Training method and device of mixed propagation interaction segmentation model, equipment and medium

Abstract

The application discloses a training method and device of a mixed propagation interaction segmentation model, equipment and a medium. The method comprises the following steps: obtaining current volume features and a current volume segmentation mask based on a current three-dimensional volume medical image, volume prompts and a previous volume segmentation mask; obtaining a current slice segmentation mask based on a current slice medical image, slice prompts and a previous slice segmentation mask; extracting slice features corresponding to a target slice medical image from the current volume features, and compiling the slice features into a slice feature dictionary; obtaining a target slice segmentation mask corresponding to the target slice medical image based on the current slice segmentation mask, the target slice medical image, the current slice medical image and the slice feature dictionary; and iteratively training the mixed propagation interaction segmentation model based on the target slice segmentation mask, the current volume segmentation mask and the current slice segmentation mask. The method fuses different propagation modes, takes into account macroscopic and microscopic information, and improves the segmentation accuracy.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a training method, apparatus, device and medium for a hybrid propagation interactive segmentation model. Background Technology

[0002] Intracerebral hemorrhage (ICH) is a leading cause of stroke-related death worldwide, accounting for 10%–20% of all stroke deaths. Due to its acute onset and rapid progression, timely diagnosis and treatment are crucial for patient survival. Accurate ICH segmentation is fundamental for correct diagnosis and personalized treatment planning; however, CT images of ICH often present significant challenges due to low contrast and blurred boundaries between the hemorrhage area and healthy tissue. Even experienced clinicians require substantial time for manual annotation, and the results are susceptible to image noise, resolution, and other factors, exhibiting high subjectivity and poor consistency.

[0003] In recent years, deep learning-based segmentation methods have provided new pathways to address the aforementioned problems. Among them, the Segment Anything Model (SAM)-based segmentation framework performs exceptionally well in natural images. When applied to the medical field, its interactive prompting mechanism significantly reduces the workload of manual annotation. For example, SAM-Med2D and MedSAM improve the segmentation accuracy of 2D slices through fine-tuning, but only focus on single slices, ignoring the spatial correlation in 3D volumes; SAM2UNet attempts to fuse multi-scale features through the UNet architecture, but still relies on a large amount of high-quality labeled data; SAMIHS improves efficiency by optimizing the interactive process, but its generalization ability in low-contrast ICH regions is limited. Although MedSAM2 performs well in 3D organ segmentation through its fast guidance mechanism, its response to the heterogeneous hemorrhage regions specific to ICH is insufficient, and its segmentation accuracy still needs improvement. Summary of the Invention

[0004] Therefore, it is necessary to provide a training method, apparatus, computer equipment, and storage medium for a hybrid propagation interactive segmentation model to address the above-mentioned technical problems, thereby solving at least one of the problems existing in the prior art.

[0005] Firstly, a training method for a hybrid propagation interactive segmentation model is provided, applied to the hybrid propagation interactive segmentation model, the method comprising:

[0006] Based on the current 3D volumetric medical image, volume cues, and the previous volume segmentation mask, the current volume features and the current volume segmentation mask are obtained.

[0007] Based on the current slice medical image, slice hints, and the previous slice segmentation mask, the current slice segmentation mask is obtained, wherein the current slice medical image is obtained by slicing the current three-dimensional volume medical image;

[0008] Extract slice features corresponding to the target slice medical image from the current volume features, and compile the slice features into a slice feature dictionary;

[0009] Based on the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary, the target slice segmentation mask corresponding to the target slice medical image is obtained;

[0010] Based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask, the hybrid propagation interactive segmentation model is iterated again until the preset convergence condition is met, thus obtaining the trained hybrid propagation interactive segmentation model.

[0011] Secondly, a training device for a hybrid propagation interaction segmentation model is provided, applied to the hybrid propagation interaction segmentation model, the device comprising:

[0012] The volume propagation unit is used to obtain the current volume features and the current volume segmentation mask based on the current 3D volume medical image, volume cues, and the previous volume segmentation mask.

[0013] The slice propagation unit is used to obtain the current slice segmentation mask based on the current slice medical image, slice prompts and the previous slice segmentation mask, wherein the current slice medical image is obtained by slicing the current three-dimensional volumetric medical image;

[0014] The feature compilation unit is used to extract slice features corresponding to the target slice medical image from the current volume features, and compile the slice features into a slice feature dictionary;

[0015] The target slice segmentation unit is used to obtain the target slice segmentation mask corresponding to the target slice medical image based on the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary;

[0016] The iteration unit is used to perform the next round of iteration on the hybrid propagation interactive segmentation model based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask, until the preset convergence condition is met, and the trained hybrid propagation interactive segmentation model is obtained.

[0017] Thirdly, a computer device is provided, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor, when executing the computer-readable instructions, implements the steps of the training method for the hybrid propagation interactive segmentation model as described above.

[0018] Fourthly, a readable storage medium is provided, which stores computer-readable instructions that, when executed by a processor, implement the steps of the training method for the hybrid propagation interactive segmentation model as described above.

[0019] The training method, apparatus, device, and medium for the aforementioned hybrid propagation interactive segmentation model include the following steps: obtaining current volume features and a current volume segmentation mask based on the current 3D volumetric medical image, volume hints, and the previous volume segmentation mask; obtaining a current slice segmentation mask based on the current slice medical image, slice hints, and the previous slice segmentation mask, wherein the current slice medical image is obtained by slicing the current 3D volumetric medical image; extracting slice features corresponding to the target slice medical image from the current volume features and compiling the slice features into a slice feature dictionary; obtaining a target slice segmentation mask corresponding to the target slice medical image based on the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary; and performing the next iteration of the hybrid propagation interactive segmentation model based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask until a preset convergence condition is met, thus obtaining the trained hybrid propagation interactive segmentation model. In this embodiment, by fusing volume propagation and slice propagation, it is possible to capture the three-dimensional features of the cerebral hemorrhage region and its spatial relationship with surrounding tissues through volume propagation, accurately grasping the overall morphology, location, and distribution of the hemorrhage site. Simultaneously, slice propagation focuses on the details of each slice, ensuring fine segmentation of subtle features such as the boundaries and textures of the hemorrhage region. This solves the problem of traditional methods struggling to simultaneously capture macroscopic and microscopic information. It not only significantly reduces segmentation errors caused by low contrast and blurred boundaries in cerebral hemorrhage images but also allows for the completion of the entire volume segmentation with only a small number of slice annotations, reducing the cost and workload of manual annotation. This provides strong technical support for the rapid and accurate diagnosis of cerebral hemorrhage and the development of personalized treatment plans, effectively promoting the improvement of clinical diagnostic efficiency and quality. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of a model architecture for a hybrid propagation interaction segmentation model in one embodiment of this application;

[0022] Figure 2 This is a flowchart illustrating the training method of the hybrid propagation interaction segmentation model in one embodiment of this application. Figure 1 ;

[0023] Figure 3 This is a flowchart illustrating the training method of the hybrid propagation interaction segmentation model in one embodiment of this application. Figure 2 ;

[0024] Figure 4 This is a schematic diagram of the segmentation results of multiple comparison models in a visualization experiment in one embodiment of this application;

[0025] Figure 5 This is a schematic diagram of the structure of a training device for a hybrid propagation interaction segmentation model in one embodiment of this application;

[0026] Figure 6 This is a schematic diagram of a computer device according to one embodiment of this application. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] The training method for the hybrid propagation interaction segmentation model provided in this embodiment can be applied to, for example... Figure 1 The model architecture of the hybrid propagation interaction segmentation model includes a volume interaction module (VIM), a slice interaction module (SIM), a feature conversion module (FCM), and a multi-propagation feature fusion module (MPFFM).

[0029] First, the 3D medical image to be segmented can be acquired and sliced ​​to obtain various sliced ​​medical images, which can be stored in the storage module. This 3D medical image, the doodle (volume hint), and the volume mask obtained from the previous iteration are input into the volume interaction module VIM for volume segmentation to obtain the volume segmentation mask and volume features. Simultaneously, the current sliced ​​medical image (S...) can be... r The graffiti r (slice hint) and the slice mask r obtained from the previous iteration are input together into the slice interaction module SIM for slice segmentation to obtain the current slice segmentation mask. Then, the target slice medical image (S) is determined. k The feature transformation module FCM can extract the target slice medical image (S) from the volumetric features. k The slice features corresponding to the target slice are compiled to obtain a slice feature dictionary. Finally, the target slice medical image, the current slice medical image, the slice feature dictionary, and the current slice segmentation mask are input into the Multi-Propagation Feature Fusion (MPFFM) module for feature fusion to obtain the final segmentation mask (M) corresponding to the target slice. k The above segmentation operation can be performed on each slice until all slices of the 3D medical image to be segmented are completed. The VIM module is used to capture 3D structural features, the SIM module is introduced to focus on slice-level segmentation, and the FCM module ensures seamless integration by converting volumetric features into slice-level representations. The MPFFM module further refines the segmentation by utilizing predicted and target slices. Furthermore, the model supports multiple interactive iterations, thus achieving progressive improvement.

[0030] The volume interaction module VIM is implemented based on the ResidualUNet3d architecture, a deep learning architecture specifically designed for 3D medical image segmentation. It combines the advantages of UNet's encoder-decoder structure and residual connections. The SIM slice interaction module SIM operates using the Deeplabv3 framework. Deeplabv3 is an advanced semantic segmentation framework that achieves excellent performance in semantic segmentation tasks through dilated convolution and atrous spatial pyramid pooling (ASPP), making it particularly suitable for scenarios requiring precise boundary and multi-scale recognition.

[0031] In one embodiment, such as Figure 2 , Figure 3 As shown, a training method for a hybrid propagation interaction segmentation model is provided, which is applied to the hybrid propagation interaction segmentation model and includes the following steps:

[0032] In step S110, based on the current three-dimensional volumetric medical image, volumetric cues, and the previous volumetric segmentation mask, the current volumetric features and the current volumetric segmentation mask are obtained.

[0033] Optionally, volume hints can be user-annotated doodles, used to represent lesion areas, providing prior information to guide VIM to focus on specific areas or correct prediction biases. The previous volume segmentation mask can be the volume segmentation mask generated in the previous iteration. First, the preprocessed image (e.g., normalized, resampled) is input into the encoder. High-level semantic features are gradually extracted and spatial dimensions are reduced through multiple residual blocks and downsampling operations (e.g., 3D max pooling or stride convolution), while retaining intermediate layer feature maps. Subsequently, the decoder uses upsampling operations (e.g., 3D deconvolution) to restore the spatial dimensions, fuses features from corresponding encoder layers through skip connections, combines low-level details with high-level semantics, and finally generates a volume segmentation mask through thresholding or argmax operations. The current volume features can be multi-scale depth features extracted from the encoder bottleneck layer or the decoder intermediate layer.

[0034] In step S120, a current slice segmentation mask is obtained based on the current slice medical image, slice prompts, and the previous slice segmentation mask, wherein the current slice medical image is obtained by slicing the current three-dimensional volume medical image;

[0035] Optionally, the slice hint refers to the user-specified slice hint (Hr), and the previous slice segmentation mask refers to the segmentation mask of the target slice obtained in the previous iteration. First, the current slice medical image, slice hint, and previous slice segmentation mask can be preprocessed separately, such as normalizing and enhancing the current slice image, and aligning the slice hint, previous slice segmentation mask, and current slice image sizes. Then, the preprocessed current slice medical image, slice hint, and previous slice segmentation mask are input into SIM (Deeplabv3 framework). Multi-scale features are extracted by the DeepLabv3 backbone network (ResNet / Xception with dilated convolutions), and contextual information from different receptive fields is captured in parallel through the ASPP module. The decoder then fuses shallow details and high-level semantics. During this process, a slice hint-guided spatial / channel attention mechanism is used to enhance the target region feature response, and consistency loss constrains the structural similarity between the current prediction and the previous mask. Finally, the slice segmentation mask is output.

[0036] In step S130, slice features corresponding to the target slice medical image are extracted from the current volume features, and the slice features are compiled into a slice feature dictionary;

[0037] It should be noted that the target slice medical image can be determined based on the index of the current slice medical image. After slicing the 3D medical image to be segmented to obtain multiple slice medical images, the slice medical images can be encoded according to certain rules, such as sorting them according to the slicing order and then encoding them sequentially as CT-1, CT-2, ..., CT-n. If there are 20 slice medical images in total, and the index of the current slice medical image (the starting slice medical image) is CT-10, then it is necessary to segment the slice medical images upwards from CT-1 to CT-9, and downwards from CT-11 to CT-20. The target slice medical images are then CT-11 and CT-9.

[0038] Optionally, the current volumetric feature may include each slice feature corresponding to the current 3D volumetric medical image. First, the index of the target slice medical image can be determined based on the index of the current slice medical image. Then, based on this index, the corresponding slice features are selected and processed in each part of the current volumetric feature, and the extracted features are compiled into a feature dictionary by the feature transformation module FCM. The feature transformation module FCM extracts relevant features for slice segmentation based on a dictionary method, and may include multiple convolutional layers.

[0039] In step S140, based on the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary, the target slice segmentation mask corresponding to the target slice medical image is obtained;

[0040] Optionally, the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary are input into the Multi-Propagation Feature Fusion Module (MPFFM). Inspired by the Spatio-Temporal Context Network (STCN), the MPFFM internally utilizes an attention mechanism to achieve effective feature propagation between slices. First, the target slice medical image S is queried by the encoder. k Current slice medical image S r Multi-scale features are generated and then utilized by the current slice of medical image S through a memory encoder. r and its advanced features f 3 r Process the current slice segmentation mask M r And calculate the target slice medical image S k With current slice medical image S r Correlation of key features between Then the slice feature dictionary F is decoded. k Correlation, processed eigenvalues ​​V rMulti-scale feature median representation V of target slice medical images k With intermediate features f 1 k and f 2 k These are integrated together to obtain the target slice segmentation mask.

[0041] It should be noted that the Multi-Propagation Feature Fusion (MPFFM) module may include a dual-branch encoder, an attention computation unit, and a multi-scale fusion decoder. The dual-branch encoder extracts multi-scale features and key-value representations from the current slice and a reference slice (combined with a mask), respectively. The attention computation unit calculates pixel-level similarity between slices using a symmetric quadratic form formula and achieves feature propagation. The multi-scale fusion decoder fuses the multi-scale features of the current slice with the propagated features and outputs a predicted mask. The module also uses a mask-guided memory mechanism to focus on the target region, balancing spatial details and semantic information, achieving efficient information integration between slices, and is suitable for fine segmentation of 3D medical images. The dual-branch encoder includes a query encoder and a memory encoder. The query encoder is typically composed of a convolutional neural network (such as a ResNet backbone) and convolutional layers to generate feature maps at different scales. The memory encoder uses a mask-guided attention mechanism to focus on the target region, balancing spatial details and semantic information, achieving efficient information integration between slices, and is suitable for fine segmentation of 3D medical images. r Information integration S r In the feature representation, the focus on the target region is enhanced.

[0042] In step S150, based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask, the hybrid propagation interactive segmentation model is iterated again until the preset convergence condition is met, and the trained hybrid propagation interactive segmentation model is obtained.

[0043] Optionally, after completing the target slice segmentation, the current volume segmentation mask is used as the previous volume segmentation mask for the next round, and input along with the volume hints and the 3D volumetric medical image into VIM for the next round of volume segmentation to obtain a new volume segmentation mask. Simultaneously, the current slice segmentation mask is used as the previous slice segmentation mask for the next round, and input along with the slice hints and the target slice medical image into SIM to obtain a new slice segmentation mask. Next, based on the index of the target slice medical image in this round, slice features are extracted from the volume features obtained in this round of segmentation, and a feature dictionary is compiled. Finally, the target slice medical image, slice medical image, slice segmentation mask, and feature dictionary are input into MPFFM to obtain the segmentation mask for the target slice in this round. The above steps are repeated until a preset convergence condition is met, such as reaching the upper limit of the number of iterations or the loss value falling below a set threshold, to obtain the trained hybrid propagation interactive segmentation model.

[0044] This application provides a training method for a hybrid propagation interactive segmentation model, comprising: obtaining current volume features and a current volume segmentation mask based on a current 3D volumetric medical image, volume hints, and a previous volume segmentation mask; obtaining a current slice segmentation mask based on a current slice medical image, slice hints, and a previous slice segmentation mask, wherein the current slice medical image is obtained by slicing the current 3D volumetric medical image; extracting slice features corresponding to a target slice medical image from the current volume features and compiling the slice features into a slice feature dictionary; obtaining a target slice segmentation mask corresponding to the target slice medical image based on the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary; and performing a next round of iteration on the hybrid propagation interactive segmentation model based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask, until a preset convergence condition is met, thereby obtaining a trained hybrid propagation interactive segmentation model. In this embodiment, by fusing volume propagation and slice propagation, it is possible to capture the three-dimensional features of the cerebral hemorrhage region and its spatial relationship with surrounding tissues through volume propagation, accurately grasping the overall morphology, location, and distribution of the hemorrhage site. Simultaneously, slice propagation focuses on the details of each slice, ensuring fine segmentation of subtle features such as the boundaries and textures of the hemorrhage region. This solves the problem of traditional methods struggling to simultaneously capture macroscopic and microscopic information. It not only significantly reduces segmentation errors caused by low contrast and blurred boundaries in cerebral hemorrhage images but also allows for the completion of the entire volume segmentation with only a small number of slice annotations, reducing the cost and workload of manual annotation. This provides strong technical support for the rapid and accurate diagnosis of cerebral hemorrhage and the development of personalized treatment plans, effectively promoting the improvement of clinical diagnostic efficiency and quality.

[0045] In one embodiment of this application, the step of extracting features based on the current three-dimensional volumetric medical image, volume cues, and the previous volume segmentation mask to obtain the current volumetric features includes:

[0046] The current 3D volumetric medical image, volumetric hints, and previous volumetric segmentation mask are preprocessed.

[0047] The preprocessed current 3D volumetric medical image, volume prompt, and previous volume segmentation mask are encoded to obtain encoded features;

[0048] The encoded features and decoded features are fused by skip connections to obtain the current volume features.

[0049] Optionally, the current 3D volumetric medical image, volume cues, and the previous volume segmentation mask are preprocessed and then concatenated along the channel dimension to form a multimodal input, which is then fed into a ResidualUNet3D architecture. In the encoder, high-level semantic features are gradually extracted and the spatial dimension is reduced through multiple residual blocks combined with 3D max pooling or stride convolution operations. Residual connections are used to alleviate gradient vanishing. Meanwhile, the volume cues are encoded to generate a spatial attention map, enhancing the feature response of the target region. The previous volume segmentation mask provides structural priors to the network through feature embedding. The decoder uses upsampling operations such as 3D deconvolution or bilinear interpolation to restore the spatial dimension. Skip connections are used to fuse the features of the corresponding level of the encoder with the current volume features, further incorporating the structural information of the previous volume segmentation mask. Finally, the final feature map is mapped to the class space through convolution, and the probability of each voxel belonging to each class is generated using the Softmax activation function to obtain a 3D probability map. After thresholding or argmax operations, the class of each voxel is determined, thereby generating the current volume segmentation mask. In addition, deep features containing global semantics can be extracted from the encoder bottleneck layer, or features combining semantic and spatial information can be extracted from the decoder intermediate layer as the current volume features.

[0050] For example, its input data consists of volume data (VD), user cues (H), and the prediction mask (PM) from the previous iteration, denoted as {VD, H, PM}. In VIM, the encoding and decoding structures are defined as {E1, E2, E3, ...} and {D}, respectively. n D n-1 ..., D3, D2, D1}. Each coding layer is derived from the output of its preceding layer (E... out i-1 ) Receive input (E in i Similarly, the input (D) of each decoding layer in i ) is achieved by using the next decoding layer (D out i+1 The output of ) and the output of the corresponding coding layer (E) out i It is constructed by connecting these elements. The final output of the decoding process is a volumetric segmentation mask (VM=D). out 1) and 3D volumetric feature map (Feature Map = {D out 1, D out 2, D out 3}).

[0051] It's important to note that the VM only functions during training, tuning module parameters to optimize the segmentation process. Simultaneously, the feature maps integrate multi-scale volumetric information, improving the accuracy and robustness of the segmentation process by providing key features for subsequent stages. After each round of interaction, the VM is updated in memory and used as input for the next iteration. In the first round of interaction, the PM is initialized with a zero-mapping structure identical to the VD dimension, ensuring a consistent starting point for the iteration process.

[0052] In one embodiment of this application, the current slice segmentation mask is obtained based on the current slice medical image, slice hints, and the previous slice segmentation mask, including:

[0053] The current slice medical image, slice prompts, and previous slice segmentation mask are preprocessed;

[0054] Based on the preprocessed current slice medical image, slice hints, and previous slice segmentation mask, multi-scale features are extracted;

[0055] The multi-scale features are fused based on an attention mechanism to obtain the current slice segmentation mask.

[0056] Optionally, the current sliced ​​medical image, slice hints (such as labeled region masks), and the previous slice segmentation mask are preprocessed (standardization, size alignment, etc.) and then concatenated along the channel dimension to form a multimodal input. Multi-scale features are extracted by the DeepLabv3 backbone network (ResNet / Xception with dilated convolutions). Subsequently, the ASPP module further captures multi-scale context by using multiple convolutional kernels with different dilation rates in parallel, along with a global average pooling branch. The decoder upsamples the high-level semantic features output by ASPP through bilinear interpolation or transposed convolutions and concatenates them with shallow detail features from the corresponding layers of the backbone network (such as shallow feature maps from ResNet). Skip connections are used to preserve details such as image edges and textures, achieving deep fusion of shallow details and high-level semantics. During the fusion process, slice hints can be used to guide spatial / channel attention mechanisms to enhance the target region feature response, and consistency loss constrains the structural similarity between the current prediction and the previous mask, ultimately outputting a slice segmentation mask.

[0057] For example, the input data consists of slice data (VD) r User-specified slice hints (H) r ) and the prediction mask of the previous interaction (PM) r ) is represented as {VD r H r PM r}. Here, r represents the index of a specific slice in the volume data. The output of this module is Mr. After each round of interaction, PM rIt will be updated in memory to provide input for subsequent interactions. In the initial interaction round, PMR starts with a zero-map of the same dimensions as VDr to establish a baseline for the segmentation process.

[0058] In one embodiment of this application, the step of extracting slice features corresponding to the target slice medical image from the current volume features and compiling the slice features into a slice feature dictionary includes:

[0059] Determine the slice index of the target slice medical image;

[0060] Based on the slice index, the corresponding initial slice feature is extracted from the current volume feature;

[0061] Perform a corresponding convolution operation on each initial slice feature, and integrate the initial slice features after the convolution operation to obtain the slice feature dictionary.

[0062] Optionally, the volumetric features include multi-size features. First, the index of the target slice medical image can be determined based on the index of the current slice medical image. Then, based on this index, the corresponding slice features are selected and processed in each part of the current volumetric features, and the extracted features are compiled into a feature dictionary through the feature transformation module FCM. The details are as follows:

[0063]

[0064] Where Conv represents the convolution operation, and D out 1, D out 2, D out 3 represents different levels of 3D volumetric feature maps (such as shallow details, mid-level semantics, and deep abstract features), and k represents the slice index.

[0065] In one embodiment of this application, determining the slice index of the target slice medical image includes:

[0066] Determine the slice index of the current slice medical image;

[0067] The slice index of the target slice medical image is determined based on the slice index of the current slice medical image.

[0068] It should be noted that the target slice medical image can be determined based on the index of the current slice medical image. After slicing the 3D medical image to be segmented to obtain multiple slice medical images, the slice medical images can be encoded according to certain rules, such as sorting them according to the slicing order and then encoding them sequentially as CT-1, CT-2, ..., CT-n. If there are 20 slice medical images in total, and the index of the current slice medical image (the starting slice medical image) is CT-10, then it is necessary to segment the slice medical images upwards from CT-1 to CT-9, and downwards from CT-11 to CT-20. The target slice medical images are then CT-11 and CT-9.

[0069] In one embodiment of this application, obtaining the target slice segmentation mask corresponding to the target slice medical image based on the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary includes:

[0070] The current slice multi-scale features corresponding to the current slice medical image and the target slice multi-scale features corresponding to the target slice medical image are generated respectively. The current slice multi-scale features include current slice high-level features, current slice key features and current slice intermediate features. The target slice multi-scale features include target slice high-level features, target slice key features and target slice intermediate features.

[0071] The segmentation mask of the current slice is encoded using the current slice medical image and the high-level features of the current slice to obtain a value representation;

[0072] Determine the correlation between the key features of the current slice and the key features of the target slice;

[0073] The slice feature dictionary, value representation, correlation, intermediate features of the current slice, and intermediate features of the target slice are integrated to obtain the target slice segmentation mask.

[0074] Optionally, the MPFFM module includes a dual-branch encoder, an attention computation unit, and a multi-scale fusion decoder. The dual-branch encoder may include a query encoder and a memory encoder. The query encoder extracts features from both the current slice medical image and the target slice medical image to generate the current slice multi-scale features {f} corresponding to the current slice medical image. 1 k f 2 k f 3 k K k V k}, and the target slice multi-scale features {f} corresponding to the target slice medical image. 1 r f 2 r f 3 r K r V r '}. Meanwhile, the memory encoder S r and its advanced features f 3 r Processing mask M r V is expressed as a calculated value. r Specifically, it can be expressed by the following formula:

[0075]

[0076] Then the following formula can be used to calculate the two slices (K). k and K r The correlations between the key features of ) are as follows:

[0077]

[0078] Finally, the predicted mask for Sk is generated by a mask decoder, which will convert F... k V k V r and With intermediate feature f 1 k and f 2 k When integrated, we get:

[0079]

[0080] In one embodiment of this application, after obtaining the target slice segmentation mask corresponding to the target slice medical image, the method further includes:

[0081] Based on the volume hints and the current volume segmentation mask, calculate the volume segmentation loss value;

[0082] Calculate the slice segmentation loss value based on the target slice segmentation mask and the current slice segmentation mask;

[0083] Based on the slice feature dictionary and the target slice segmentation mask, the feature transformation loss value is calculated;

[0084] Based on the volume segmentation loss value, slice segmentation loss value, and feature transformation loss value, the total loss value is obtained;

[0085] If the total loss value is greater than the preset loss threshold, then the next iteration of the hybrid propagation interactive segmentation model is performed based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask.

[0086] Optionally, the model can be updated using three different sub-losses, which may include volume segmentation loss (VSL), slice segmentation loss (SSL), and feature transformation loss (FCL). Each sub-loss includes L dice and L ce L dice This can be expressed by the following formula:

[0087]

[0088] Where N represents the number of samples, p i g represents the segmentation mask for the i-th voxel in the predicted segmentation result (such as the current volume segmentation mask, the current slice segmentation mask, or the target slice segmentation mask). i This indicates the actual label.

[0089] Among them, L ce This can be expressed by the following formula:

[0090]

[0091] The formula for calculating each sub-loss can be expressed as:

[0092] L=αL dice +βL ce ;

[0093] Both α and β were experimentally set to 0.5.

[0094] It should be noted that the input data for each sub-loss is different. For Volume Segmentation Loss (VSL), the input is the current volume segmentation mask and the true volume label. For Slice Segmentation Loss (SSL), the input is... r and S r Between and M k and S k The relationship is determined between these factors. For Feature Transform Loss (FCL), the loss is F... k Each element in M k The sum of the losses.

[0095] Therefore, the total loss (Ltotal) can be obtained by weighted summation of the three sub-losses mentioned above, which can be specifically expressed as:

[0096]

[0097] Wherein, λ1, λ2 and λ3 are weighting coefficients, which can be set to 5, 1 and 4 respectively.

[0098] To further verify the practical technical effectiveness of this application, the following experiments were designed. First, a dataset was collected, which could be a private dataset from a collaborator, such as a hospital, and could include 286 cases. Each case included the original volumetric CT image and the corresponding mask for the ICH region. These volumetric data were used after windowing thresholding and normalization. Although the number of slices varied depending on the case, all slices maintained a consistent resolution of 512×512 pixels to ensure consistency in processing and analysis. The dataset also included the Physionet dataset, a public dataset containing 75 cases. To ensure meaningful analysis, only cases with three or more slices containing hemorrhage areas were selected. The provided preprocessing methods were used to standardize the data. For the segmentation task, the dataset included the original CT images and the corresponding segmentation masks. Furthermore, each slice of the volumetric data had a uniform resolution of 512×512 pixels.

[0099] To evaluate model performance, several metrics were used, including Dice, Jaccard, Hausdorff distance (HD), and mean absolute error (MAE). During training, the model was configured for 500 epochs with a learning rate of 1×10⁻⁶. - 5. Weights were decayed to 1×10⁻⁷ to mitigate overfitting, along with the Adam optimizer. This setup included CUDA version 11.8, PyTorch version 2.5.0, and an NVIDIA 3090Ti GPU. Graffiti was generated for each slice in the dataset. To ensure generalization, S... r and S k The mask is randomly selected from the volume, allowing prediction of any slice based on the starting slice. During evaluation, a graffiti is generated for a given slice, and volumetric segmentation is performed iteratively across the entire volume. Comparative experiments were conducted with other state-of-the-art methods and ablation studies to analyze the effectiveness of these two segmentation strategies. Furthermore, visualizations of sample results and performance variations with the number of interactions are provided to illustrate the impact of final predictions and intelligent interactions. All experiments were performed using five-fold cross-validation.

[0100] The aforementioned hybrid propagation interaction segmentation models were initially evaluated against several state-of-the-art models, including SAMed, SAM-Med2D, SAM2-UNet, SAMIHS, MedSAM, and MedSAM2. To ensure fair comparison, default parameters from the open-source code of each comparison method were used, aligning the data volume with the number of training epochs. Experimental results are shown in Table 1 below:

[0101] Table 1: Comparative Experiment Results

[0102]

[0103] As shown in Table 1 above, the hybrid propagation interaction segmentation model trained in this application outperforms all other models on every metric. Specifically, the hybrid propagation interaction segmentation model shows significant improvements on both our private and public datasets, increasing the scores of Dice, Jaccard, HD, and MAE by at least 0.1116, 0.1028, 0.03, and 0.0007 on the private dataset, and by 0.0489, 0.0107, 0.66, and 0.0013 on the Physionet dataset. Even with fewer annotations, the hybrid propagation interaction segmentation model consistently achieves the best results, demonstrating its superiority.

[0104] To evaluate the effectiveness of the hybrid propagation segmentation strategy, ablation experiments can be conducted by training models using only volume propagation (referred to as Volume-Pro) or only slice propagation (referred to as Slice-Pro).

[0105] The details are shown in Table 2 below:

[0106] Table 2: Ablation Experiment Results

[0107]

[0108] As shown in Table 2 above, the results demonstrate that the hybrid propagation interactive segmentation model significantly outperforms the two individual propagation methods. Specifically, the hybrid strategy improves the scores of Dice, Jaccard, HD, and MAE by at least 0.1243, 0.1277, 0.73, and 0.00143, respectively, on the private dataset, and by 0.1116, 0.0942, 1.33, and 0.01255, respectively, on the Physionet dataset. These results highlight the significant contribution of the training method presented in this application to the overall performance of the hybrid propagation interactive segmentation model, emphasizing its effectiveness in improving segmentation results.

[0109] Furthermore, a visualization experiment was designed, which provides strong evidence for the effectiveness of this approach, such as... Figure 4 As shown, most methods perform well in simpler cases, such as the method in (b), with high segmentation accuracy. However, in more complex cases, our method outperforms others in both accuracy and competitiveness. For example, in challenging cases with multiple hemorrhage points (a and c), our method demonstrates superior performance. Even in samples with very small outliers (such as (d)), our method still achieves significant results.

[0110] Finally, interactive experiments were conducted to simulate the interaction process between the user and the hybrid propagation interactive segmentation model. The model generates updated segmentation results based on each new doodle provided by the user, and then uses these results for the next interaction. Details are shown in Table 3 below:

[0111] frequency 1 2 3 4 5 6 7 8 Dice 0.6133 0.7763 0.7770 0.7831 0.7820 0.7856 0.7868 0.7865 Jaccard 0.4499 0.6471 0.6469 0.6537 0.6527 0.6567 0.6583 0.6580 HD 4.50 3.61 3.61 3.66 3.64 3.60 3.56 3.56 MAE 0.0065 0.0043 0.0043 0.0043 0.0042 0.0042 0.0042 0.0042

[0112] As shown in Table 3 above, each interaction updates the segmentation results, demonstrating the most significant improvement in the initial stage. After the second interaction, the results begin to stabilize, but overall there is still an improvement. These results highlight the effectiveness of our interactive approach and its ability to progressively refine the segments through additional user input.

[0113] In this embodiment, by fusing volume propagation and slice propagation, it is possible to capture the three-dimensional features of the cerebral hemorrhage region and its spatial relationship with surrounding tissues through volume propagation, accurately grasping the overall morphology, location, and distribution of the hemorrhage site. Simultaneously, slice propagation focuses on the details of each slice, ensuring fine segmentation of subtle features such as the boundaries and textures of the hemorrhage region. This solves the problem of traditional methods struggling to simultaneously capture macroscopic and microscopic information. It not only significantly reduces segmentation errors caused by low contrast and blurred boundaries in cerebral hemorrhage images but also allows for the completion of the entire volume segmentation with only a small number of slice annotations, reducing the cost and workload of manual annotation. This provides strong technical support for the rapid and accurate diagnosis of cerebral hemorrhage and the development of personalized treatment plans, effectively promoting the improvement of clinical diagnostic efficiency and quality.

[0114] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0115] In one embodiment, a training apparatus for a hybrid propagation interactive segmentation model is provided, which corresponds one-to-one with the training method for the hybrid propagation interactive segmentation model in the above embodiments. For example... Figure 5 As shown, the training device for this hybrid propagation interactive segmentation model includes a volume propagation unit 10, a slice propagation unit 20, a feature compilation unit 30, a target slice segmentation unit 40, and an iteration unit 50. Detailed descriptions of each functional module are as follows:

[0116] The volume propagation unit 10 is used to obtain the current volume features and the current volume segmentation mask based on the current three-dimensional volume medical image, volume cues and the previous volume segmentation mask;

[0117] The slice propagation unit 20 is used to obtain the current slice segmentation mask based on the current slice medical image, slice prompts and the previous slice segmentation mask, wherein the current slice medical image is obtained by slicing the current three-dimensional volume medical image;

[0118] The feature compilation unit 30 is used to extract slice features corresponding to the target slice medical image from the current volume features, and compile the slice features into a slice feature dictionary;

[0119] The target slice segmentation unit 40 is used to obtain the target slice segmentation mask corresponding to the target slice medical image based on the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary;

[0120] The iteration unit 50 is used to perform the next round of iteration on the hybrid propagation interactive segmentation model based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask, until the preset convergence condition is met, and the trained hybrid propagation interactive segmentation model is obtained.

[0121] In one embodiment of this application, the volume propagation unit 10 is further configured to:

[0122] The current 3D volumetric medical image, volumetric hints, and previous volumetric segmentation mask are preprocessed.

[0123] The preprocessed current 3D volumetric medical image, volume prompt, and previous volume segmentation mask are encoded to obtain encoded features;

[0124] The encoded features and decoded features are fused by skip connections to obtain the current volume features.

[0125] In one embodiment of this application, the slice propagation unit 20 is further configured to:

[0126] The current slice medical image, slice prompts, and previous slice segmentation mask are preprocessed;

[0127] Based on the preprocessed current slice medical image, slice hints, and previous slice segmentation mask, multi-scale features are extracted;

[0128] The multi-scale features are fused based on an attention mechanism to obtain the current slice segmentation mask.

[0129] In one embodiment of this application, the feature compilation unit 30 is further configured to:

[0130] Determine the slice index of the target slice medical image;

[0131] Based on the slice index, the corresponding initial slice feature is extracted from the current volume feature;

[0132] Perform a corresponding convolution operation on each initial slice feature, and integrate the initial slice features after the convolution operation to obtain the slice feature dictionary.

[0133] In one embodiment of this application, the feature compilation unit 30 is further configured to:

[0134] Determine the slice index of the current slice medical image;

[0135] The slice index of the target slice medical image is determined based on the slice index of the current slice medical image.

[0136] In one embodiment of this application, the target slice segmentation unit 40 is further configured to:

[0137] The current slice multi-scale features corresponding to the current slice medical image and the target slice multi-scale features corresponding to the target slice medical image are generated respectively. The current slice multi-scale features include current slice high-level features, current slice key features and current slice intermediate features. The target slice multi-scale features include target slice high-level features, target slice key features and target slice intermediate features.

[0138] The segmentation mask of the current slice is encoded using the current slice medical image and the high-level features of the current slice to obtain a value representation;

[0139] Determine the correlation between the key features of the current slice and the key features of the target slice;

[0140] The slice feature dictionary, value representation, correlation, intermediate features of the current slice, and intermediate features of the target slice are integrated to obtain the target slice segmentation mask.

[0141] In one embodiment of this application, the apparatus further includes: a loss calculation unit, used for:

[0142] Based on the volume hints and the current volume segmentation mask, calculate the volume segmentation loss value;

[0143] Calculate the slice segmentation loss value based on the target slice segmentation mask and the current slice segmentation mask;

[0144] Based on the slice feature dictionary and the target slice segmentation mask, the feature transformation loss value is calculated;

[0145] Based on the volume segmentation loss value, slice segmentation loss value, and feature transformation loss value, the total loss value is obtained;

[0146] If the total loss value is greater than the preset loss threshold, then the next iteration of the hybrid propagation interactive segmentation model is performed based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask.

[0147] In this embodiment, by fusing volume propagation and slice propagation, it is possible to capture the three-dimensional features of the cerebral hemorrhage region and its spatial relationship with surrounding tissues through volume propagation, accurately grasping the overall morphology, location, and distribution of the hemorrhage site. Simultaneously, slice propagation focuses on the details of each slice, ensuring fine segmentation of subtle features such as the boundaries and textures of the hemorrhage region. This solves the problem of traditional methods struggling to simultaneously capture macroscopic and microscopic information. It not only significantly reduces segmentation errors caused by low contrast and blurred boundaries in cerebral hemorrhage images but also allows for the completion of the entire volume segmentation with only a small number of slice annotations, reducing the cost and workload of manual annotation. This provides strong technical support for the rapid and accurate diagnosis of cerebral hemorrhage and the development of personalized treatment plans, effectively promoting the improvement of clinical diagnostic efficiency and quality.

[0148] Specific limitations regarding the training device for the hybrid propagation interactive segmentation model can be found in the limitations on the training method for the hybrid propagation interactive segmentation model described above, and will not be repeated here. Each module in the training device for the aforementioned hybrid propagation interactive segmentation model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0149] In one embodiment, a computer device is provided, which may be a terminal device, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes a readable storage medium storing computer-readable instructions. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer-readable instructions implement a training method for a hybrid propagation interactive segmentation model. The readable storage medium provided in this embodiment includes both non-volatile and volatile readable storage media.

[0150] In this application embodiment, a computer device is provided, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor. When the processor executes the computer-readable instructions, it implements the steps of the training method of the hybrid propagation interactive segmentation model as described above.

[0151] In one embodiment of the application, a readable storage medium is provided, which stores computer-readable instructions. When the computer-readable instructions are executed by a processor, they implement the steps of the training method of the hybrid propagation interactive segmentation model described above.

[0152] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware with computer-readable instructions. These computer-readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium. When executed, these computer-readable instructions can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can 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 a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0153] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0154] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A training method for a hybrid propagation interaction segmentation model, characterized in that, The method, applied to a hybrid propagation interaction segmentation model, includes: Based on the current 3D volumetric medical image, volume cues, and the previous volume segmentation mask, the current volume features and the current volume segmentation mask are obtained. Here, volume cues refer to the annotation information of the user on the 3D volumetric medical image, which is used to provide prior information. Based on the current slice medical image, slice hints, and the previous slice segmentation mask, the current slice segmentation mask is obtained. The current slice medical image is obtained by slicing the current three-dimensional volumetric medical image. The slice hints refer to the annotation information of the user on the slice medical image, which is used to guide the spatial / channel attention mechanism to enhance the feature response of the target region. Extracting slice features corresponding to the target slice medical image from the current volume features and compiling the slice features into a slice feature dictionary includes: determining the slice index of the target slice medical image; extracting corresponding initial slice features from the current volume features based on the slice index; performing a corresponding convolution operation on each initial slice feature and integrating the initial slice features after the convolution operation to obtain the slice feature dictionary; wherein, the target slice medical image is determined based on the index of the current slice medical image by selecting a preset number of slice medical images adjacent to the index; Based on the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary, the target slice segmentation mask corresponding to the target slice medical image is obtained; Based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask, the hybrid propagation interactive segmentation model is iterated again until the preset convergence condition is met, thus obtaining the trained hybrid propagation interactive segmentation model.

2. The training method for the hybrid propagation interaction segmentation model as described in claim 1, characterized in that, The process of obtaining the current volume features and the current volume segmentation mask based on the current 3D volumetric medical image, volume cues, and the previous volume segmentation mask includes: The current 3D volumetric medical image, volumetric hints, and previous volumetric segmentation mask are preprocessed. The preprocessed current 3D volumetric medical image, volume prompt, and previous volume segmentation mask are encoded to obtain encoded features; The encoded features and decoded features are fused by skip connections to obtain the current volume features.

3. The training method for the hybrid propagation interaction segmentation model as described in claim 1, characterized in that, Based on the current slice medical image, slice hints, and the previous slice segmentation mask, the current slice segmentation mask is obtained, including: The current slice medical image, slice prompts, and previous slice segmentation mask are preprocessed; Based on the preprocessed current slice medical image, slice hints, and previous slice segmentation mask, multi-scale features are extracted; The multi-scale features are fused based on an attention mechanism to obtain the current slice segmentation mask.

4. The training method for the hybrid propagation interaction segmentation model as described in claim 1, characterized in that, The determination of the slice index of the target slice medical image includes: Determine the slice index of the current slice medical image; The slice index of the target slice medical image is determined based on the slice index of the current slice medical image.

5. The training method for the hybrid propagation interaction segmentation model as described in claim 1, characterized in that, The step of obtaining the target slice segmentation mask corresponding to the target slice medical image based on the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary includes: The current slice multi-scale features corresponding to the current slice medical image and the target slice multi-scale features corresponding to the target slice medical image are generated respectively. The current slice multi-scale features include current slice high-level features, current slice key features and current slice intermediate features. The target slice multi-scale features include target slice high-level features, target slice key features and target slice intermediate features. The segmentation mask of the current slice is encoded using the current slice medical image and the high-level features of the current slice to obtain a value representation; Determine the correlation between the key features of the current slice and the key features of the target slice; The slice feature dictionary, value representation, correlation, intermediate features of the current slice, and intermediate features of the target slice are integrated to obtain the target slice segmentation mask.

6. The training method for the hybrid propagation interaction segmentation model as described in any one of claims 1-5, characterized in that, After obtaining the target slice segmentation mask corresponding to the target slice medical image, the process further includes: Based on the volume hints and the current volume segmentation mask, calculate the volume segmentation loss value; Calculate the slice segmentation loss value based on the target slice segmentation mask and the current slice segmentation mask; Based on the slice feature dictionary and the target slice segmentation mask, the feature transformation loss value is calculated; Based on the volume segmentation loss value, slice segmentation loss value, and feature transformation loss value, the total loss value is obtained; If the total loss value is greater than the preset loss threshold, then the next iteration of the hybrid propagation interactive segmentation model is performed based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask.

7. A training device for a hybrid propagation interaction segmentation model, characterized in that, The apparatus, applied to a hybrid propagation interaction segmentation model, comprises: The volume propagation unit is used to obtain the current volume features and the current volume segmentation mask based on the current 3D volume medical image, volume cues, and the previous volume segmentation mask. The volume cues refer to the annotation information of the user on the 3D volume medical image, which is used to provide prior information. The slice propagation unit is used to obtain the current slice segmentation mask based on the current slice medical image, slice prompts and the previous slice segmentation mask. The current slice medical image is obtained by slicing the current three-dimensional volume medical image. The slice prompts refer to the annotation information of the user on the slice medical image, which is used to guide the spatial / channel attention mechanism to enhance the feature response of the target region. A feature compilation unit is configured to extract slice features corresponding to the target slice medical image from the current volume features and compile the slice features into a slice feature dictionary, including: determining the slice index of the target slice medical image; extracting corresponding initial slice features from the current volume features based on the slice index; performing a corresponding convolution operation on each initial slice feature and integrating the initial slice features after the convolution operation to obtain the slice feature dictionary; wherein, the target slice medical image is determined based on the index of the current slice medical image by selecting a preset number of slice medical images adjacent to the index; The target slice segmentation unit is used to obtain the target slice segmentation mask corresponding to the target slice medical image based on the current slice segmentation mask, the target slice medical image, the current slice medical image, and the slice feature dictionary; The iteration unit is used to perform the next round of iteration on the hybrid propagation interactive segmentation model based on the target slice segmentation mask, the current volume segmentation mask, and the current slice segmentation mask, until the preset convergence condition is met, and the trained hybrid propagation interactive segmentation model is obtained.

8. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, characterized in that, When the processor executes the computer-readable instructions, it implements the steps of the training method for the hybrid propagation interactive segmentation model as described in any one of claims 1 to 6.

9. A readable storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by the processor, they implement the steps of the training method for the hybrid propagation interactive segmentation model as described in any one of claims 1 to 6.

Images