Medical image automatic segmentation method based on simplified meta learning

Abstract

The invention provides a medical image automatic segmentation method based on simplified meta learning, and belongs to the technical field of artificial intelligence and medical image processing. Comprising the following steps: step 1, preprocessing a medical image; 2, constructing a double-branch structure composed of a visual Transform branch and a convolutional network branch, respectively extracting global semantic and local detail features, and realizing information complementation through a weighted fusion module; 3, performing model training under a simplified meta-learning framework; and step 4, inputting a to-be-segmented image into the trained model to generate a segmentation result, and outputting a binary mask through a set threshold. According to the method, the advantages of global modeling and local optimization are fused, and the adaptability, generalization and segmentation precision of the model are remarkably improved through a lightweight meta-learning mechanism.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation technology, and specifically relates to an automatic medical image segmentation method based on simplified meta-learning. Background Technology

[0002] Medical image segmentation is a crucial step in clinical auxiliary diagnosis, disease monitoring, and treatment planning. Traditional segmentation methods, such as thresholding, region growing, and edge detection, are often limited by noise interference, lighting variations, and the complexity of tissue structures, resulting in blurred target boundaries and poor robustness. The rise of deep learning technology has driven the development of medical image segmentation, especially convolutional neural network (CNN) models such as U-Net. Through encoder-decoder structures and skip connection mechanisms, these models effectively extract multi-scale local features, demonstrating outstanding performance in tasks such as extracting skin lesions, tumors, and organ contours. However, CNNs are limited by a fixed receptive field, making it difficult to capture global semantic relationships, and they are prone to incomplete segmentation or misclassification in complex scenes.

[0003] Visual Transformer (ViT) possesses global modeling capabilities through its self-attention mechanism, complementing the shortcomings of CNNs in contextual understanding. However, it suffers from high computational cost in high-resolution medical images and insufficient precision in depicting detail boundaries. To address this issue, researchers have proposed a hybrid structure combining global features and local details, and introduced meta-learning concepts to enhance the model's rapid adaptability. Traditional meta-learning often relies on complex second-order gradient calculations in multi-task optimization, resulting in high training overhead and unsuitability for resource-constrained medical scenarios.

[0004] How to solve the above-mentioned technical problems is the challenge facing this invention. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic medical image segmentation method based on simplified meta-learning. By constructing a lightweight inner and outer loop optimization mechanism, the method achieves rapid transfer and adaptation capabilities of the model in multimodal and multi-task segmentation. This method effectively integrates global semantics and local detail features, reduces computational complexity while maintaining high segmentation accuracy, and has the characteristics of full automation. No manual intervention is required during the segmentation process, providing an efficient and scalable solution for automated medical image analysis.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: an automatic medical image segmentation method based on simplified meta-learning, comprising the following steps:

[0007] Step 1: Preprocess the medical image, adjust the image to a uniform resolution, and perform data augmentation operations; Step 1, preprocessing the medical image, adjusting the image to a uniform resolution, and performing data augmentation operations includes the following steps:

[0008] Step 1.1: Image Normalization Processing: The pixel value range of the medical image is normalized to the [0,1] interval. The specific processing method is as follows: Input medical image I, where H is the height, W is the width, and C is the number of channels. Then, a normalization operation is performed to ensure that the pixel values ​​are evenly distributed, which facilitates subsequent feature extraction. The normalization formula is:

[0009]

[0010] in: Indicates the image in The original pixel value at that location, and These represent the minimum and maximum pixel values ​​in the original image, respectively. These are the normalized pixel values. Finally, the normalized image is output. This process helps alleviate the problems of low contrast and noise interference in medical images.

[0011] Step 1.2: Image resizing. Adjust the medical image to a fixed target resolution to ensure consistent input dimensions. Input the normalized medical image. Then, bilinear interpolation is used to scale the image. The specific process is as follows:

[0012]

[0013] in: and These are weighting coefficients, calculated based on the linear relationship between the target pixel coordinates and the original pixel coordinates. Represents the pixel coordinates in the target image. The final output is an image with uniform size, using neighboring original pixels. This is to match the model input requirements and avoid calculation errors caused by inconsistent dimensions.

[0014] Step 1.3: Data Augmentation. Perform data augmentation operations on the images and their masks, including the following random transformations, to increase dataset diversity and improve the model's robustness to complex lesions:

[0015] 1) Randomly flip the image and its corresponding label horizontally with probability P=0.5:

[0016]

[0017] Where W is the image width. This operation simulates mirror symmetry, improving the model's ability to identify lesions that are symmetrical.

[0018] 2) Randomly flip the image and label vertically with probability P=0.5:

[0019]

[0020] Where H is the image height, this operation enhances the model's adaptability to variations in the vertical direction.

[0021] 3) Randomly rotate the image and label within an angle range (±15 degrees), and use bilinear interpolation to maintain the integrity of edge information and avoid pixel loss caused by rotation.

[0022] 4) Randomly crop a region in the image and readjust it to the target resolution. This operation focuses on the local area, simulating local magnification of clinical images, and improves the detection accuracy of small target lesions.

[0023] 5) Add Gaussian noise to simulate noise interference in real medical images and increase the model's robustness to noise:

[0024]

[0025] in: This represents Gaussian noise with a mean of 0 and a variance of σ, where σ is dynamically adjusted based on the image noise level.

[0026] 6) During data augmentation, the label images must undergo the same transformation operation to ensure that the labels are aligned with the image space and to avoid training bias caused by label offset after augmentation.

[0027] Step 1.4: Output the preprocessing results. After normalization, resizing, and data augmentation, the images and their labels are converted to a uniform format for easy batch input into the model.

[0028]

[0029] Where: B is the batch size.

[0030] Preprocessed image and tags This will be used as input to the subsequent fusion and segmentation model for training and inference.

[0031] Step 2: Combining the visual Transformer branch and the convolutional network branch, and fully utilizing global semantic features and local detail features, a highly efficient medical image segmentation model is constructed. This specifically includes the following two core processes:

[0032] Step 2.1: Visual Transformer branch extracts global semantic features. The visual Transformer branch is mainly responsible for extracting global semantic information from the input medical image, using a self-attention mechanism to capture long-distance dependencies, and is suitable for scenarios where the contrast between lesions and the background in the image is low. The specific process is as follows: First, input the preprocessed image. The input image is segmented into a series of non-overlapping image patches through a convolutional layer. Each patch is mapped to a high-dimensional feature representation to preserve the global context.

[0033]

[0034] Then output features Where: N is the number of image patches, and D is the embedding feature dimension.

[0035] Next, position encoding is added to each image patch to introduce position information, enabling the Transformer to distinguish the spatial location of each image patch and avoid the loss of position information:

[0036]

[0037] Features with added position encoding The input is processed through a multi-layer Transformer block. First, a self-attention mechanism is used to capture long-distance dependencies between image patches and extract global contextual information. Then, a feedforward network (MLP layer) is used to further enhance feature representation capabilities. This process freezes pre-trained weights to reduce computational burden. Finally, after processing through multiple Transformer blocks, global semantic features are obtained and output.

[0038]

[0039] Finally, a 1×1 convolutional layer is used to process the high-dimensional features output by the Transformer. Project to a lower channel dimension to match subsequent fusion requirements:

[0040]

[0041] Output global features .

[0042] Step 2.2: Convolutional network branches extract local detail features. The convolutional network branches are primarily responsible for capturing the details and local features of the input image to supplement the edge and texture information lacking in the Transformer branches. The specific process is as follows:

[0043] First, input the preprocessed image. The input image is processed through stacked convolutional blocks, with each layer progressively extracting multi-scale local features:

[0044]

[0045] In each convolutional block, local detail features are extracted through convolution operations. Then, a normalization layer accelerates training and improves feature stability. Finally, an activation function is used to introduce non-linear feature representation capabilities. Furthermore, downsampling is used between each convolutional layer to progressively reduce the spatial resolution of the feature map and increase the number of channels. The final local features are then output.

[0046]

[0047] This branch preserves shallow features through skip connections, ensuring accurate segmentation of small targets and irregularly shaped lesions.

[0048] Step 2.3: Weighted Feature Fusion. A weighted fusion module combines global features from the visual Transformer branch and local features from the convolutional network branch. The specific operation is as follows:

[0049] Input the global features from step 2.1 and the local features of step 2.2 It also uses an attention mechanism to dynamically assign importance weights to global and local features to balance their contributions in complex scenarios.

[0050]

[0051] Finally, the fused features are output:

[0052]

[0053] Step 3: Train the overall segmentation model using a supervised learning method with a binary cross-entropy loss function, incorporating an inner and outer loop optimization strategy to simplify the meta-learning framework during training. This includes the following sub-steps:

[0054] Step 3.1: The decoder restores the spatial information of the fused image. The decoder is responsible for progressively upsampling the fused features from Step 2.3 to restore the original resolution of the image and generate a segmentation mask to address the issues of blurred lesion boundaries and irregular shapes in dermoscopic images. The specific process is as follows:

[0055] Input fused features Upsampling is performed layer by layer through decoder blocks, with each block combining transposed convolutions and convolutions to refine features. First, transposed convolutions (deconvolutions) are used to progressively increase the resolution of the feature maps, while simultaneously extracting richer spatial detail information.

[0056]

[0057] in, Indicates the decoder's first The input feature map of the layer is used to progressively restore the image size, avoiding information loss. Next, skip connections are introduced to directly pass shallow features from the encoder to the corresponding decoding layer, preserving low-level details such as lesion edges.

[0058]

[0059] in, These are the features after upsampling. It is the encoder number The output features of the layers, this cascaded operation ensures that the model accurately locates small objects in segmentation.

[0060] In the final layer of the decoder, 1×1 convolutions are used to map features to a segmentation mask, and a sigmoid activation function is applied for normalization to generate a probability mask.

[0061]

[0062] in, Mask represents the probability value of each pixel belonging to the lesion area.

[0063] Step 3.2: Inner Loop Optimization Strategy. In the simplified meta-learning framework, the preprocessed dataset is first divided into multiple sub-tasks, each containing a small number of sample images to quickly adapt to the segmentation requirements of a specific task. The specific process is as follows:

[0064] The input is a small task dataset after partitioning. The training set is further split into K-fold subsets, with each fold used alternately for training and validation.

[0065] For each sub-task, use a small number of samples to perform a gradient descent update of the model's local parameters. This allows for rapid adaptation to the current task without retraining the entire network.

[0066]

[0067] in, It is the inner loop learning rate. This is the loss function, and the update focuses on task-specific features. This step has low computational complexity, mitigating the catastrophic forgetting problem.

[0068] Step 3.3: Outer Loop Optimization Strategy. The outer loop integrates update information from multiple sub-tasks, calculates the average gradient of parameter differences between tasks using a first-order approximation method, and updates the global model parameters. The specific process is as follows:

[0069] Collect the parameter update differences for each task after the inner loop. Then calculate the average:

[0070]

[0071] in, It's the number of tasks. It is the first The local parameters for each task are then updated using the average gradient.

[0072]

[0073] in, It is the outer loop learning rate.

[0074] Step 3.4: Overall Supervised Training Setup. The binary cross-entropy loss function (BCE) is used to optimize the binary classification accuracy between foreground and background, combined with the AdamW optimizer for end-to-end training. The specific process is as follows:

[0075] Input preprocessed batch images and labels, and calculate BCE loss:

[0076]

[0077] in, It's a real label. It is a prediction probability. It refers to the number of pixels.

[0078] Step 4: Input the medical image to be segmented into the trained model. The model generates segmentation results through an adaptation mechanism within a simplified meta-learning framework. Evaluation metrics include mean intersection-over-union ratio (CIU), Dice similarity coefficient, accuracy, sensitivity, and specificity. This includes the following sub-steps:

[0079] Step 4.1: Input the image to be segmented. Input the test image into the trained model, and then extract features through a dual encoder and fusion module.

[0080] Step 4.2: Generate Segmentation Mask. Recover features using the decoder and apply a sigmoid function to generate a probability mask, outputting a binary segmentation result. The specific process is as follows: Input fused features After upsampling by the decoder and skip connections, the final mask is calculated:

[0081]

[0082] in, This indicates the predicted lesion area.

[0083] Step 4.3: Evaluate segmentation performance. Quantify the model's performance using multiple metrics, calculating the average across the dataset. Specific metrics are as follows:

[0084] Mean Intersection over Union (IoU):

[0085]

[0086] Dice similarity coefficient:

[0087]

[0088] Accuracy:

[0089]

[0090] Specificity:

[0091]

[0092] Sensitivity:

[0093]

[0094] Among them, TP, FP, FN, and TN represent true positive, false positive, false negative, and true negative, respectively.

[0095] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0096] 1. By introducing an inner and outer loop optimization strategy to simplify the meta-learning framework, this invention can effectively alleviate the catastrophic forgetting problem of traditional medical image segmentation methods in multi-task scenarios, improve the model's rapid adaptability in different tasks, and thus significantly improve segmentation accuracy and generalization ability.

[0097] 2. The outer loop optimization designed in this invention adjusts the global parameters by calculating the average value of the parameter update differences between tasks and using a first-order approximation method. This avoids the high overhead of second-order gradient calculation in traditional meta-learning, reduces computational complexity, and makes the model more suitable for clinical application scenarios with limited hardware resources.

[0098] 3. In this invention, an inner loop optimization strategy is combined to perform rapid gradient updates using a small number of sample images, thereby achieving local parameter adaptation for specific tasks, enhancing the model's robustness to multimodal medical images, and avoiding performance degradation caused by task switching.

[0099] 4. This invention, through effective image preprocessing and K-fold cross-validation strategy, combined with the optimized design of a simplified meta-learning framework, enables the model to exhibit excellent stability and generalization ability when processing medical images of different resolutions, noise interference, or varying quality. It is particularly suitable for segmentation tasks involving blurred boundaries or small target lesions.

[0100] 5. This invention optimizes the model training process by employing a binary cross-entropy loss function and a cosine annealing learning rate scheduler, reducing the error between the predicted segmentation results and the true labels, and ensuring that the model has high practicality and reliability in actual clinical diagnosis. Attached Figure Description

[0101] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0102] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0103] Figure 2 This is a flowchart of the visual Transformer branch of the SAM encoder in this invention.

[0104] Figure 3 This is a simplified flowchart of the meta-learning framework in this invention.

[0105] Figure 4 This is a schematic diagram of the weighted fusion module in this invention. Detailed Implementation

[0106] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0107] Example 1

[0108] In this embodiment, the experiment was implemented using Python on a Windows 11 platform. The computer platform used in this embodiment has a 4060ti 8GB GPU and 32GB of RAM. During the experiment, the model was tested and evaluated on the ISIC2018 dataset and compared with other methods using the same dataset.

[0109] This embodiment provides a method for automatic medical image segmentation based on simplified meta-learning, including the following steps:

[0110] Step 1: Preprocess the medical images by adjusting them to a uniform resolution and performing data augmentation operations. For example... Figure 1As shown, after inputting a medical image, it is first preprocessed to standardize the image data format and enhance the diversity of the data samples. First, the pixel values ​​of the input medical image are mapped to a uniform range. Then, bilinear interpolation is used to adjust the input image to a uniform resolution. Next, a random transformation operation is performed on the image, and the same transformation is applied to the image mask to ensure data alignment with the label space. Finally, the preprocessed medical image and its corresponding label are output.

[0111] Step Two: Combining the visual Transformer branch and the convolutional network branch, and fully utilizing global semantic features and local detail features, a highly efficient medical image segmentation model is constructed. For example... Figure 1 As shown, after the image preprocessing module, the image is input in parallel to the visual Transformer branch and the convolutional network branch.

[0112] In the Visual Transformer branch, the preprocessed medical image is first input. For example... Figure 2 As shown, the image is segmented into fixed-size patches through a patch embedding operation and embedded in a high-dimensional space. Then, positional encoding is added, and the patch embedding vectors are passed layer by layer through multiple Transformer blocks, extracting global semantic features through a self-attention mechanism. Finally, channel projection and bilinear interpolation operations are used to adjust the number of channels in the feature map to match the size, and the global features are output.

[0113] In the convolutional network branch, the preprocessed medical image is first input, and then local detail features are extracted layer by layer through multiple convolutional operations. For example... Figure 3 As shown, the encoder captures detailed features such as image edges and textures through small receptive field convolution operations, while gradually reducing the resolution and increasing the number of channels through downsampling operations. Finally, the output of each layer of the encoder is passed to the corresponding position of the decoder through skip connections, and outputs local features.

[0114] Finally, as Figure 4 As shown, the input consists of global features from the visual Transformer branch and local features from the convolutional network branch. The weights α are dynamically assigned through the attention mechanism, and the advantages of global and local features are combined to generate fused features and output them.

[0115] Step 3: Train the overall segmentation model using a supervised learning method with a binary cross-entropy loss function. During training, incorporate an inner and outer loop optimization strategy that simplifies the meta-learning framework to minimize the error between the predicted segmentation result and the true label. For example... Figure 1As shown, the fused features output from step two are input to the decoder module. The decoder recovers the spatial resolution of the feature map through layer-by-layer upsampling operations. In each decoding layer, skip connections are used to combine the intermediate features of the convolutional network encoder with the current decoder features, ensuring the fusion of high-level semantics and low-level detail information. The last layer of the decoder maps the number of feature channels to 1 channel through a 1×1 convolution, generating a probability map. Finally, the Sigmoid activation function is used to normalize the output, generating a pixel-level segmentation probability map.

[0116] like Figure 3 As shown, a simplified meta-learning framework is introduced during training. First, the dataset is divided into multiple small tasks, and K-fold cross-validation is used. The inner loop performs fast gradient updates for each small task using a small number of samples to achieve local parameter adaptation. The outer loop integrates the update differences across multiple tasks, calculates the average gradient using a first-order approximation, and adjusts the global parameters. The overall training uses the AdamW optimizer with an initial learning rate of 0.001, a cosine annealing scheduler decaying to 1e-6, 300 training epochs, a batch size of 8, and saves the optimal weights on the validation set every 20 epochs. This process is repeated three times on a single RTX 4060Ti GPU, and the average value is taken.

[0117] Step 4: Input the image to be segmented into the trained model, and generate segmentation results through the adaptation mechanism of the simplified meta-learning framework. Set a threshold T=0.5 to convert the probability map into a binary segmentation result. The generated segmentation result is a binary mask image, and the segmentation result is compared with the true label. Calculate metrics such as the Dice coefficient and IoU to evaluate the segmentation performance.

[0118] The present invention was compared with the current mainstream skin disease image segmentation models on the ISIC2018 dataset, and the results are shown in Table 1.

[0119] Table 1. Performance comparison of segmentation models on the ISIC2018 dataset.

[0120]

[0121] Table 1 compares the performance of different segmentation models on the ISIC2018 dataset. Evaluation metrics include Intersection over Union (IoU), Dice similarity coefficient (Dice), accuracy (Acc), specificity (Spe), and sensitivity (Sen). The results show that the method presented in this invention outperforms other comparative models in all five metrics: IoU, Dice, Acc, Spe, and Sen. Specifically, the IoU reaches 0.840, an improvement of approximately 7.8% compared to UNet, indicating a significant improvement in the segmentation accuracy of the target region and background. The Dice metric is 0.904, an improvement of 1.6% to 3.2% compared to other methods, demonstrating the advantage of this method in preserving the shape and integrity of the target region. Acc reaches 0.963, Spe is 0.981, and Sen is 0.908, all slightly higher than MALUNet and SANet, showing the superior performance of this method in overall pixel segmentation accuracy, background noise removal, and small target detection.

[0122] This invention combines global and local feature fusion with a simplified meta-learning framework, ensuring global semantic understanding of the segmentation while enhancing multi-task adaptability and generalization ability. As a result, it achieves optimal segmentation results on multiple evaluation metrics, verifying the effectiveness and superiority of the proposed method.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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. A method for automatic medical image segmentation based on simplified meta-learning, characterized in that, Includes the following steps: Step 1: Preprocess and segment the medical image dataset; Step 2: Construct an optimized structure for the simplified meta-learning framework, and use the inner loop to quickly adjust parameters for each sub-task using a limited number of sample images; Step 3: The outer loop optimizes and integrates the update information of multiple tasks, and optimizes the global model parameters by calculating the average difference of parameters of each task; Step 4: Train the overall segmentation model using a supervised learning method equipped with a binary cross-entropy loss function, and gradually reduce the learning rate to a preset minimum value using a cosine annealing learning rate scheduler.

2. The automatic medical image segmentation method based on simplified meta-learning according to claim 1, characterized in that, Step one includes the following steps: Step 1.1: Image normalization processing. The pixel values ​​of the original medical image are normalized and mapped to a uniform range to eliminate grayscale differences caused by different devices or acquisition conditions. Step 1.2: Unify resolution adjustment. Use bilinear interpolation or equivalent interpolation methods to scale the normalized image to the preset target resolution. Step 1.3: Data augmentation and label synchronization preprocessing, performing a series of random augmentation operations on the image and its corresponding segmentation label mask; Step 1.4: Output standardized samples. After normalization, size adjustment and enhancement, output medical images with uniform shape and format and their corresponding labels.

3. The automatic medical image segmentation method based on simplified meta-learning according to claim 1, characterized in that, The specific steps in step two are as follows: Step 2.1: Visual Transformer Branch. Input the preprocessed medical image into the visual Transformer branch. Through patch segmentation and embedding, position encoding and multi-layer Transformer module, the long-distance dependence and global semantic relationship between different regions in the image are captured by the self-attention mechanism. The global semantic features are then output for subsequent fusion through channel projection or dimensionality reduction transformation. Step 2.2: Convolutional Network Branch. The same preprocessed image is input into the convolutional network branch in parallel. By stacking small receptive field convolutional modules, normalization layers and nonlinear activation functions, multi-scale local detail information is extracted layer by layer. Shallow spatial details are preserved for use by the decoder through downsampling and skip connections. Step 2.3: Weighted fusion module. The global semantic features of the visual Transformer branch and the local detail features of the convolutional network branch are input into the weighted fusion module. Attention or gating mechanism is used to dynamically estimate the importance weights of global and local features. The two features are then combined linearly or non-linearly according to their weights to produce a fused feature representation that contains both global context and local boundary information. Step 2.4: The features are passed to the decoder. The fused high-quality feature map is passed to the decoder module. The decoder restores the spatial resolution and generates a predicted segmentation probability map for loss calculation through layer-by-layer upsampling, feature refinement and skip connection fusion operations.

4. The automatic medical image segmentation method based on simplified meta-learning according to claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Initialize model parameters, and initialize the parameters for each branch of the segmentation model; Step 3.2: Define the training objective. The objective is pixel-level binary classification. Applicable supervised loss functions such as binary cross-entropy are used as optimization objectives to measure the difference between the model's predicted probability map and the true label, and to serve as the basis for gradient updates. Step 3.3: Sample Input and Forward Inference. The preprocessed training samples are input into the model. The model extracts features in the visual Transformer branch and the convolutional network branch, and obtains the predicted segmentation probability map through weighted fusion and the decoder module, which serves as the direct output of the loss calculation. Step 3.4: Loss calculation. Based on the predicted probability map and true labels obtained in Step 3.3, calculate the loss value sample by sample or batch by batch according to the defined loss function to measure the performance of the current model on the task. Step 3.5: Inner loop and parameter update. Following the simplified meta-learning framework, the training data is divided into multiple small tasks. Within each small task, a limited number of samples are used to perform several steps of fast gradient update to obtain short-term adaptation of the local parameters of the task, thereby improving the model's specialized performance for this type of sample. Step 3.6: Outer loop and global update. Collect the local parameter change information obtained after the inner loop of each sub-task, integrate the update direction of each task using a first-order approximation or averaging strategy, and update the global model parameters accordingly. The outer loop updates the gradient-based optimizer and gradually adjusts the learning rate in combination with learning rate scheduling until the training termination condition is met. Step 3.7: Repeat training until convergence. Repeat the inner and outer loop training process from Step 3.3 to Step 3.6 until the preset termination condition is reached, and save the model weights corresponding to the best validation performance for inference.

5. The automatic medical image segmentation method based on simplified meta-learning according to claim 1, characterized in that, Step 4 includes the following steps: Step 4.1: Input the image to be segmented. Input the medical image to be analyzed into the trained segmentation model. First, perform basic preprocessing operations such as normalization and size adjustment to ensure that the input format is consistent with the training stage. Step 4.2: Generate segmentation results. The model utilizes the adaptive capability of the simplified meta-learning framework to quickly extract and fuse features from the input image and output the corresponding probability segmentation map. Step 4.3: Performance evaluation and output. The predicted mask is compared with the real label. The segmentation performance of the model is evaluated using indicators such as Dice coefficient, intersection-over-union ratio (IoU), accuracy, sensitivity, and specificity. For unlabeled data, the predicted mask is directly output as an auxiliary diagnostic result and applied to clinical analysis or data labeling assistance systems.

Images