represent the label value and predicted value of pixel i in the segmentation result obtained from the features before decoding, respectively. and represents the predicted value of pixel i in the segmentation result obtained from the encoded features and the features before decoding, respectively, and M is the total number of pixels in the image.

6. A three-dimensional medical image segmentation system based on temporal information enhancement, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a processor to perform the method as described in any one of claims 1-5.

Images