An AI-based automatic segmentation method for 3D images of mitral valve structure

Abstract

An AI-based automatic 3D image segmentation method for mitral valve structure includes the following steps: First, the original 3D dataset of the heart structure is acquired, labeled, and normalized and segmented. Then, the image data is denoised and enhanced to expand the training data and improve the model's robustness. Subsequently, a 3D encoder-decoder architecture is constructed, combining a multi-scale feature fusion module and a strategy of progressively increasing the kernel size to optimize image detail preservation and the recognition of subtle mitral valve structures. In model design, a composite scaling strategy is used to adjust the model's depth, width, and kernel size. A composite loss function is introduced during model training. Finally, a deep supervision mechanism is introduced at different decoding stages, and the model's performance on multiple datasets is evaluated using five-fold cross-validation. This invention achieves high-precision and high-efficiency automatic 3D image segmentation of the mitral valve structure, possessing significant clinical application value.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an automatic segmentation method for 3D images of the mitral valve structure based on artificial intelligence. Background Technology

[0002] The mitral valve is a vital structure in the heart, and its function is essential for maintaining a normal cardiovascular system. Accurate 3D image segmentation of the mitral valve structure is crucial for diagnosing mitral valve disease, assessing the severity of lesions, and developing treatment plans. However, due to the complexity and diversity of the mitral valve structure, traditional manual or semi-automatic segmentation methods are time-consuming and labor-intensive, and easily affected by subjective factors, making it difficult to guarantee the consistency and repeatability of segmentation results.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially deep learning, in the field of medical image processing, automated image segmentation technology has made significant progress. However, in the 3D image segmentation of the fine structure of the heart, existing technologies still face many challenges, especially in the segmentation of mitral valve images, where the following problems exist:

[0004] 1. Structural complexity: The mitral valve is composed of complex structures such as the anterior leaflet, posterior leaflet, and chordae tendineae. These structures often appear as thin-walled, small features in 3D images, making them difficult to capture accurately.

[0005] 2. Variation of lesions: Mitral stenosis, regurgitation, prolapse and other lesions can cause significant changes in valve morphology, increasing the difficulty of segmentation.

[0006] 3. Unstable image quality: Heartbeat, patient breathing, and differences in different imaging devices can lead to unstable image quality, affecting the accuracy of segmentation.

[0007] 4. Data scarcity: The difficulty in obtaining high-quality labeled data limits the training effect of deep learning models.

[0008] 5. Model generalization ability: Existing models often perform well on specific datasets, but their generalization ability is insufficient when faced with new and unseen data.

[0009] 6. Computational efficiency: 3D image segmentation requires processing a large amount of data, which places high demands on computing resources and affects the real-time performance and clinical application of the algorithm.

[0010] Existing methods, such as edge detection and region growing algorithms based on traditional image processing techniques, often perform poorly when dealing with complex mitral valve structures. While deep learning-based methods, such as U-Net and its variants, have achieved good results in 2D medical image segmentation tasks and perform reasonably well in processing 3D data of large organs, they often exhibit problems such as insufficient information utilization, insufficient sensitivity to small targets, and decreased robustness to multiple morphologies of the same target when processing the fine structures of small organs.

[0011] Therefore, there is a need for an automatic segmentation method that can fully utilize 3D spatial information, adapt to the structural characteristics of the mitral valve, and possess good generalization ability. To address the problems of existing technologies, this invention proposes an efficient and accurate automatic 3D image segmentation scheme for the mitral valve structure, incorporating the latest deep learning techniques. This scheme combines the advantages of mechanisms such as Transformer attention and ConvNeXt, and optimizes them for the specific characteristics of the mitral valve structure, aiming to provide more reliable imaging evidence for the diagnosis and treatment of mitral valve-related diseases. Summary of the Invention

[0012] This invention proposes an automatic segmentation method for 3D images of the mitral valve structure based on artificial intelligence. The specific steps are as follows:

[0013] Step 1: Obtain the original 3D dataset of the heart structure and obtain labels through manual or semi-automatic annotation; after normalization, the original 3D dataset is uniformly segmented to ensure that all blocks are of the same size, generating an image dataset and label dataset suitable for 3D segmentation;

[0014] Step 2: Preprocess the mitral valve image data, including image denoising and enhancement; expand the training data volume by enhancing the data to improve the robustness of the model. The enhancement process includes rotation, translation, scaling and flipping.

[0015] Step 3: After image data preprocessing, a 3D encoder-decoder architecture is constructed using 3D ConvNeXt; upsampling and downsampling are performed using convolutional blocks with inverse bottleneck structures to effectively preserve image details at different scales and enhance the ability to recognize the fine structure of the mitral lobe; at the same time, a multi-scale feature fusion module is introduced to effectively integrate features at different scales and improve the model's ability to recognize target objects.

[0016] Step 4: Use a method of gradually increasing kernel size to optimize the transition from small kernel to large kernel; generate the initial weights of large convolutional kernels by performing trilinear interpolation on small convolutional kernels to overcome the performance saturation problem that may occur when training large models on small datasets.

[0017] Step 5: In the model design, a composite scaling strategy is adopted to comprehensively adjust the depth, width, and kernel size of the model; the parameters are optimized according to the specific structure and complexity of the mitral lobe, thereby achieving high-precision image segmentation;

[0018] Step 6: During model training, a composite loss function, DC_CE_FOCAL_CLS_ANA_loss, is introduced. This function combines Soft Dice loss, cross-entropy loss, Focal loss, CLS loss, and a custom anatomical-based mitral valve category loss to improve segmentation accuracy and robustness.

[0019] DC_CE_FOCAL_CLS_loss=α·Soft Dice Loss+β·Cross Entropy Loss+γ

[0020] ·Focal Loss+ε·Classification Loss+θ·Anatomy Loss

[0021] Where α, β, γ, ε, and θ are weighting coefficients.

[0022] Step 7: After designing the loss function, use mixed precision training and gradient checkpointing techniques to reduce the memory consumption of the graphics processing unit (GPU) during training; use the AdamW optimizer, combined with the above composite loss function, for supervised training to improve the segmentation accuracy; the training process includes an initial training phase, an intermediate training phase, and a fine training phase, gradually adjusting the learning rate and data augmentation strategies to optimize model performance.

[0023] Step 8: Introduce a deep supervision mechanism at different decoding stages of the model, using multi-scale labeled images for supervision and calculating the loss;

[0024] Step 9: Based on the loss value calculated in Step 6, after training in Step 8 using the backpropagation algorithm and the AdamW optimizer set in Step 7, perform five-fold cross-validation on multiple datasets to evaluate the model's performance on different tasks and datasets; verify the model's segmentation effect by calculating the Dice similarity coefficient and Hausdorff distance, and evaluate the model's performance on the public test set to verify the model's generalization ability and robustness.

[0025] Furthermore, according to an artificial intelligence-based 3D image automatic segmentation method for mitral valve structure of the present invention, in step 1, both the image data and the label data contain three-dimensional image information. Considering the complex three-dimensional structure of the mitral valve, we adopt a patch-based strategy to segment the 3D data into partially overlapping small blocks to ensure that the model can capture local and global information. To ensure that all blocks are of the same size, appropriate overlap padding is performed when the data cannot be divided by the specified block size. This method is particularly suitable for processing delicate structures such as the mitral valve leaflets and chordae tendineae.

[0026] According to the present invention, an automatic segmentation method for 3D images of mitral valve structure based on artificial intelligence, step 3 specifically includes the following steps:

[0027] Step 3.1: Select a large-size convolutional kernel and apply strided convolution to extract high-level features containing long-range spatial dependencies;

[0028] Step 3.2: Use deconvolution layers to upsample the feature map layer by layer or use interpolation methods such as bilinear interpolation to perform upsampling operations, and use pooling layers or stride convolution layers to compress features to perform downsampling operations.

[0029] Step 3.3: Use weighted averaging or cascading operation schemes to fuse feature maps of different levels and scales; use a multi-scale feature fusion module to capture local details and global semantic information.

[0030] According to the present invention, an automatic segmentation method for 3D images of mitral valve structure based on artificial intelligence, step 5 specifically includes the following steps:

[0031] Step 5.1: Adjust the model depth by increasing or decreasing the number of layers;

[0032] Step 5.2: Adjust the number of convolutional kernels in each layer to change the width of the model;

[0033] Step 5.3: Adjust the size of the convolution kernel according to the specific task requirements to optimize the receptive field and computational efficiency of the model.

[0034] According to the present invention, an automatic segmentation method for 3D images of mitral valve structure based on artificial intelligence, step 6 further includes the following steps:

[0035] Step 6.1: Use Soft Dice loss to measure the overlap between the segmentation result and the label, which is suitable for handling class imbalance problems.

[0036]

[0037] Where, p i It is the predicted probability, gi It is the actual label, and ∈ is a very small constant used to avoid the denominator being zero;

[0038] Step 6.2: Apply cross-entropy loss for pixel-level classification, and calculate the difference between the segmentation result and the label.

[0039]

[0040] Where C is the number of categories. It is the predicted probability of category c. It is an indicator function for whether the actual label belongs to category c;

[0041] Step 6.3: Apply Focal loss to enhance the weights of hard-to-classify samples, mitigating the impact of class imbalance on model training.

[0042]

[0043] Where γ is an adjustment factor used to control the weights of easy and difficult samples;

[0044] Step 6.4: Introduce a classification loss for specific mitral lobe structures to improve the recognition accuracy of these structures. The classification loss adjusts the weights of each category based on their features to achieve more accurate classification results.

[0045]

[0046] Where, p i It is the predicted probability, g i It's a tag;

[0047] Step 6.5: Introduce an anatomical-based constraint term, Anatomy Loss, into the loss function to further improve the accuracy and anatomical plausibility of the segmentation results.

[0048]

[0049] Where P i and A j represents the segmentation probability values ​​of the anterior and posterior mitral lobes at positions i and j, respectively;

[0050] Step 6.6: Adjust the weights of each loss term according to different tasks and datasets to achieve the optimal segmentation effect.

[0051] According to the present invention, an automatic segmentation method for 3D images of mitral valve structure based on artificial intelligence, step 8 further includes the following steps:

[0052] Step 8.1: Enlarge or reduce the size of the tag image to match the size of the feature map in the corresponding decoding stage. Ideally, bilinear interpolation should be used for scaling.

[0053]

[0054] Step 8.2: Convert the label image into a multi-channel image, with each target structure occupying one channel;

[0055] One-hot(g,C)=[g==c1,g==c2,...,g==C C ]

[0056] Where g is the label image, C is the number of categories, and C i For class i;

[0057] Step 8.3: Add noise to the label images of different channels, including Gaussian noise, Poisson noise or salt and pepper noise, and perform dilation and erosion operations on the label images after adding noise to obtain morphological gradients;

[0058] Step 8.4: Perform convolution operations on the feature images of each decoding stage to output multi-channel images, and calculate the loss by combining them with the morphological gradient of the label image; sum the loss values ​​of different decoding stages to obtain the overall loss value for deep supervision.

[0059] Losss d =LossFunc(Conv(x) d ), Morphological Gradient (g d ))

[0060] Where, x d For the feature image of the d-th decoding stage, g d For the corresponding label image, LossFunc is the loss function in step 6.

[0061] The beneficial effects of this invention are:

[0062] This invention proposes an AI-based automatic 3D image segmentation method for mitral valve structures. Through innovative network architecture design and training strategies, it achieves significant technical effects and clinical value. Combining the local feature capture capabilities of CNNs and the global contextual understanding capabilities of Transformers, it significantly improves the segmentation accuracy of mitral valve structures, especially fine structures such as the leaflets and chordae tendineae. The large convolutional kernel design and multi-scale deep supervision mechanism increase the receptive field and improve adaptability to images of different resolutions and qualities. The progressive learning strategy enables the model to better handle morphological changes caused by mitral valve lesions, enhancing the algorithm's robustness.

[0063] The designed composite loss function, especially the anatomical-based constraint term, ensures that the segmentation results conform to the anatomical characteristics of the mitral valve, improving the model's generalization ability on different datasets. Borrowing the adaptive design concept of nnUNet, the model can automatically adjust its network structure according to the characteristics of the input data, further enhancing its applicability in different clinical scenarios. The optimized network architecture and training strategy, along with the patch-based strategy and mixed-precision training technique, significantly improve segmentation efficiency and reduce memory consumption while maintaining high accuracy, making real-time clinical applications possible.

[0064] This method provides more reliable imaging evidence for the accurate diagnosis, disease assessment, and surgical planning of mitral valve-related diseases. The increased automation significantly reduces the workload of manual annotation, improving clinical efficiency. High-precision 3D segmentation results provide a foundation for patient-specific cardiac function assessment and treatment planning, promoting the development of personalized medicine in cardiology. The innovative combination of CNN and Transformer network architectures, progressive learning strategies, and morphological loss provides new research ideas for medical image segmentation. High-quality segmentation results provide strong support for research on mitral valve pathological mechanisms and the development of novel interventional treatment devices. In summary, the automatic 3D image segmentation method for mitral valve structure proposed in this invention not only achieves significant technological breakthroughs, improving segmentation accuracy, efficiency, and robustness, but also has important clinical value, and is expected to bring significant progress to the diagnosis, treatment, and research of mitral valve-related diseases. Attached Figure Description

[0065] Figure 1 This is a flowchart of the artificial intelligence-based automatic segmentation method for 3D images of the mitral valve structure according to the present invention;

[0066] Figure 2 This is a schematic diagram of the data processing and acquisition block described in step 1 of the present invention;

[0067] Figure 3 This is a schematic diagram illustrating the capture of long-distance semantics described in steps 3, 4, and 5 of the present invention;

[0068] Figure 4 This is a schematic diagram of the loss function logic in step 6.5 of the present invention;

[0069] Figure 5a This is a schematic diagram of the label described in this invention;

[0070] Figure 5b This is a schematic diagram of the 2D cross-sectional gradient of the mitral valve described in this invention;

[0071] Figure 5cThis is a schematic diagram of the gradient of the mitral valve 3D structure described in this invention. Detailed Implementation

[0072] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that the embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Furthermore, it should be understood that after reading the disclosure of this invention, those skilled in the art can make various modifications or alterations to the invention, and these equivalent forms also fall within the scope of protection defined by this invention.

[0073] This invention aims to address the shortcomings of current fine-structure segmentation models for the heart by proposing an artificial intelligence-based automatic 3D image segmentation method for the mitral valve structure. A flowchart is shown below. Figure 1 As shown, it includes the following steps:

[0074] Step 1: Obtain the original 3D dataset of the left ventricular structure and obtain labels through manual annotation by doctors or semi-automatic annotation. The acquired original 3D dataset undergoes format conversion and data cleaning to ensure data consistency and integrity. Image pixel values ​​are normalized to ensure that pixel values ​​in different images are distributed within the same range. Both the data and the labeled images contain 3D image information. Considering the complex 3D structure of the mitral valve, we adopt a patch-based strategy, dividing the 3D data into partially overlapping small patches to ensure the model can capture local and global information. This method is particularly suitable for processing delicate structures such as the mitral valve leaflets and chordae tendineae. Specifically, each pixel value is subtracted from the image mean and then divided by the standard deviation to obtain standardized image data. To improve the robustness and generalization ability of the model, the original 3D dataset is uniformly segmented to obtain image data blocks and corresponding label data blocks of consistent size, such as... Figure 2 As shown. The specific method is to divide the image into equal parts based on its dimensions. If the image size cannot be divided evenly by the block size, then overlapping and padding are used to fill the gaps. For example, if each block is 64x64x64 pixels and the original image size is 64x128x128 pixels, it can be divided into 1*2*2=4 blocks. If the original image size is 150x150x150 pixels, appropriate overlapping and padding are needed to ensure that each block is the same size, thus dividing it into 3*3*3=27 blocks.

[0075] Step 2: Enhance the image data. Specific methods include, but are not limited to:

[0076] Rotation: The image is rotated at random angles to enhance the model's ability to recognize mitral valve structures in different orientations.

[0077] Translation: Randomly translate the image to allow the model to adapt to different mitral valve structures.

[0078] Scaling: Randomly scale the image to simulate mitral valve structures of different sizes.

[0079] Flip: Randomly flip the image horizontally or vertically to improve the model's ability to recognize symmetrical structures.

[0080] Noise: Randomly add appropriate noise to enhance the model's ability to recognize datasets of different quality.

[0081] Through the above various random operations, diverse training samples are generated step by step, thereby increasing the robustness of the model.

[0082] Step 3: Constructing a 3D ConvNeXt encoder-decoder architecture. In this step, we combine the features of Transformer and the advantages of CNN to design a network architecture adapted to the structural characteristics of the mitral valve. Specifically, this architecture retains the detail-capturing ability of CNN and incorporates a spatial attention mechanism to better capture the thin-walled structure of the mitral valve leaflets and the fine features of the chordae tendineae. Simultaneously, a large convolutional kernel design from ConvNeXt is introduced to increase the receptive field, which helps to identify larger-scale structures such as the aorta, left ventricle, left myocardium, and other structures. Upsampling and downsampling are performed using convolutional blocks with inverse bottleneck structures to retain image details at different scales, enhancing the ability to recognize the fine structures of the mitral valve. Furthermore, a multi-scale feature fusion module is employed to effectively integrate features at different scales, improving the model's ability to recognize target objects, such as... Figure 3 As shown. The specific steps are as follows:

[0083] Step 3.1: Use large convolution kernels to perform convolution operations on each voxel and its neighborhood in the 3D image data, thereby extracting features over a large spatial range to capture long-distance spatial dependencies while maintaining computational efficiency.

[0084] Step 3.2: During upsampling, deconvolutional layers or interpolation methods are used to progressively enlarge the feature map, gradually restoring spatial resolution, and convolutional layers are combined for feature extraction. During downsampling, pooling layers or convolutional layers are used to compress the feature map, reducing spatial dimensionality. Residual connections are introduced within each convolutional block to ensure that the original input information is preserved during upsampling and downsampling, thereby enhancing the expressive power of the feature map. This maintains the semantic richness of the feature map, improves gradient fluidity, and avoids information loss.

[0085] Step 3.3: Introduce a multi-scale feature fusion module to extract high-resolution local features at shallower layers and low-resolution global features at deeper layers, improving the model's ability to recognize target objects. Use convolutional or pooling layers to process and fuse feature maps of different scales. By performing weighted fusion or cascading operations on multi-scale features, comprehensive utilization of multi-scale information is achieved.

[0086] Step 4: This invention introduces a strategy of gradually increasing the kernel size, i.e., a progressive learning strategy. Considering the complexity of the mitral valve structure, we first train the model using small convolutional kernels to capture detailed features, and then gradually increase the kernel size to obtain a wider range of contextual information. This method is particularly helpful in handling the complex morphology of the mitral valve leaflets at the junction of the atria and ventricles. By performing trilinear interpolation on the small convolutional kernels, the initial weights of the large convolutional kernels are generated, overcoming the performance saturation problem that may occur when training a large model on a small dataset. The specific steps are as follows:

[0087] Step 4.1: Use small convolutional kernels for initial training, such as 3x3x3 small convolutional kernels, to focus on the fine structure of the mitral valve leaflets and chordae tendineae, ensuring the convergence of the model on limited data.

[0088] Step 4.2: Gradually increase the convolution kernel size to 5x5x5, then to 7x7x7, to capture a wider range of contextual information, which helps to identify the morphology of large structures and capture the semantic relationship between the mitral valve and adjacent large structural organs. The weights of the small convolution kernel are extended to the large convolution kernel through trilinear interpolation.

[0089] Step 4.3: Continue training based on the large convolutional kernel, using the initial weights obtained from the previous interpolation to maintain the previously learned feature information and prevent overfitting and performance saturation.

[0090] Step 5: In model design, a composite scaling strategy is employed to comprehensively adjust the model's depth, width, and kernel size. Parameter optimization is performed based on the specific structure and complexity of the mitral lobe to achieve high-precision image segmentation. The specific steps are as follows:

[0091] Step 5.1: Adjust the depth of the model by increasing or decreasing the number of layers.

[0092] Step 5.2: Adjust the number of convolutional kernels in each layer to change the width of the model.

[0093] Step 5.3: Adjust the size of the convolution kernel according to the specific task requirements to optimize the receptive field and computational efficiency of the model.

[0094] Step 6: During model training, we designed a composite loss function specific to the mitral valve structure. In addition to the Dice loss, cross-entropy loss, global loss, and classification loss, we also introduced an anatomy-based constraint term loss, such as... Figure 4 As shown, this ensures that the segmentation results conform to the anatomical features of the mitral valve. AnatomyLoss loss is used to calculate a penalty term to ensure that the relative positional relationship between the anterior and posterior leaflets is reasonable, helping the model automatically focus on, capture, and correct the relationship between the anterior and posterior leaflets of the mitral valve.

[0095] DC_CE_FOCAL_CLS_loss=α·Soft Dice Loss+β·Cross Entropy Loss+γ

[0096] ·Focal Loss+ε·Classofocation Loss+θ·Anatomy Loss

[0097] Where α, β, γ, ε, and θ are weighting coefficients.

[0098] The specific steps are as follows:

[0099] Step 6.1: Use Soft Dice loss to measure the overlap between the segmentation results and the labels, which is suitable for handling class imbalance problems.

[0100]

[0101] Where, p i It is the predicted probability, g i It is the actual label, and ∈ is a very small constant used to avoid the denominator being zero.

[0102] Step 6.2: Use cross-entropy loss for pixel-level classification and calculate the difference between the segmentation result and the label.

[0103]

[0104] Where C is the number of categories. It is the predicted probability of category c. It is an indicator function for whether the actual label belongs to category c.

[0105] Step 6.3: Use Focal loss to enhance the weights of hard-to-classify samples and mitigate the impact of class imbalance on model training.

[0106]

[0107] γ is an adjustment factor used to control the weights of easy and difficult samples.

[0108] Step 6.4: Introduce a classification loss for specific mitral lobe structures to improve the recognition accuracy of those structures. The classification loss adjusts the weights of each category based on the features of different categories to achieve a more accurate classification result.

[0109]

[0110] Where, p i It is the predicted probability, g i It's a tag.

[0111] Step 6.5: To further improve the accuracy and anatomical rationality of the segmentation results, this invention introduces an anatomical-based constraint term, Anatomy Loss, into the loss function. This constraint term aims to ensure that the segmentation results conform to the anatomical characteristics of the mitral valve, especially the reasonable relative positional relationship between the anterior and posterior flaps. Specifically, by introducing a penalty term, segmentation results that do not conform to common anatomical knowledge are penalized, thereby guiding the model to generate segmentation results that are more consistent with the actual anatomical structure.

[0112]

[0113] Where P i and A j represents the segmentation probability values ​​of the anterior and posterior mitral lobes at positions i and j, respectively.

[0114] Step 6.6: Adjust the weights of each loss term according to different tasks and datasets to achieve the optimal segmentation effect.

[0115] Step 7: After designing the loss function, employ mixed-precision training and gradient checkpointing techniques to reduce GPU memory consumption during training. Utilize the AdamW optimizer, combined with the aforementioned composite loss function, for supervised training to improve segmentation accuracy. The training process includes initial training, intermediate training, and fine-tuning phases, progressively adjusting the learning rate and data augmentation strategies to optimize model performance.

[0116] Step 8: Considering the significant differences in the mitral valve structure at different scales—for example, the leaflets appear as saddle-shaped sheets at small scales, while at large scales they form a clear boundary with the heart chambers; and if prolapse, regurgitation, or other lesions occur, they may appear as non-uniformly distributed or cavitary structures—a deep supervision mechanism is introduced at different decoding stages of the model. Multi-scale labeled images are used for supervision, and the loss is calculated. This method helps the model learn different features of the mitral valve at different levels. The specific steps are as follows:

[0117] Step 8.1: Enlarge or reduce the size of the tag image so that its size is the same as the size of the feature map in the corresponding decoding stage. It is preferable to use bilinear interpolation for scaling.

[0118]

[0119] Step 8.2: Convert the label image into a multi-channel image, with each target structure occupying one channel, such as... Figure 5a .

[0120] One-hot(g,C)=[g==c1,g==c2,...,g==c C ]

[0121] Where g is the label image, C is the number of categories, and C i It belongs to the i-th class.

[0122] Step 8.3: Add Gaussian noise, Poisson noise, salt-and-pepper noise, etc., to the label images of different channels, and extract morphological gradients. Perform dilation and erosion operations on the noisy label images to obtain the two-dimensional morphological gradients, as shown below. Figure 5b Three-dimensional morphological gradients, such as Figure 5c .

[0123] Step 8.4: Perform convolution operations on the feature images of each decoding stage to output multi-channel images, and calculate the loss by combining them with the morphological gradient of the label image. Accumulate the loss values ​​from different decoding stages to obtain the overall loss value for deep supervision.

[0124] Losss d =LossFunc(Conv(x) d ), Morphological Gradient (g d ))

[0125] Where, x d For the feature image of the d-th decoding stage, g d For the corresponding label image, LossFunc is the loss function in step 6.

[0126] Step 9: Based on the loss value calculated in Step 6, and using the backpropagation algorithm and the optimizer set in Step 7, after training in Step 8, perform five-fold cross-validation on multiple datasets to evaluate the model's performance on different tasks and datasets. Verify the model's segmentation effectiveness by calculating the Dice similarity coefficient and Hausdorff distance. Evaluate the model's generalization ability and robustness on a public test set.

[0127] The embodiments of the present invention have been described above. However, the present invention is not limited to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic segmentation method for 3D images of mitral valve structure based on artificial intelligence, characterized in that... Includes the following steps: Step 1: Obtain the original 3D dataset of the heart structure and obtain annotation labels through manual or semi-automatic annotation methods; After normalization, the original 3D dataset is uniformly segmented to ensure that all blocks are of the same size, generating image datasets and label datasets suitable for 3D segmentation. Step 2: Preprocess the mitral valve image data, including image denoising and enhancement; expand the training data volume by enhancing the data to improve the robustness of the model. The enhancement process includes rotation, translation, scaling and flipping. Step 3: After image data preprocessing is completed, a 3D encoder-decoder architecture is constructed using 3D ConvNeXt; Upsampling and downsampling are performed using convolutional blocks with inverse bottleneck structures to effectively preserve image details at different scales and enhance the ability to recognize the fine structure of the mitral lobe. At the same time, a multi-scale feature fusion module is introduced to effectively integrate features at different scales and improve the model's ability to recognize target objects. Step 4: Use a method of gradually increasing kernel size to optimize the transition from small kernel to large kernel; generate the initial weights of large convolutional kernels by performing trilinear interpolation on small convolutional kernels to overcome the performance saturation problem that may occur when training large models on small datasets. Step 5: In the model design, a composite scaling strategy is adopted to comprehensively adjust the depth, width, and kernel size of the model; the parameters are optimized according to the specific structure and complexity of the mitral lobe, thereby achieving high-precision image segmentation; Step 6: During model training, a composite loss function, DC_CE_FOCAL_CLS_ANA_loss, is introduced. This loss function combines Soft Dice loss, cross-entropy loss, Focal loss, CLS loss, and a custom-defined anatomy-based constraint loss. The composite loss function is expressed as follows: in, These are weighting coefficients; Among them, Soft Dice loss is used to measure the overlap between the segmentation result and the label, cross-entropy loss is used for pixel-level classification and to calculate the difference between the segmentation result and the label, Focal loss is used to enhance the weight of hard-to-classify samples, classification loss is used to improve the recognition accuracy of specific mitral valve structures, and an anatomical-based constraint term is introduced into the loss function to further improve the accuracy and anatomical rationality of the segmentation result. Step 7: After designing the loss function, use mixed precision training and gradient checkpointing techniques to reduce the GPU memory consumption during training; use the AdamW optimizer in conjunction with the above composite loss function for supervised training to improve segmentation accuracy; the training process includes an initial training phase, an intermediate training phase, and a fine-tuning phase, gradually adjusting the learning rate and data augmentation strategies to optimize model performance. Step 8: Introduce a deep supervision mechanism at different decoding stages of the model, using multi-scale label images for supervision and calculating the loss. Specifically, the label images are enlarged or reduced to the same size as the feature maps of the corresponding decoding stages; the label images are converted into multi-channel images, with each target structure occupying one channel; noise is added to the label images in different channels, and dilation and erosion operations are performed to obtain morphological gradients; convolution operations are performed on the feature images of each decoding stage to output multi-channel images, and the loss is calculated by combining these images with the morphological gradients of the label images. The loss values ​​from different decoding stages are then summed to obtain the overall loss value for deep supervision. Step 9: Based on the composite loss function defined in Step 6 and the loss value calculated in Step 8, train the model using the backpropagation algorithm and the AdamW optimizer set in Step 7. Then, perform five-fold cross-validation on multiple datasets to evaluate the model's performance on different tasks and datasets. Validate the model's segmentation effect by calculating the Dice similarity coefficient and Hausdorff distance. Evaluate the model's performance on a public test set to verify its generalization ability and robustness.

2. The method for automatic mitral valve structure 3D image segmentation based on artificial intelligence according to claim 1, characterized in that, In step 1, both the image data and the label data contain three-dimensional image information. Considering the complex three-dimensional structure of the mitral valve, a patch-based strategy is adopted to divide the 3D data into partially overlapping small blocks to ensure that the model can capture local and global information. To ensure that all blocks are of the same size, appropriate overlap padding is performed when the data cannot be divided by the specified block size.

3. The method for automatic mitral valve structure 3D image segmentation based on artificial intelligence according to claim 2, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Select a large-size convolutional kernel and apply strided convolution to extract high-level features containing long-range spatial dependencies; Step 3.2: Use deconvolution layers to enlarge the feature map layer by layer or use bilinear interpolation to complete the upsampling operation, and use pooling layers or stride convolution layers to compress the features to complete the downsampling operation; Step 3.3: Use weighted averaging or cascading operation schemes to fuse feature maps of different levels and scales; use a multi-scale feature fusion module to capture local details and global semantic information.

4. The method for automatic mitral valve structure 3D image segmentation based on artificial intelligence according to claim 3, characterized in that, Step 5 specifically includes the following steps: Step 5.1: Adjust the model depth by increasing or decreasing the number of layers; Step 5.2: Adjust the number of convolutional kernels in each layer to change the width of the model; Step 5.3: Adjust the size of the convolution kernel according to the specific task requirements to optimize the receptive field and computational efficiency of the model.

5. The method for automatic mitral valve structure 3D image segmentation based on artificial intelligence according to claim 4, characterized in that, Step 6 further includes the following steps: Step 6.1: Use Soft Dice loss to measure the overlap between the segmentation result and the label. The calculation formula is as follows: in, It is the predicted probability. It's a real label. It is a very small constant used to avoid the denominator being zero; Step 6.2: Apply cross-entropy loss for pixel-level classification. The calculation formula is as follows: in It is the number of categories. It is the predicted category The probability, Is the actual label a category? Indicator functions; Step 6.3: Apply Focal loss to enhance the weights of hard-to-classify samples. The calculation formula is as follows: in It is a regulating factor used to control the weights of easy and difficult samples; Step 6.4: Introduce a classification loss for a specific mitral valve structure, the calculation formula of which is: in, It is the predicted probability. It's a tag; Step 6.5: The calculation formula for the anatomical-based constraint term is as follows: in and Let i and j represent the segmentation probability values ​​of the anterior and posterior mitral lobes at positions i and j, respectively. Step 6.6: Adjust the weights of each loss term according to different tasks and datasets to achieve the optimal segmentation effect.

6. The method for automatic segmentation of 3D images of mitral valve structure based on artificial intelligence according to claim 5, characterized in that, Step 8 further includes the following steps: Step 8.1: The scaling of the label image is performed using bilinear interpolation, as detailed below: Step 8.2: Converting the label image into a multi-channel image includes conversion using One-hot encoding, as shown below: in, For the label image, For the number of categories, For class i; Step 8.3: The noise includes Gaussian noise, Poisson noise, or salt-and-pepper noise; Step 8.4: Sum the loss values ​​from different decoding stages to obtain the overall loss value for deep supervision; where the loss value for each decoding stage is expressed as: in, For the feature image of the d-th decoding stage, For the corresponding label image, This is the loss function for step 6.

Images