Medical image segmentation system and method based on deep learning

Abstract

The invention relates to the technical field of medical image processing, and discloses a medical image segmentation system and method based on deep learning, and the system comprises a data preprocessing module which is used for receiving original medical image data, and carrying out the preprocessing operation of normalization, denoising, smoothing and the like, so as to improve the image quality, reduce the noise interference, and improve the image segmentation efficiency. The analyzability of the image is enhanced; and the deep learning model training module comprises a constructed deep learning network structure, and the network structure at least comprises a convolution layer, a pooling layer and a full connection layer. By adopting a deep learning network structure of a U-Net architecture or a variant thereof, multi-scale feature extraction, extraction of high-level semantic features by down-sampling and fusion of high-level and low-level features by up-sampling are realized, so that a target region in a medical image is positioned and segmented more accurately, the boundary and region of tissues and organs are marked clearly, a more reliable basis is provided for medical diagnosis, and the medical image segmentation efficiency is improved. And the high clinical requirement on segmentation precision is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a medical image segmentation system and method based on deep learning. BACKGROUND

[0002] In the field of clinical medicine, the rapid development of medical imaging technology provides key basis for the diagnosis, treatment and research of diseases. Medical image segmentation, as a core task in medical image analysis, aims to accurately separate specific tissues, organs or lesion regions from complex medical images, which has important value for early detection, accurate diagnosis and treatment plan of diseases.

[0003] Traditional medical image segmentation techniques mainly include threshold-based methods, region growing-based methods and edge detection-based methods, etc. Threshold-based methods separate the target region from the background region in the image by setting appropriate gray threshold, however, this kind of method has high requirements for image quality and uniformity of target region gray distribution, and when facing medical images with uneven gray scale and large noise interference, the segmentation effect is often not ideal, and the target region may be missed or misclassified. Region growing-based methods start from a seed point and gradually expand the region according to the pre-set similarity criteria, but the selection of seed point is sensitive, and inappropriate seed point selection may lead to incorrect growth path and boundary, and it is difficult to accurately define the region boundary when dealing with images of organs with complex internal structure. Edge detection-based methods achieve segmentation by identifying edge features in the image, but the edges in medical images are often affected by noise, gray gradient and other factors, resulting in discontinuous or false edges, which affects the accuracy and integrity of segmentation.

[0004] With the continuous progress of computer technology, machine learning methods have gradually emerged in the field of medical image segmentation. Early machine learning methods mainly rely on manual extraction of image features such as texture features and shape features, and then use classification algorithms to classify target regions and non-target regions. However, the process of manually extracting features is not only time-consuming and laborious, but also the extracted features often have subjectivity and limitations, making it difficult to accurately describe the complex tissue and organ features in medical images, which limits the further improvement of segmentation performance.

[0005] In addition, the quality of the segmentation result is crucial for clinical application. After outputting the segmentation result, the existing methods often lack effective post-processing means to optimize the result, resulting in problems such as isolated points and non-smooth boundaries in the segmentation result, which affects the accuracy and reliability of the segmentation result. At the same time, in the aspect of visual display, most of the existing medical image segmentation systems can only provide single two-dimensional slice display or simple three-dimensional reconstruction, which cannot meet the needs of doctors or researchers for all-round and multi-angle observation and analysis of medical images and their segmentation results, and is not conducive to their accurate diagnosis and research of diseases.

[0006] Therefore, the person skilled in the art proposes a medical image segmentation system and method based on deep learning to solve the above problems. SUMMARY

[0007] In view of the deficiencies of the prior art, the present application provides a medical image segmentation system and method based on deep learning, which solves the problems proposed in the background art.

[0008] To achieve the above purpose, the present application is implemented by the following technical solutions: a medical image segmentation system based on deep learning, comprising:

[0009] A data preprocessing module is used to receive original medical image data and perform normalization, denoising and smoothing preprocessing operations thereon to improve image quality, reduce noise interference and enhance the analyzability of the image.

[0010] A deep learning model training module includes a constructed deep learning network structure, which at least includes a convolutional layer, a pooling layer and a fully connected layer, for feature extraction and learning of the preprocessed medical image data, and through training on a large amount of labeled and unlabeled medical image data, the model can automatically learn the feature representation of different tissues and organs in the medical image to achieve accurate segmentation of the medical image.

[0011] A segmentation result generation module is used to input the medical image data to be segmented into the trained deep learning model, and through forward propagation calculation of the model, the segmentation result is outputted in the form of a binary image or a multi-classification image, clearly indicating the boundaries and regions of different tissues and organs in the medical image.

[0012] A post-processing module is used to further optimize the segmentation result, including removing isolated points, filling holes and smoothing boundaries.

[0013] A display module is used to visually display the original medical image, the preprocessed image, the segmentation result and the three-dimensional reconstructed image after segmentation.

[0014] a storage module for storing raw medical image data, preprocessing parameters, parameters of the deep learning model, segmentation results, and related intermediate data;

[0015] In the specific implementation of the deep learning network structure, a plurality of convolutional neural networks are used in the convolutional layer, each convolutional neural network includes a plurality of convolution kernels for extracting different scale and level features of the image data, a maximum pooling or average pooling method is used in the pooling layer to reduce the spatial dimension of the feature map, reduce the amount of calculation and extract the main features, the full connection layer maps the feature map to the output space to generate the final segmentation result, the encoder part of the U-Net architecture or its variant includes a plurality of convolutional layers and pooling layers for extracting high-level semantic features of the image, and the decoder part includes a plurality of convolutional layers and up-sampling layers for fusing high-level semantic features and low-level features to achieve accurate positioning and segmentation of the target region.

[0016] Preferably, the deep learning network structure in the deep learning model training module adopts the U-Net architecture or its variant, which realizes multi-scale feature extraction of medical images through up-sampling and down-sampling, extracts high-level semantic features of the image in the down-sampling process, and fuses the high-level semantic features with the low-level features in the up-sampling process, so as to more accurately locate and segment the target region in the medical image and improve the accuracy and detail performance of the segmentation.

[0017] Preferably, the data preprocessing module further includes a sub-module for data enhancement of medical image data, and the data enhancement method includes one or more combinations of random rotation, random translation, random scaling, random flipping, and adding Gaussian noise, so as to expand the size of the training data set by performing data enhancement operation on the original medical image data.

[0018] Preferably, the deep learning model training module further includes a model evaluation sub-module for periodically evaluating the deep learning model during the training process, and the evaluation index includes one or more combinations of segmentation accuracy, recall rate, F1 value, and Dice coefficient.

[0019] Preferably, the storage module adopts a distributed storage architecture, which can efficiently store and manage large-scale medical image data and model parameters, support fast read-write and query operations of data, and has data backup and recovery functions to ensure the safety and reliability of the data.

[0020] Preferably, the display module supports a plurality of visualization display modes, including two-dimensional slice display, three-dimensional reconstruction display, and fusion display, and users can select different display modes according to needs to observe and analyze the medical image and its segmentation results from all angles.

[0021] A medical image segmentation method based on deep learning, comprising the following steps:

[0022] Step one: Obtain the original medical image data and use the data preprocessing module to perform normalization, denoising and smoothing and other preprocessing operations to improve the image quality;

[0023] Step two: Construct a deep learning network structure, which at least includes one convolutional layer, one pooling layer and one fully connected layer, use the preprocessed medical image data to train the deep learning model, learn the feature representation of different tissues and organs in the medical image by learning a large number of labeled and unlabeled medical image data;

[0024] Step three: input the medical image data to be segmented into the trained deep learning model, and output the segmentation result through the forward propagation calculation of the model, which is presented in the form of binary image or multi-classification image, indicating the boundaries and regions of different tissues and organs in the medical image;

[0025] Step four: use the post-processing module to optimize the segmentation result, including removing isolated points, filling holes, smoothing boundaries and other operations to improve the accuracy and integrity of the segmentation result; the specific optimization method of the post-processing module is:

[0026] The step of removing isolated points is performed on two-dimensional slices and three-dimensional body data respectively to ensure the connectivity of the segmentation result in different dimensions;

[0027] The step of filling holes sets different filling strategies according to the morphological characteristics of the target region, and adopts overall filling for regular-shaped target regions and local filling for irregular-shaped target regions;

[0028] The step of smoothing the boundary sets the intensity of Gaussian filtering according to the detail requirements of the segmentation result, and adopts lower intensity filtering for the segmentation result that needs to retain small structures;

[0029] Step five: visualize the original medical image, preprocessed image, segmentation result and three-dimensional reconstructed image through the display module to assist doctors or researchers in observing and analyzing the medical image.

[0030] Preferably, in step two, the deep learning network structure adopts U-Net architecture or its variants, which realizes multi-scale feature extraction of medical images through upsampling and downsampling, extracts high-level semantic features in the downsampling process, and fuses high-level semantic features and low-level features in the upsampling process, so as to more accurately locate and segment the target region in the medical image and improve the segmentation precision and detail performance.

[0031] Preferably, after step one, a data enhancement step is further included, and the data enhancement method at least includes one or more combinations of random rotation, random translation, random scaling, random flipping and adding Gaussian noise, the scale of the training data set is expanded by the data enhancement operation on the original medical image data, and the generalization ability and adaptability to different image conditions of the deep learning model are improved.

[0032] Preferably, in step two, the deep learning model is also periodically evaluated, and the evaluation index at least includes one or more combinations of segmentation accuracy, recall rate, F1 value and Dice coefficient, and the training parameters and optimization strategy of the model are adjusted according to the change trend of the evaluation index.

[0033] The application provides a deep learning-based medical image segmentation system and method.

[0034] Advantages:

[0035] 1、The application adopts a deep learning network structure of U-Net architecture or its variant, realizes multi-scale feature extraction, extracts high-level semantic features through downsampling, and fuses high-level and low-level features through upsampling, so as to more accurately locate and segment the target region in the medical image, clearly mark the boundaries and regions of tissues and organs, and provide more reliable basis for medical diagnosis, and meet the high requirements of clinical segmentation accuracy.

[0036] 2、The application introduces a data enhancement step, uses random rotation, translation, scaling, flipping and adding Gaussian noise and other means to expand the scale of the training data set, provides more diverse data for model training. At the same time, the model is periodically evaluated and the training strategy is adjusted according to the index, so that the model better adapts to different image conditions, enhances the generalization ability, ensures stable output of high-quality segmentation results when processing complex and diverse actual medical images, and improves the reliability and practicability of the system.

[0037] 3、The application uses a post-processing module to remove isolated points, fill holes, smooth boundaries and other optimization processing on the segmentation results, effectively improves the integrity and accuracy of the segmentation results. At the same time, the display module supports multiple visualization display modes, such as two-dimensional slices, three-dimensional reconstruction and fusion display, users can select as needed, and observe and analyze medical images and their segmentation results from all directions and multiple angles, so as to bring better visual experience to doctors or researchers, assist them to more accurately observe and analyze medical images, and help disease diagnosis and treatment plan making. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The system overall architecture of the application is shown in the figure;

[0039] Figure 2 The deep learning model training flowchart of the application is shown in the figure;

[0040] Figure 3 Flow chart of the medical image segmentation method of the present application;

[0041] Figure 4 Flow chart of the display module visualization of the present application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0043] Please refer to the drawings in the specification of the present application Figure 1 - the drawings in the specification of the present application Figure 4 The embodiments of the present application provide a medical image segmentation system based on deep learning, which comprises:

[0044] The data preprocessing module is used for receiving original medical image data and performing preprocessing operations such as normalization, denoising and smoothing on the original medical image data, so as to improve image quality, reduce noise interference and enhance the analyzability of the image. The data preprocessing module further comprises a data enhancement submodule for performing data enhancement on the medical image data. The data enhancement method comprises one or more combinations of random rotation, random translation, random scaling, random flipping and adding Gaussian noise. The scale of the training data set is expanded by performing data enhancement operations on the original medical image data.

[0045] Specifically, first, the image is normalized to unify the image data in different ranges to a specific interval, so that the data is standardized and convenient for subsequent processing. At the same time, denoising and smoothing algorithms are used to remove noise in the image. These noises may be caused by interference in the imaging device or the imaging process, which will lead to a decrease in image quality and affect the segmentation effect. Denoising and smoothing processing can effectively reduce noise interference, make the image clearer and enhance its analyzability, providing a high-quality data basis for the training and segmentation of the deep learning model. In addition, the data preprocessing module also includes a data enhancement submodule, which expands the original medical image data by methods such as random rotation, translation, scaling, flipping and adding Gaussian noise. Data enhancement technology can generate more training samples with diversity and in accordance with the actual distribution without increasing the actual data acquisition cost, effectively expanding the scale of the training data set, helping to improve the generalization ability of the deep learning model, so that it can more stably output accurate segmentation results when facing medical images under different conditions, and thus improve the performance and reliability of the entire medical image segmentation system.

[0046] The deep learning model training module includes a constructed deep learning network structure, which at least contains a convolutional layer, a pooling layer and a fully connected layer, is used for feature extraction and learning of pre-processed medical image data, and through training of a large amount of labeled and unlabeled medical image data, the model can automatically learn the feature representation of different tissues and organs in the medical image to realize accurate segmentation of the medical image; the deep learning network structure in the deep learning model training module adopts a U-Net architecture or a variant thereof, which realizes multi-scale feature extraction of the medical image through upsampling and downsampling, extracts high-level semantic features of the image in the downsampling process, and fuses the high-level semantic features with low-level features in the upsampling process, so as to more accurately locate and segment the target region in the medical image and improve the accuracy and detail performance of the segmentation.

[0047] Specifically, the deep learning model training module is responsible for constructing and training a deep learning network for medical image segmentation. The network structure at least contains a convolutional layer, a pooling layer and a fully connected layer. The convolutional layer extracts image features through a convolution kernel, the pooling layer is used for dimension reduction and reduction of overfitting, and the fully connected layer maps features to an output space to realize feature extraction and learning. Through training of a large amount of labeled and unlabeled medical image data, the model can automatically learn the feature representation of different tissues and organs to realize accurate segmentation.

[0048] The module adopts a U-Net architecture or a variant thereof to realize multi-scale feature extraction through upsampling and downsampling. The downsampling gradually extracts high-level semantic features of the image, and the upsampling fuses the high-level semantic features with low-level features to more accurately locate and segment the target region. Its algorithm includes the following contents:

[0049] Suppose the input image data is X, the operation of the convolutional layer can be represented as:

[0050] Y = σ (W * X + b)

[0051] Where W is the convolution kernel, b is the bias term, * represents the convolution operation, and σ is the activation function. The pooling layer adopts maximum pooling or average pooling to reduce the spatial dimension of the feature map. The fully connected layer maps the feature vector to the output space, and its calculation formula is:

[0052] Y = σ (W * X + b)

[0053] Where W is the weight matrix, b is the bias term, and σ is the activation function. In the training process, the network parameters are updated through the backpropagation algorithm to minimize the loss function:

[0054]

[0055] Where y i is the true label, is the predicted probability. The feature extraction and fusion process of the U-Net architecture is achieved through an encoder-decoder structure, where the encoder is responsible for down-sampling to extract high-level features, and the decoder is responsible for up-sampling to fuse multi-scale features, thereby improving segmentation accuracy and detail performance.

[0056] As a preferred solution, the above technical solution can be replaced by a DeepSeek inference model, the specific content of which is as follows:

[0057] The overall architecture of the system still maintains the original modular design, including data preprocessing module, deep learning model training module, segmentation result generation module, post-processing module, display module and storage module. On this basis, the DeepSeek inference model is integrated as an important part of the deep learning model training module.

[0058] In the deep learning model training module, the backbone network uses a pre-trained visual base model such as DeepSeek-VL or SwinTransformer. These pre-trained models can provide powerful feature extraction capabilities for medical image segmentation tasks. The multi-scale feature fusion module is used to integrate features at different levels to better understand the information in medical images. The decoder part uses a MaskDecoder based on Transformer, which supports text and visual Prompt input, allowing the model to interactively segment according to different prompt information.

[0059] When integrating the DeepSeek inference model into the system, after the encoder extracts the image features, these features are input into the DeepSeek inference model for further processing and reasoning. DeepSeek inference model can use the knowledge obtained from pre-training on large-scale data to enhance or adjust the image features, making them more suitable for subsequent segmentation tasks.

[0060] During processing, the DeepSeek inference model receives feature maps from the encoder and possibly text prompts, and outputs processed feature maps or intermediate representations. These features processed by the DeepSeek inference model will be fused with the original features and then input together into the decoder for final segmentation mask generation.

[0061] In the inference process, the system first preprocesses the input medical image data, including normalization, denoising and smoothing operations, to improve image quality. Then, the preprocessed image data is input into the encoder of the deep learning model to extract multi-scale feature representations. These features are then fed into the DeepSeek inference model, combined with possible text prompt information, for feature enhancement or adjustment. The enhanced features are fused with the original features and input into the Transformer-based decoder to generate the final segmentation results.

[0062] In the image preprocessing stage, the system performs normalization, denoising and smoothing on the original medical image data. Taking normalization as an example, the formula is as follows:

[0063]

[0064] where X is the original image data pixel value. μ is the mean of the image data, used to adjust the data to the appropriate range. σ is the standard deviation of the image data, used to measure the dispersion of the data. X norm The normalized image data pixel value is X. Normalization can make the image data in the same scale, speed up the model training and improve the performance.

[0065] In the encoder, a convolutional neural network is used to extract multi-scale features. The core formula of convolution operation is as follows:

[0066]

[0067] where X: input feature map (for the input layer, it is the preprocessed image data). W is the convolution kernel parameter, used to extract features. b is the bias term, used to adjust the convolution result. K and M are the size of the convolution kernel. σ is the activation function, used to introduce nonlinearity. Y is the output feature map after convolution operation.

[0068] The model evaluation submodule is also set in the deep learning model training module, which is used to evaluate the deep learning model regularly during training. The evaluation indicators include at least one or more combinations of segmentation accuracy, recall rate, F1 value and Dice coefficient.

[0069] The segmentation result generation module is used to input the medical image data to be segmented into the trained deep learning model, and output the segmentation result through the forward propagation calculation of the model. The segmentation result is presented in the form of a binary image or a multi-classification image, clearly indicating the boundaries and regions of different tissues and organs in the medical image.

[0070] Specifically, the deep learning model training module includes a model evaluation submodule that periodically evaluates the model performance during training. Evaluation indicators include segmentation accuracy, recall rate, F1 value, Dice coefficient, etc., which comprehensively reflect the model segmentation capability. Accuracy is the proportion of correctly segmented pixels, recall rate focuses on the recall degree of the target area, F1 value considers both, and Dice coefficient measures the similarity between the segmentation result and the true label. These indicators help optimize model parameters and training strategies to improve segmentation results.

[0071] The segmentation result generation module is used for practical application of the model. The medical image data to be segmented is input into the trained model, and the segmentation result in the form of binary or multi-classification image is output after forward propagation calculation. The output segmentation result clearly identifies the boundaries and regions of the tissues and organs in the medical image. In the binary image, the target area is displayed in white and the background is displayed in black; in the multi-classification image, different colors are used to distinguish multiple tissues and organs. This module realizes the core function of the medical image segmentation system and provides intuitive and reliable image analysis results for doctors.

[0072] The post-processing module further optimizes the segmentation result, including removing isolated points, filling holes, smoothing boundaries, etc.

[0073] The display module is used to visualize the original medical image, preprocessed image, segmentation result, and three-dimensional reconstructed image after segmentation, etc. The display module supports multiple visualization display modes, including two-dimensional slice display, three-dimensional reconstruction display, and fusion display, etc. Users can choose different display modes according to their needs to observe and analyze the medical image and its segmentation result from all angles.

[0074] The storage module is used to store the original medical image data, preprocessing parameters, deep learning model parameters, segmentation results, and related intermediate data. The storage module adopts a distributed storage architecture, which can efficiently store and manage large-scale medical image data and model parameters, support fast read-write and query operations, and has data backup and recovery functions to ensure data security and reliability.

[0075] Specifically, the storage module is responsible for storing the original medical image data, preprocessing parameters, deep learning model parameters, segmentation results, and related intermediate data. It adopts a distributed storage architecture, which can efficiently store and manage large-scale medical image data and model parameters, support fast read-write and query operations, and has data backup and recovery functions to ensure data security and reliability, ensuring stable operation of the system. This module provides comprehensive data support and protection for the entire medical image segmentation system.

[0076] The specific implementation of the deep learning network structure is: the convolutional layer adopts a multi-layer convolutional neural network, each layer of the convolutional neural network includes a plurality of convolutional kernels, which are used to extract different scale and level features of the image data, the pooling layer adopts a maximum pooling or average pooling manner, which is used to reduce the spatial dimension of the feature map, reduce the calculation amount and extract the main features, the full connection layer maps the feature map to an output space, which is used to generate the final segmentation result, the encoder part of the U-Net architecture or its variant includes a plurality of convolutional layers and pooling layers, which are used to extract high-level semantic features of the image, and the decoder part includes a plurality of convolutional layers and up-sampling layers, which are used to fuse the high-level semantic features and low-level features (such as edge, texture and other detail features) to realize accurate positioning and segmentation of the target region.

[0077] Specifically, in the U-Net architecture or its variant, the encoder part is composed of a plurality of convolutional layers and pooling layers, and its task is to gradually extract high-level semantic features of the image, which can provide overall information and contextual relationship about the target region in the image. The decoder part contains a plurality of convolutional layers and up-sampling layers, and the role of the up-sampling layer is to gradually restore the feature map to a spatial dimension close to the original image, while fusing the high-level semantic features from the encoder with low-level features (such as edge, texture and other detail features), so as to realize accurate positioning and segmentation of the target region, and ensure that the segmentation result is not only accurate in position, but also clear in boundary and rich in details. Such network structure design enables the deep learning model to effectively learn the feature representation of different tissues and organs in medical images, and achieve high-precision segmentation effect.

[0078] A medical image segmentation method based on deep learning includes the following steps:

[0079] Step one: obtaining original medical image data, and using a data preprocessing module to perform normalization, denoising and smoothing and other preprocessing operations to improve image quality; after step one, a data enhancement step is also included, and the data enhancement method includes one or more combinations of random rotation, random translation, random scaling, random flipping and adding Gaussian noise, so as to expand the scale of the training data set and improve the generalization ability and adaptability of the deep learning model to different image conditions by performing data enhancement operations on the original medical image data.

[0080] Specifically, after obtaining the original medical image data, preprocessing operations including normalization, denoising and smoothing are required to improve the image quality and provide a better data basis for subsequent segmentation operations. At the same time, a data enhancement step is added after the preprocessing step, using methods such as random rotation, translation, scaling, flipping and adding Gaussian noise, which can increase data diversity, expand the scale of the training data set, help the model adapt to different image conditions, improve its generalization ability, and make the model perform more stably when facing various actual medical images.

[0081] Step two: build a deep learning network structure, which at least includes one convolutional layer, one pooling layer and one fully connected layer. The deep learning model is trained using pre-processed medical image data. By learning a large amount of labeled and unlabeled medical image data, the model automatically learns the feature representation of different tissues and organs in the medical image. The deep learning network structure adopts U-Net architecture or its variants, which realizes multi-scale feature extraction of medical images through upsampling and downsampling. The high-level semantic features are extracted in the downsampling process, and the high-level semantic features and low-level features are fused in the upsampling process, so as to more accurately locate and segment the target region in the medical image, and improve the segmentation accuracy and detail performance. At the same time, the deep learning model is also regularly evaluated, and the evaluation indexes at least include one or more combinations of segmentation accuracy, recall rate, F1 value and Dice coefficient. According to the change trend of the evaluation indexes, the training parameters and optimization strategy of the model are adjusted.

[0082] Specifically, building a deep learning network structure is the core link of the medical image segmentation system. The network structure at least includes one convolutional layer, one pooling layer and one fully connected layer. The convolutional layer can automatically extract features in the medical image, such as edges, textures, etc.; the pooling layer is used to reduce the spatial dimension of the feature map, reduce the amount of calculation, and enhance the robustness of the model to image translation, rotation and other changes; the fully connected layer integrates the extracted features and outputs the final segmentation result. By training the deep learning model using pre-processed medical image data, the model can learn the feature representation of different tissues and organs from a large amount of labeled and unlabeled data, thereby realizing accurate segmentation of medical images.

[0083] The network structure adopting U-Net architecture or its variants can further improve the segmentation effect. U-Net architecture realizes multi-scale feature extraction through upsampling and downsampling. In the downsampling process, the model gradually extracts high-level semantic features of the image, such as the overall shape and position of the tissue and organ; in the upsampling process, the high-level semantic features are fused with low-level features (such as edge, texture and other detail features), so as to more accurately locate and segment the target region in the medical image, and improve the accuracy and detail performance of the segmentation. This fusion mechanism makes the segmentation result not only has accurate boundary, but also retains rich detail information, providing more reliable basis for medical diagnosis.

[0084] Meanwhile, to ensure the performance and reliability of the model, the deep learning model is periodically evaluated during the training process. Evaluation indicators include segmentation accuracy, recall rate, F1 value, and Dice coefficient, etc. Segmentation accuracy measures the proportion of correctly segmented pixels to the total pixels; recall rate reflects the coverage of the target area by the model; F1 value is the harmonic mean of accuracy and recall, considering the balance of the two; Dice coefficient is used to measure the similarity between the segmentation result and the true label. By analyzing the trend of these evaluation indicators, the training status of the model can be understood in a timely manner. If some indicators show a decline or fluctuation, the training parameters and optimization strategies of the model can be adjusted accordingly, such as adjusting the learning rate, adding regularization terms, etc., to avoid overfitting or underfitting phenomena, and ensure that the model can stably output high-quality segmentation results when processing different medical images.

[0085] Step three: input the medical image data to be segmented into the trained deep learning model, and output the segmentation result through the forward propagation calculation of the model. The segmentation result is presented in the form of a binary image or a multi-classification image, indicating the boundaries and regions of different tissues and organs in the medical image;

[0086] Step four: use the post-processing module to optimize the segmentation result, including removing isolated points, filling holes, smoothing boundaries, etc., to improve the accuracy and integrity of the segmentation result; the specific optimization method of the post-processing module is:

[0087] The step of removing isolated points is performed on two-dimensional slices and three-dimensional volume data respectively to ensure the connectivity of the segmentation result in different dimensions;

[0088] The step of filling holes sets different filling strategies according to the morphological characteristics of the target area. For regular-shaped target areas, overall filling is adopted, and for irregular-shaped target areas, local filling is adopted;

[0089] The step of smoothing the boundary sets the intensity of Gaussian filtering according to the detail requirements of the segmentation result. For segmentation results that need to retain small structures, lower intensity filtering is adopted;

[0090] Specifically, the step of removing isolated points is performed on two-dimensional slices and three-dimensional volume data respectively to ensure the connectivity of the segmentation result in different dimensions. In two-dimensional slices, the isolated points that are not connected with the surrounding pixels are identified by the connected region labeling algorithm, and are labeled as background, thereby eliminating the scattered pixel points of false segmentation. In three-dimensional volume data, similar connectivity analysis methods are used to identify and remove isolated voxels from the three-dimensional space, ensuring the connectivity and integrity of the target area in three-dimensional structure.

[0091] The step of filling the holes sets different filling strategies according to the morphological characteristics of the target region. For a regular-shaped target region, a whole filling method is adopted, that is, the morphological closing operation and other technologies are used to completely fill the hole region, so that the contour of the target region is more complete. For an irregularly shaped target region, a local filling strategy is adopted, and the holes are filled according to the local shape and structural characteristics of the target region, so as to avoid distortion of the segmentation result caused by excessive filling.

[0092] The step of smoothing the boundary sets the intensity of the Gaussian filter according to the detail requirements of the segmentation result. Gaussian filter is a commonly used smoothing technique, which smoothes the image boundary by convolution operation with Gaussian kernel. For the segmentation result that needs to retain small structures, a lower intensity filter is used, that is, a smaller standard deviation Gaussian kernel is used to smooth the boundary while retaining the detailed features of the target region as much as possible. For the case where the detail requirement is low, the filter strength is appropriately increased, and a larger standard deviation Gaussian kernel is used to obtain a smoother boundary.

[0093] Step five: the original medical image, the preprocessed image, the segmentation result and the three-dimensional reconstructed image after segmentation are visualized and displayed through the display module, which assists doctors or researchers in observing and analyzing medical images.

[0094] Specifically, in the two-dimensional slice display mode, the user can view the image at any slice position to observe the cross-sectional characteristics of the target region in detail. Three-dimensional reconstruction display allows the user to observe the three-dimensional structure of the target region from all directions and understand its spatial morphology and positional relationship. The fusion display mode combines the original image and the segmentation result, so that the user can refer to the information of both at the same time, and further verify the accuracy of the segmentation.

[0095] Through these diversified visualization display modes, doctors can more accurately diagnose diseases, such as determining the location, size and shape of tumors; researchers can more deeply analyze image data, and promote the development of medical image analysis technology.

[0096] In summary: the present application adopts the deep learning network structure of U-Net architecture or its variants, realizes multi-scale feature extraction, down-sampling to extract high-level semantic features, and up-sampling to fuse high and low-level features, so as to more accurately locate and segment the target region in the medical image, clearly mark the boundaries and regions of tissues and organs, provide more reliable basis for medical diagnosis, and meet the high requirements of clinical segmentation accuracy.

[0097] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A medical image segmentation system based on deep learning, characterized in that: include: The data preprocessing module is used to receive raw medical image data and perform preprocessing operations such as normalization, denoising and smoothing to improve image quality, reduce noise interference and enhance image analyzability; A deep learning model training module, including a constructed deep learning network structure, which includes at least one convolutional layer, one pooling layer, and one fully connected layer, for extracting and learning features from preprocessed medical image data. By training on a large amount of labeled and unlabeled medical image data, the model can automatically learn the feature representations of different tissues and organs in medical images to achieve accurate segmentation of medical images; Among them, the specific implementation method of the deep learning network structure is: the convolution layer adopts a multi-layer convolutional neural network, each layer of the convolutional neural network includes multiple convolution kernels, which are used to extract features of different scales and levels of image data; the pooling layer adopts maximum pooling or average pooling to reduce the spatial dimension of the feature map, reduce the amount of calculation and extract the main features; the fully connected layer maps the feature map to the output space to generate the final segmentation result; the encoder part of the U-Net architecture or its variant includes multiple convolution layers and pooling layers to extract high-level semantic features of the image; the decoder part includes multiple convolution layers and upsampling layers to fuse high-level semantic features with low-level features to achieve accurate positioning and segmentation of the target area; The segmentation result generation module is used to input the medical image data to be segmented into the trained deep learning model, and after the forward propagation calculation of the model, output the segmentation result. The segmentation result is presented in the form of a binary image or a multi-classification image, clearly marking the boundaries and regions of different tissues and organs in the medical image; The post-processing module further optimizes the segmentation results, including removing isolated points, filling holes, smoothing boundaries, etc. The display module is used to visualize the original medical images, pre-processed images, segmentation results, and 3D reconstructed images after segmentation; The storage module is used to store raw medical image data, preprocessing parameters, deep learning model parameters, segmentation results, and related intermediate data.

2. A medical image segmentation system based on deep learning according to claim 1, characterized in that: The deep learning network structure in the deep learning model training module adopts the U-Net architecture or its variant, which realizes multi-scale feature extraction of medical images through up and down sampling. During the downsampling process, high-level semantic features of the image are gradually extracted, and during the upsampling process, high-level semantic features are fused with low-level features, thereby more accurately locating and segmenting the target area in the medical image, improving the segmentation accuracy and detail representation ability.

3. The deep learning-based medical image segmentation system according to claim 1, characterized in that: The data preprocessing module also includes a submodule for performing data enhancement on medical imaging data. The data enhancement method includes at least one or more combinations of random rotation, random translation, random scaling, random flipping and adding Gaussian noise. By performing data enhancement operations on the original medical imaging data, the scale of the training data set is expanded.

4. The deep learning-based medical image segmentation system according to claim 1, characterized in that: The deep learning model training module is also provided with a model evaluation submodule for regularly evaluating the deep learning model during the training process. The evaluation indicators include at least one or more combinations of segmentation accuracy, recall rate, F1 value and Dice coefficient.

5. The deep learning-based medical image segmentation system according to claim 1, characterized in that: The storage module adopts a distributed storage architecture, which can efficiently store and manage large-scale medical imaging data and model parameters, support fast data reading, writing and query operations, and has data backup and recovery functions to ensure data security and reliability.

6. The deep learning-based medical image segmentation system according to claim 1, characterized in that: The display module supports multiple visualization display modes, including two-dimensional slice display, three-dimensional reconstruction display, and fusion display. Users can select different display modes according to their needs to observe and analyze medical images and their segmentation results in an all-round and multi-angle manner.

7. A medical image segmentation method based on deep learning A medical image segmentation method based on deep learning, according to a medical image segmentation system based on deep learning according to any one of claims 1 to 6, characterized in that: The following steps are involved: Step 1: Obtain the original medical image data and use the data preprocessing module to perform preprocessing operations such as normalization, denoising, and smoothing to improve image quality; Step 2: Construct a deep learning network structure, which includes at least one convolutional layer, one pooling layer, and one fully connected layer. Use preprocessed medical imaging data to train the deep learning model. By learning from a large amount of labeled and unlabeled medical imaging data, the model automatically learns the characteristic representations of different tissues and organs in medical images. Step 3: Input the medical image data to be segmented into the trained deep learning model. After the model's forward propagation calculation, the segmentation result is output. The segmentation result is presented in the form of a binary image or a multi-classification image, marking the boundaries and regions of different tissues and organs in the medical image. Step 4: Use the post-processing module to optimize the segmentation results, including removing isolated points, filling holes, smoothing boundaries, etc., to improve the accuracy and completeness of the segmentation results. The specific optimization methods of the post-processing module are as follows: The step of removing isolated points is performed on 2D slices and 3D volume data separately to ensure the connectivity of segmentation results in different dimensions; The step of filling holes sets different filling strategies according to the morphological characteristics of the target area. For regular-shaped target areas, overall filling is used, while for irregular-shaped target areas, local filling is used. The step of smoothing the boundary sets the intensity of the Gaussian filter according to the detail requirements of the segmentation result. For segmentation results that need to retain fine structures, a lower intensity filter is used; Step 5: The original medical image, pre-processed image, segmentation result and segmented 3D reconstructed image are visualized through the display module to assist doctors or researchers in observing and analyzing medical images.

8. The medical image segmentation method based on deep learning according to claim 7, characterized in that: In step 2, the deep learning network structure adopts the U-Net architecture or its variant, and realizes multi-scale feature extraction of medical images through up and down sampling. The downsampling process extracts high-level semantic features, and the upsampling process fuses high-level semantic features with low-level features, thereby more accurately locating and segmenting the target area in the medical image, improving the segmentation accuracy and detail representation ability.

9. The medical image segmentation method based on deep learning according to claim 7, characterized in that: After step one, a data enhancement step is also included. The data enhancement method includes at least one or more combinations of random rotation, random translation, random scaling, random flipping and adding Gaussian noise. By performing data enhancement operations on the original medical image data, the scale of the training data set is expanded, and the generalization ability of the deep learning model and its adaptability to different imaging conditions are improved.

10. The medical image segmentation method based on deep learning according to claim 7, characterized in that: In step 2, the deep learning model is also evaluated regularly. The evaluation indicators include at least one or more combinations of segmentation accuracy, recall rate, F1 value and Dice coefficient. According to the changing trend of the evaluation indicators, the training parameters and optimization strategy of the model are adjusted.

Images