Interactive three-dimensional medical image segmentation method and system, and electronic equipment

Abstract

The invention provides an interactive three-dimensional medical image segmentation method and system and electronic equipment, and the method comprises the steps: obtaining a click record of a user on a two-dimensional slice of a three-dimensional medical image, and converting the click record into three-dimensional interaction weight data distributed in a three-dimensional space; original three-dimensional medical image data and the three-dimensional interaction weight data are combined into multi-channel input data; and inputting the multi-channel input data into a three-dimensional convolutional neural network for segmentation processing to obtain a three-dimensional segmentation result. According to the method, the two-dimensional interaction operation is converted into the weight distribution in the three-dimensional space, so that direct mapping from two-dimensional interaction to accurate segmentation of the three-dimensional space is realized, and the continuity and consistency of the segmentation result in the three-dimensional space are ensured; the problems of boundary jump and space fracture caused by a traditional two-dimensional interactive segmentation method are effectively solved, and efficient and accurate three-dimensional medical image segmentation is achieved.

Description

Technical Field

[0001] This invention relates to the field of interactive three-dimensional medical image segmentation technology, specifically to an interactive three-dimensional medical image segmentation method and system, and electronic equipment. Background Technology

[0002] Medical image segmentation is a crucial foundation for clinical work such as computer-aided diagnosis, surgical navigation, and radiotherapy planning. Its goal is to accurately delineate the contours of specific organs, tissues, or lesions from three-dimensional medical imaging data such as CT and MRI. However, due to inherent problems in medical images, such as noise, artifacts, low contrast between different tissues, and blurred boundaries, fully automated segmentation methods often fail to meet clinical needs in terms of generalization and accuracy. Based on different implementation paths and objectives, medical image segmentation methods can be mainly divided into dedicated segmentation algorithms and interactive general-purpose segmentation algorithms.

[0003] Specialized segmentation algorithms are designed for fully automated segmentation of specific organs or lesions, but they suffer from poor generalization ability. For example, a model trained to segment the liver cannot be directly applied to segmenting the heart or tumors. They lack flexibility and have limited ability to handle complex situations. When encountering anatomical variations, rare pathologies, or image quality issues not fully covered in the training data, the model's performance significantly degrades, producing incorrect segmentation results, and lacks an effective mechanism for immediate correction. Therefore, while specialized segmentation algorithms are efficient for specific tasks, they struggle to meet the universal clinical need for flexible and accurate segmentation of different organs and various lesions.

[0004] Interactive general segmentation algorithms do not limit the target category for segmentation. They introduce manual guidance (such as clicking and drawing lines) to guide the segmentation process in real time, and have the potential to handle arbitrary anatomical structures. Existing technologies include interactive general segmentation based on traditional image processing (such as region growing and graph cut), whose interaction forms are relatively simple and rely on low-level image features such as grayscale and gradients. They are easily affected by image noise, resulting in segmentation results with holes, breaks, or boundary leaks. Users usually need to make a lot of fine parameter adjustments or place seed points multiple times, which is cumbersome. Existing technologies also include interactive general segmentation based on deep learning, but they are mostly limited to simple positive and negative clicks and represent interactive information as simple two-dimensional binary images. This cannot effectively carry richer and more efficient interactive intentions such as selection boxes and drawing, resulting in low efficiency of interactive information transmission. Users need to click multiple times to achieve the expected effect. At the same time, they lack consistent modeling of the three-dimensional spatial context, which leads to spatial discontinuities in the segmented three-dimensional objects, with step-like jumps at the boundaries, and the inability to generate high-quality three-dimensional models. In summary, existing technologies lack an interactive segmentation scheme that can simultaneously satisfy the requirements of flexible and efficient interaction, spatial continuity of segmentation results, low model training cost, and efficient processing of large 3D images.

[0005] Therefore, the existing technology still needs further development. Summary of the Invention

[0006] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide an interactive three-dimensional medical image segmentation method, system, and electronic device to solve the problems existing in the prior art.

[0007] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides an interactive three-dimensional medical image segmentation method, comprising: S100: Obtain the click records made by the user on the two-dimensional slices of the three-dimensional medical image, and convert the click records into three-dimensional interaction weight data distributed in three-dimensional space; S200: Combine the original three-dimensional medical image data and the three-dimensional interactive weight data into multi-channel input data; S300. The multi-channel input data is input into a three-dimensional convolutional neural network for segmentation processing to obtain a three-dimensional segmentation result.

[0008] Specifically, the three-dimensional interactive weight data marks the positive and negative segmentation regions indicated by the user in three-dimensional space.

[0009] Specifically, the click record includes at least one of point selection, box selection, drawing, and lasso operations, and each operation includes positive segmentation and negative segmentation.

[0010] Specifically, converting the click records into three-dimensional interaction weight data distributed in three-dimensional space includes: For point operations, a three-dimensional normal distribution function is generated with the three-dimensional coordinates of the user's operation point as the center, and the value of the three-dimensional normal distribution function in three-dimensional space is used as the weight value.

[0011] Specifically, the three-dimensional normal distribution function generated by the point operation is: ; in, Let be the voxel coordinates of any point in three-dimensional space. The spatial coordinates of the user's operation point in three-dimensional space. Σ represents 1 / 2 of the voxel spacing in three-dimensional space.

[0012] Specifically, converting the click records into three-dimensional interaction weight data distributed in three-dimensional space includes: For selection, drawing, or lasso operations, on the two-dimensional slice operated by the user, the pixel weight values ​​within the operated area are set as the first weight value, and the pixel weight values ​​outside the operated area are set as the second weight value.

[0013] Specifically, the multi-channel input data includes: The original 3D medical image channel, the baseline segmentation channel, and the positive and negative segmentation channels corresponding to the clicked records, respectively.

[0014] Specifically, the step of inputting the multi-channel input data into a three-dimensional convolutional neural network for segmentation processing to obtain a three-dimensional segmentation result includes: The original three-dimensional medical image channel, the baseline segmentation channel, and the positive and negative segmentation channels corresponding to the click records are input into the three-dimensional convolutional neural network. The 3D segmentation result obtained in the current step is used as the reference segmentation channel for the next segmentation process. It is then combined with the positive and negative segmentation channels generated by the user's subsequent new click records and input again into the 3D convolutional neural network to generate an updated 3D segmentation result, until a 3D segmentation result that meets the user's requirements is obtained.

[0015] Specifically, the method further includes: When training the three-dimensional convolutional neural network, a preset interactive simulation algorithm is used to simulate the interactive operations performed by the user on the two-dimensional slices of the three-dimensional medical image, and click records are generated. Then, the click records are converted into three-dimensional interactive weight data distributed in three-dimensional space.

[0016] Specifically, the method further includes: When the size of the original three-dimensional medical image data is larger than the fixed input size of the three-dimensional convolutional neural network, the original three-dimensional medical image data is segmented, and the sub-regions of the original three-dimensional medical image are filled with boundaries. The starting position of the image input to the three-dimensional convolutional neural network is dynamically adjusted to align it with the boundary of the original three-dimensional medical image.

[0017] According to a second aspect of the present invention, an interactive three-dimensional medical image segmentation system is provided, comprising: The interaction acquisition module is used to acquire the click records made by the user on the two-dimensional slices of the three-dimensional medical image, and convert the click records into three-dimensional interaction weight data distributed in three-dimensional space; The data combination module is used to combine the original three-dimensional medical image data and the three-dimensional interactive weight data into multi-channel input data; The segmentation processing module is used to input the multi-channel input data into a three-dimensional convolutional neural network for segmentation processing to obtain a three-dimensional segmentation result.

[0018] According to a third aspect of the present invention, an electronic device is provided, comprising: a memory; and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the above-described interactive three-dimensional medical image segmentation method.

[0019] Beneficial effects: This invention provides an interactive three-dimensional medical image segmentation method and system. By converting user click records on two-dimensional slices into three-dimensional spatially distributed three-dimensional interactive weight data, and combining it with the original three-dimensional image data to form a multi-channel input, the system then uses a three-dimensional convolutional neural network for segmentation processing. This achieves a direct mapping from two-dimensional interactive guidance to accurate three-dimensional spatial segmentation, ensuring the continuity and consistency of the segmentation results in three-dimensional space. It effectively solves the boundary jump and spatial breakage problems caused by traditional two-dimensional interactive segmentation methods. At the same time, through end-to-end learning of multi-channel fused data by the three-dimensional convolutional neural network, the segmentation accuracy and robustness of complex anatomical structures are significantly improved, thereby greatly improving the efficiency of medical image segmentation and realizing efficient, accurate, and universal three-dimensional medical image segmentation. Attached Figure Description

[0020] Figure 1 This is a flowchart of the interactive three-dimensional medical image segmentation method provided in a specific embodiment of the present invention; Figure 2 This is a schematic diagram of the system composition of the interactive three-dimensional medical image segmentation system provided in a specific embodiment of the present invention; Figure 3 This is an interactive segmentation diagram provided in a specific embodiment of the present invention; Figure 4 This is a flowchart of user interaction operations provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of input multi-channel data provided in a specific embodiment of the present invention; Figure 6 This is a flowchart of the training of a three-dimensional convolutional neural network provided in a specific embodiment of the present invention; Figure 7 This is a flowchart illustrating user interactive operations during convolutional neural network training, provided in a specific embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Other similar embodiments obtained by those skilled in the art based on the embodiments in this application without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0022] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0023] Example 1 Please see Figure 1 This embodiment provides an interactive three-dimensional medical image segmentation method, including the following steps: S100: Obtain the click records made by the user on the two-dimensional slices of the three-dimensional medical image, and convert the click records into three-dimensional interaction weight data distributed in three-dimensional space.

[0024] like Figure 3 As shown, in this embodiment, the click recording includes at least one of point selection, bounding box selection, drawing, and lasso operations, and each operation includes positive and negative segmentation. These point, bounding box, drawing, and lasso interactive operations are all performed on two-dimensional slices of a three-dimensional medical image. Furthermore, these operations are not limited to the axial direction and can also be performed in the coronal and sagittal planes. The three-dimensional interaction weight data is used to label the positive and negative segmentation regions indicated by the user in three-dimensional space. Based on this, the user's click recording operation flow is as follows: Figure 4 As shown, the specific operation steps are as follows: Step 1: Load DICOM medical image data. First, import the 3D medical image data to be processed. Step 2: Select the ROI region. In the coronal, sagittal, and axial directions, select a region of interest (ROI) that contains the target structure, and use its center as the reference point for segmentation. Step 3: Select any two-dimensional section. Choose one section from the coronal, sagittal, or axial directions as the current operation plane to facilitate subsequent interactive operations. Step 4: Perform interactive segmentation. On the selected slice, the user can use any of eight interactive methods to perform initial segmentation. These methods are divided into two categories: Positive segmentation (positive hints): includes point selection, box selection, drawing, and lasso, used to mark target areas; Negative segmentation (negative hints): This also includes point selection, box selection, drawing, and lasso, used to exclude non-target areas. The algorithm generates preliminary segmentation results based on user input. Step 5: Iteratively optimize the segmentation results. If the initial segmentation is not accurate enough, the user can continue to adjust it using positive or negative segmentation on different cross-sections, repeating the above interactive process until a satisfactory segmentation result is obtained.

[0025] It is understood that this embodiment supports the flexible use of various traditional interactive tools (such as brushes, erasers, etc.) in multiple planes (axial, coronal, sagittal) for precise segmentation, so as to achieve efficient and intuitive medical image segmentation.

[0026] Furthermore, in the interactive 3D medical image segmentation method of this embodiment, for point operations, a 3D normal distribution function is generated with the 3D coordinates of the user's operation point as the center. The values ​​of the 3D normal distribution function in 3D space are used as weight values. The 3D normal distribution function generated by the point operation is as follows: ; in, Let be the voxel coordinates of any point in three-dimensional space. The spatial coordinates of the user's operation point in three-dimensional space. Σ represents 1 / 2 of the voxel spacing in three-dimensional space.

[0027] Furthermore, for selection, drawing, or lasso operations, on the two-dimensional slice operated by the user, the pixel weight values ​​within the operated area are set as the first weight value, and the pixel weight values ​​outside the operated area are set as the second weight value.

[0028] Preferably, for bounding box selection operations, the pixel weight within the selected area can be set to 1, and the pixel weight outside the selected area can be set to 0. For drawing operations, the pixel weight within the drawing area can be set to 1, and the pixel weight outside the drawing area can be set to 0. For lasso operations, the pixel weight within the lasso area can be set to 1, and the pixel weight outside the lasso area can be set to 0. In practical applications, this weight setting method can clearly distinguish between the operated area and the unoperated area, making subsequent segmentation processing more accurate and efficient.

[0029] S200 combines the original 3D medical image data and 3D interactive weight data into multi-channel input data.

[0030] like Figure 5 As shown, in this embodiment, the multi-channel input data includes: the original three-dimensional medical image channel, the baseline segmentation channel, and the positive segmentation channel and negative segmentation channel corresponding to the click record, respectively.

[0031] Understandably, the raw 3D medical image channel contains unprocessed medical image information, preserving the original features of the image and providing the basic data source for subsequent segmentation processing. The baseline segmentation channel, on the other hand, is generated based on existing segmentation results or preset segmentation criteria. It provides a reference benchmark for the entire segmentation process, helping to more accurately determine the segmentation boundaries and range. The baseline segmentation channel is a 3D matrix of the same size as the raw 3D medical image. The values ​​in the matrix represent the system's latest understanding of whether each voxel belongs to the target segmentation region before the current segmentation iteration. When the user performs the first interaction, there are no historical segmentation results yet. At this time, the baseline segmentation channel can be a matrix full of zeros, indicating an initial state of no segmentation, or it can be an initial coarse segmentation result generated by a simple algorithm (such as thresholding). After the first interaction, the current segmentation result can be... The 3D segmentation result generated by the 3D convolutional neural network in step N is directly used as the baseline segmentation channel for the next segmentation process (N+1). This makes the segmentation process a dynamic and iteratively evolving system. The baseline segmentation channel uses the previous segmentation result as spatial prior knowledge input into the network, allowing the network to clearly understand which regions have already been segmented, thus avoiding redundant calculations and focusing attention on the newly added or to-be-corrected regions indicated by the current interaction. The positive and negative segmentation channels correspond to positive and negative clicks in the click records, respectively. The positive segmentation channel records the region information that the user expects to be included in the segmentation result, while the negative segmentation channel records the region information that the user expects to be excluded from the segmentation result. By combining the data from these four channels to form multi-channel input data, the information carried by different types of data can be fully utilized, providing richer and more comprehensive input for subsequent segmentation algorithms, thereby improving the accuracy and reliability of segmentation.

[0032] S300: Input the multi-channel input data into a three-dimensional convolutional neural network for segmentation processing to obtain the three-dimensional segmentation result.

[0033] like Figure 5 As shown, the 3D convolutional neural network includes an Encoder and a Decoder structure. The original 3D medical image channel, the baseline segmentation channel, and the positive and negative segmentation weight channels corresponding to the click records are input into the 3D convolutional neural network. The 3D segmentation result obtained in the current step is used as the baseline segmentation channel for the next segmentation process. It is then combined with the positive and negative segmentation channels generated by the new click records subsequently input by the user and input into the 3D convolutional neural network again to generate an updated 3D segmentation result, until a 3D segmentation result that meets the user's requirements is obtained.

[0034] It's important to note that in the processing of a 3D convolutional neural network, the Encoder structure is responsible for feature extraction and dimensionality reduction of the input multi-channel data. Through a series of convolutions, pooling, and other operations, it gradually captures key feature information in the image, converting the original high-dimensional data into a low-dimensional feature representation. The Decoder structure, on the other hand, performs upsampling and deconvolution operations on the low-dimensional features output by the Encoder, gradually restoring the data's dimensionality and ultimately generating a 3D segmentation result close to the size of the original image. This iterative segmentation method fully utilizes the results of each segmentation step, continuously optimizing the segmentation boundaries and range. As iterations proceed, the 3D segmentation result becomes increasingly closer to the user's actual needs, effectively improving segmentation accuracy and efficiency. Furthermore, the powerful feature learning and processing capabilities of the 3D convolutional neural network provide strong technical support for the entire interactive 3D medical image segmentation method.

[0035] See Figure 6 In this embodiment, the steps for training using a three-dimensional convolutional neural network are as follows: Step 1: Data Preparation Collect raw 3D medical image data (such as CT or MRI scan sequences in DICOM format). The data should include multiple patient samples, covering different lesion degrees and anatomical structures, to ensure the model's generalization ability. Step 2: Data Cleaning Preprocess the raw data to remove noise (e.g., using Gaussian filtering), standardize intensity values ​​(e.g., normalize to [0,1] or Z-score standardization), fill in missing slices or irregular volumes, unify image size (e.g., crop or interpolate to the same resolution), handle annotation errors or blurred boundaries, and ensure consistent input data quality to avoid affecting model training performance.

[0036] Step 3: Construct the training dataset The cleaned images are integrated with their corresponding labels to form training data, which is usually stored in formats such as NIfTI, HDF5, or TensorFlow Dataset for easy and efficient reading later. The dataset is divided proportionally into a training set, a validation set (optional), and a test set; Step 4: Random initialization of the neural network The core of a three-dimensional convolutional neural network architecture (such as 3D U-Net, V-Net, etc.) consists of an Encoder and a Decoder. Initialize network parameters (weights and biases) to ensure stable gradient propagation in the early stages of training; Step 5: Random sampling of data In each training iteration (epoch), a batch of 3D images and their labels are randomly selected from the training dataset; Each sample is typically a three-dimensional volume block, such as 64×64×64 voxels, to reduce GPU memory usage and improve training efficiency; data augmentation (such as rotation, flipping, elastic deformation) is used to increase sample diversity and prevent overfitting. Step 6: Interactive Simulation (Core Training Phase) The sampled 3D medical images are acquired, and interactive simulation operations are performed on 2D slices of the 3D medical images. These interactive operations can include one of the following: point selection, bounding box selection, drawing, and lasso operations. Click records are obtained and converted into 3D interactive weight data distributed in 3D space. This data is then input into an initialized 3D convolutional neural network (3DCNN) model to perform the following process: (1) Forward propagation The image is processed through multi-layer 3D convolution and pooling operations to extract high-level features in the Encoder. The feature map is then upsampled by the Decoder and fused with skip connections from the Encoder to restore spatial details. The output is a predicted segmentation map (or classification result) of the same size as the input. (2) Calculate the loss function The loss function used in this embodiment is a combination of cross-entropy loss and Dice loss. The specific calculation method is as follows: ; ; ; in, Represents cross-entropy loss, Represents the Dice loss function. y The true pixel label is represented by 0 or 1, p is represented by the predicted probability label, which is 0 to 1, i=[0, 1], N is represented by the total number of pixels in the image, and C is represented by the total number of categories. In this embodiment, C=2. (3) Backpropagation Based on the gradient of the loss function, the network parameters are updated in reverse using an optimization algorithm to minimize the loss; Step 7: Iteratively update parameters In step 6 above, the forward propagation-loss calculation-backward propagation process is repeated continuously to form a training loop. After each batch is completed, the model parameters are updated step by step, and training continues for multiple epochs until the stopping condition is met (such as loss convergence or validation set performance no longer improving). Step 8: Save the model Once training is complete, the final trained model parameters are saved as a file. The saved model can be used for subsequent inference tasks, such as automatic segmentation or diagnosis of 3D medical images of new patients.

[0037] It is understood that, in this embodiment, the training process of the 3D convolutional neural network is a closed-loop iterative process. Starting with high-quality data preparation, it proceeds through cleaning, modeling, random sampling, interactive forward and backward propagation, ultimately achieving continuous optimization of model parameters and saving a usable model. This process is particularly suitable for automated segmentation, detection, or classification tasks of 3D medical images and has good potential for clinical applications.

[0038] See Figure 7 In some specific embodiments, when training a three-dimensional convolutional neural network, a preset interaction simulation algorithm is used to simulate user interactions on two-dimensional slices of a three-dimensional medical image, and click records are generated. These click records are then converted into three-dimensional interaction weight data distributed in three-dimensional space. The specific implementation process is as follows: Step 1: Input Data Preparation Input the original 3D medical image data as the basic training data, and input the current segmentation prediction result of the 3D convolutional neural network, that is, the initial segmentation result of the first phase, as the baseline segmentation result. Step 2: Error Region Calculation Based on the difference between the segmentation results predicted by the 3D convolutional neural network and the actual annotations, the positive and negative error regions are calculated. (1) Positive error region: The region that exists in the actual annotation but is missing in the prediction result, i.e. the missing segmentation region, which simulates the positive sample operation that the user needs to add; (2) Negative error region: The region that exists in the prediction result but does not exist in the actual annotation, i.e., the oversegmented region, which simulates the negative sample operation that the user needs to delete; Step 3: Connected Component Filtering From the calculated error region, select the connected regions that satisfy the following conditions: (1) Volume threshold condition: The volume of the connected region must be greater than 8 voxels to ensure that the simulation operation has sufficient spatial significance; (2) Spatial distribution conditions: Spatially discrete connected regions are preferred to simulate the correction process of real users for multiple independent error regions; Step 4: Slice Selection Within the selected connected region, select operation slices based on the following criteria: (1) Center position reference: The center of the volume of the connected region is used as the reference point; (2) Position deviation range: On three orthogonal directions (x, y, z axes), slices within ±20% of the volume center are selected to simulate the user's operating habits from different perspectives; Step 5: Interactive Operation Simulation On the selected two-dimensional slice, one of the aforementioned four interactive operations (point selection, box selection, drawing, lasso) is randomly selected for simulation. Random selection ensures that the model can learn to respond to various interactive methods and determine the specific operation position and range within the error region. Step 6: Adjusting Operation Weights A dynamic weight adjustment strategy is adopted to reduce the probability of selecting interaction types that have been used in the current simulation by 10%. This mechanism ensures that in subsequent training rounds, the system will tend to select interaction types that have been used less frequently before. Step 7: Neural Network Input Construction Based on the current image data, segmentation prediction results, and simulated interactive operations, the input of the neural network is constructed: the simulated operations are converted into corresponding three-dimensional interactive weight data, and the original image, current segmentation prediction, positive sample weight channel, and negative sample weight channel are combined into 10-channel input data. Step 8: Network parameter update The constructed multi-channel input data is fed into a 3D convolutional neural network. The forward propagation yields a new segmentation prediction. The combined loss between the prediction result and the ground label (cross-entropy loss + Dice loss) is calculated. The network parameters are then updated through the backpropagation algorithm. Step 9: Iterative Optimization The updated segmentation prediction results are used as input for the next round of training. Then, steps 2 to 8 are repeated for the next round of interactive simulation and parameter update. Step 10: Training Termination Conditions The current training round will be terminated if one of the following conditions is met: (1) The predicted segmentation result and the Dice coefficient of the actual annotation reach a predetermined threshold (e.g., 0.95). (2) The decrease in the loss function over multiple iterations is less than the set value; (3) Reach the maximum number of iterations.

[0039] By adopting the above technical solution, this method not only improves the model's responsiveness to user prompts, but also provides an efficient training paradigm for developing interactive medical image segmentation systems.

[0040] In some specific embodiments, such as Figure 5As shown, when the size of the original 3D medical image data is larger than the fixed input size of the 3D convolutional neural network, the original 3D medical image data is segmented, and the boundaries of the sub-regions of the original 3D medical image are filled. The starting position of the image input to the 3D convolutional neural network is dynamically adjusted to align it with the boundary of the original 3D medical image. In this way, it is ensured that the 3D convolutional neural network can process input data of arbitrary size, while preserving the boundary information of the image to the maximum extent and avoiding the information loss caused by traditional cropping methods.

[0041] It should be noted that this embodiment provides an interactive three-dimensional medical image segmentation method. Through the above-mentioned interactive three-dimensional medical image segmentation method, users can perform simple interactive operations on two-dimensional slices to achieve accurate segmentation of three-dimensional medical images, thereby improving the efficiency and accuracy of medical image analysis.

[0042] Example 2 Please see Figure 2 This embodiment provides an interactive three-dimensional medical image segmentation system, including an interactive acquisition module 100, a data combination module 200, and a segmentation processing module 300.

[0043] The interaction acquisition module 100 is used to acquire click records made by the user on two-dimensional slices of a three-dimensional medical image, and convert these click records into three-dimensional interaction weight data distributed in three-dimensional space. Specifically, as shown... Figure 3 As shown, the interaction acquisition module 100 acquires user actions on a two-dimensional slice using different interactive methods, including clicking, selecting, drawing, or lassoing. Users can use these methods to mark positive sample areas (i.e., target areas) and negative sample areas (i.e., background areas). When a user clicks or performs other interactive operations on the two-dimensional slice, the interaction acquisition module 100 records the positional information of these operations and maps this two-dimensional positional information to three-dimensional space, generating corresponding three-dimensional interaction weight data. The three-dimensional interaction weight data includes the spatial coordinates of each interaction point and its corresponding weight value. The weight value for positive sample areas is positive, and the weight value for negative sample areas is negative. The magnitude of the weight value can be adjusted according to the intensity or determinism of the user interaction.

[0044] The data combination module 200 is used to combine raw 3D medical image data and 3D interactive weight data into multi-channel input data. For example... Figure 5As shown, the multi-channel input data includes the original 3D medical image channel, the baseline segmentation channel, and the positive and negative segmentation channels. The original 3D medical image channel contains the original 3D medical image data, such as 3D voxel data acquired by imaging equipment such as CT and MRI; the baseline segmentation channel contains the initial segmentation result or the segmentation result of the previous iteration; the positive and negative segmentation channels contain 3D interactive weight data generated by user interaction, where the positive segmentation channel corresponds to the target region marked by the user, and the negative segmentation channel corresponds to the background region marked by the user. These channels are combined into a unified multi-channel input data structure for subsequent neural network processing.

[0045] The segmentation processing module 300 is used to input multi-channel input data into a three-dimensional convolutional neural network for segmentation processing to obtain three-dimensional segmentation results. For example... Figure 5 As shown, the 3D convolutional neural network includes an Encoder and a Decoder. The Encoder extracts features from the input data through multi-layer 3D convolution and pooling operations. The Decoder restores the spatial resolution of the features and generates the final segmentation result through upsampling and deconvolution operations. The segmentation result is a probability map with the same size as the input 3D medical image, representing the probability that each voxel belongs to the target region. By setting a threshold (usually 0.5), the probability map can be converted into a binary segmentation mask, where regions larger than the threshold are identified as target regions, and regions smaller than the threshold are identified as background regions.

[0046] In a preferred embodiment, the interactive acquisition module 100 further includes an interactive interface submodule, used to provide a graphical interface for user interaction, display two-dimensional slices of three-dimensional medical images, and receive user interaction operations, such as... Figure 4 As shown, the interactive interface submodule provides functions for loading DICOM images, selecting regions of interest (ROI), selecting 2D cross-sections, and performing interactive operations. Users can perform a series of operations through these functions until they obtain accurate segmentation results. The interactive interface submodule supports a variety of interaction methods, including but not limited to clicking, selecting boxes, drawing, and lassoing, enabling users to flexibly mark target and background areas.

[0047] In another preferred embodiment, the interaction acquisition module 100 further includes a weight conversion submodule, which is used to convert two-dimensional interaction operations into three-dimensional interaction weight data. The weight conversion submodule adopts a Gaussian decay model to expand the user's interaction points on the two-dimensional slice into a Gaussian distribution in three-dimensional space, so that the area around the interaction point also has a certain weight value. The weight value decays as the distance from the interaction point increases. This method can enhance the robustness of the interaction and reduce the accuracy requirements of the segmentation system on user interaction.

[0048] In another preferred embodiment, the data combination module 200 further includes a data preprocessing submodule for preprocessing the original three-dimensional medical image data, including operations such as intensity normalization, noise filtering, and contrast enhancement. Intensity normalization standardizes medical image data acquired from different devices and with different parameters to the same numerical range, typically [0,1] or [-1,1]. Noise filtering reduces noise in the image using methods such as Gaussian filtering or median filtering. Contrast enhancement improves the image contrast using methods such as histogram equalization or adaptive histogram equalization, making the target area and background area more clearly distinguishable.

[0049] In a preferred embodiment, the 3D convolutional neural network in the segmentation processing module 300 adopts a U-Net structure, including a downsampling path and an upsampling path, as well as skip connections between corresponding layers. The downsampling path gradually reduces the spatial resolution of the feature map through successive convolution and pooling operations, while increasing the number of channels, to extract high-level semantic features of the image; the upsampling path gradually restores the spatial resolution of the feature map through successive upsampling and convolution operations, while reducing the number of channels, to generate fine segmentation results; the skip connections directly connect the feature maps in the downsampling path to the corresponding layers in the upsampling path to preserve detailed information and improve segmentation accuracy.

[0050] In another preferred embodiment, the training process of the three-dimensional convolutional neural network is as follows: Figure 6 As shown, the process includes steps such as data preparation, interactive simulation, model training, and parameter updating. The data preparation phase collects annotated 3D medical image datasets; the interactive simulation phase generates simulated user interaction data based on real annotations; the model training phase trains the network using data with simulated interactions; and the parameter updating phase updates the network parameters based on the training results. Training employs a combination of cross-entropy loss and Dice loss functions to balance class imbalance and improve segmentation accuracy.

[0051] It should be noted that this embodiment provides an interactive three-dimensional medical image segmentation system, including an interactive acquisition module 100, a data combination module 200, and a segmentation processing module 300. By converting user click records on two-dimensional slices into three-dimensional spatially distributed three-dimensional interactive weight data, and combining it with the original three-dimensional image data into a multi-channel input, the system is then segmented by a three-dimensional convolutional neural network. This achieves a direct mapping from two-dimensional interactive guidance to accurate three-dimensional spatial segmentation, ensuring the continuity and consistency of the segmentation results in three-dimensional space. It effectively solves the boundary jump and spatial breakage problems caused by traditional two-dimensional interactive segmentation methods, significantly improves the segmentation accuracy and robustness of complex anatomical structures, and thus greatly improves the efficiency of medical image segmentation, achieving efficient and accurate universal three-dimensional medical image segmentation.

[0052] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor. The memory stores computer-readable instructions that, when executed by the processor, implement the interactive three-dimensional medical image segmentation method described in Embodiment 1. The method includes acquiring user click records on two-dimensional slices of a three-dimensional medical image and converting them into three-dimensional interactive weight data; combining the original three-dimensional medical image data and the three-dimensional interactive weight data into multi-channel input data; and inputting the multi-channel input data into a three-dimensional convolutional neural network for segmentation processing to obtain a three-dimensional segmentation result. This computer device can be broadly considered as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0053] Furthermore, the electronic device in this embodiment may also be equipped with a display screen for displaying two-dimensional slices and segmentation results of the three-dimensional medical image, as well as input devices such as a mouse, stylus, or touchscreen for receiving user interaction. By executing stored computer-readable instructions, the electronic device can respond to user interaction in real time and continuously optimize the segmentation results of the three-dimensional medical image based on the user's click records, thereby improving the efficiency and accuracy of medical image analysis.

[0054] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0055] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0056] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. An interactive three-dimensional medical image segmentation method, characterized in that, include: S100: Obtain the click records made by the user on the two-dimensional slices of the three-dimensional medical image, and convert the click records into three-dimensional interaction weight data distributed in three-dimensional space; S200: Combine the original three-dimensional medical image data and the three-dimensional interactive weight data into multi-channel input data; S300. The multi-channel input data is input into a three-dimensional convolutional neural network for segmentation processing to obtain a three-dimensional segmentation result.

2. The interactive three-dimensional medical image segmentation method according to claim 1, characterized in that, The three-dimensional interactive weight data is used to define the positive and negative segmentation regions indicated by the user in three-dimensional space.

3. The interactive three-dimensional medical image segmentation method according to claim 1, characterized in that, The click records include at least one of point selection, box selection, drawing, and lasso operations, and each operation includes positive segmentation and negative segmentation.

4. The interactive three-dimensional medical image segmentation method according to claim 3, characterized in that, The step of converting the click records into three-dimensional interaction weight data distributed in three-dimensional space includes: For point operations, a three-dimensional normal distribution function is generated with the three-dimensional coordinates of the user's operation point as the center, and the value of the three-dimensional normal distribution function in three-dimensional space is used as the weight value.

5. The interactive three-dimensional medical image segmentation method according to claim 4, characterized in that, The three-dimensional normal distribution function generated by the point operation is: ； in, Let be the voxel coordinates of any point in three-dimensional space. The spatial coordinates of the user's operation point in three-dimensional space. Σ represents 1 / 2 of the voxel spacing in three-dimensional space.

6. The interactive three-dimensional medical image segmentation method according to claim 3, characterized in that, The step of converting the click records into three-dimensional interaction weight data distributed in three-dimensional space includes: For selection, drawing, or lasso operations, on the two-dimensional slice operated by the user, the pixel weight values ​​within the operated area are set as the first weight value, and the pixel weight values ​​outside the operated area are set as the second weight value.

7. The interactive three-dimensional medical image segmentation method according to claim 2, characterized in that, The multi-channel input data includes: The original 3D medical image channel, the baseline segmentation channel, and the positive and negative segmentation channels corresponding to the clicked records, respectively.

8. The interactive three-dimensional medical image segmentation method according to claim 7, characterized in that, The step of inputting the multi-channel input data into a three-dimensional convolutional neural network for segmentation processing to obtain a three-dimensional segmentation result includes: The original three-dimensional medical image channel, the baseline segmentation channel, and the positive and negative segmentation channels corresponding to the click records are input into the three-dimensional convolutional neural network. The 3D segmentation result obtained in the current step is used as the reference segmentation channel for the next segmentation process. It is then combined with the positive and negative segmentation channels generated by the user's subsequent new click records and input again into the 3D convolutional neural network to generate an updated 3D segmentation result, until a 3D segmentation result that meets the user's requirements is obtained.

9. The interactive three-dimensional medical image segmentation method according to claim 1, characterized in that, The method further includes: When training the three-dimensional convolutional neural network, a preset interactive simulation algorithm is used to simulate the interactive operations performed by the user on the two-dimensional slices of the three-dimensional medical image, and click records are generated. Then, the click records are converted into three-dimensional interactive weight data distributed in three-dimensional space.

10. The interactive three-dimensional medical image segmentation method according to claim 1, characterized in that, The method further includes: When the size of the original three-dimensional medical image data is larger than the fixed input size of the three-dimensional convolutional neural network, the original three-dimensional medical image data is segmented, and the sub-regions of the original three-dimensional medical image are filled with boundaries. The starting position of the image input to the three-dimensional convolutional neural network is dynamically adjusted to align it with the boundary of the original three-dimensional medical image.

11. An interactive three-dimensional medical image segmentation system, characterized in that, include: The interaction acquisition module is used to acquire the click records made by the user on the two-dimensional slices of the three-dimensional medical image, and convert the click records into three-dimensional interaction weight data distributed in three-dimensional space; The data combination module is used to combine the original three-dimensional medical image data and the three-dimensional interactive weight data into multi-channel input data; The segmentation processing module is used to input the multi-channel input data into a three-dimensional convolutional neural network for segmentation processing to obtain a three-dimensional segmentation result.

12. An electronic device, characterized in that, include: Memory; The processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the interactive three-dimensional medical image segmentation method according to any one of claims 1 to 10.

Images