A method for mitochondria segmentation and classification based on deep learning

Abstract

The application provides a kind of deep learning-based mitochondrion segmentation and classification method, solve the robustness and accuracy lower of existing technical method and model for mitochondrion morphological analysis processing, and the problem of qualitative and quantitative analysis cannot be solved, qualitative and quantitative analysis of mitochondrion super-resolution image is realized by the "segmentation-classification" idea of image segmentation and image classification combination, wherein, image segmentation part embodies quantitative, deep learning can quantify the morphological profile of mitochondrion in each image, and the fluorescence intensity of mitochondrion false color image is quantified.Image classification part embodies qualitative and quantitative, deep learning can classify mitochondrion of different morphologies to qualitatively (circular or filamentous), and the proportion of circular mitochondrion is counted to quantitatively the overall pathological degree of mitochondrion in the current cell.

Description

Technical Field

[0001] This invention belongs to the field of biomedical image processing technology, specifically relating to a segmentation and classification method for mitochondrial super-resolution images based on deep learning. Background Technology

[0002] Mitochondria are among the most important organelles in eukaryotic cells. They perform vital cellular functions by producing adenosine triphosphate (ATP), earning them the nickname "the cell's power plant." Simultaneously, they play a crucial role in cellular life processes, including the regulation of disease states. Significant heterogeneity exists in the distribution of mitochondria within cells, and variations in mitochondrial morphology are associated with multiple important cellular functions and cell death pathways. The morphological characteristics of mitochondria can evolve from complex, branched filamentous structures to round structures. This division process is associated with enhanced phagocytosis, increased uncoupling, and increased production of reactive oxygen species (ROS). Changes in mitochondrial morphology and structure may predict early events in tumor progression and the occurrence of cancer; furthermore, fragmented mitochondria after division can lead to the proliferation and metastasis of cancer cells. This suggests that mitochondrial phenotype and morphology may serve as important evidence for cancer diagnosis, analysis of cancer cell metabolic status, and the identification of antitumor therapeutic targets and responses.

[0003] The structure of mitochondria in living cells is highly dynamic. The point spread function of a microscope causes three-dimensional blurring, resulting in defocused structures appearing at the focal point with varying intensities and blurred outlines, leading to poor signal-to-noise ratios in the images. Furthermore, the variability of mitochondrial motion and the high-intensity spots produced by multiple overlapping structures further complicate the information contained in the images. These factors make it difficult to directly visualize and analyze intracellular mitochondrial structures after super-resolution fluorescence imaging, necessitating image segmentation of mitochondria and quantification of their spatial distribution and morphology. However, manually segmenting the numerous small and dynamic mitochondria within cells is an extremely time-consuming and labor-intensive task. Moreover, the large number and diverse morphologies of mitochondria in images make it difficult to subjectively quantify their health status, and manually analyzing massive amounts of biomedical images is both time-consuming and labor-intensive.

[0004] In recent years, researchers have used various automated image processing and analysis techniques to segment, quantify, and analyze mitochondria in a large number of generated microscopic images. However, there is still a lack of a well-developed method to perform qualitative and quantitative analysis of mitochondrial super-resolution images to explore the relationship between morphological changes and cellular life activities, and to assist in the research on the prevention, diagnosis, and treatment of organelle dysfunction-related diseases (such as tumors, diabetes, and neurodegenerative diseases). Summary of the Invention

[0005] Deep learning algorithms have achieved great success in biomedical image analysis. Deep learning builds neural networks similar to the human brain's analytical learning mechanisms and mimics these mechanisms to interpret data. This characteristic is particularly important for biomedical images, which often exhibit complex textures and significant variations due to different cellular features and batch variations in sample processing. However, the feature extraction process of deep learning is a "black box problem"—the underlying data and code logic is incomprehensible to humans, making it somewhat uninterpretable.

[0006] In view of this, in order to solve the problems of low robustness and accuracy of existing techniques and models for mitochondrial morphological analysis, as well as the inability to perform qualitative and quantitative analysis, this invention explores a segmentation and classification method for mitochondrial super-resolution images based on deep learning, aiming to improve the work efficiency and accuracy of medical staff and scientific and technological workers.

[0007] To achieve the above objectives, the technical solution provided by this invention is:

[0008] A method for establishing a segmentation and classification model for mitochondrial super-resolution images based on deep learning is characterized by the fact that the model includes an image segmentation model and an image classification fusion model from input to output.

[0009] The image segmentation model is an improved version of U-Net, U-RNet+.

[0010] The image classification and fusion model includes an image classification model and an image fusion model set sequentially; wherein, the image classification model is the improved CNN-11 model; and the image fusion model is XGBoost.

[0011] This model is established through the following steps:

[0012] S1: Obtain training sample data

[0013] Mitochondrial super-resolution fluorescence images were captured using structured illumination microscopy (SIM) for training image segmentation models. As a type of super-resolution imaging technology, structured illumination super-resolution fluorescence microscopy has the advantages of fast imaging speed and low phototoxicity. Moreover, its resolution of hundreds of nanometers is just right to meet the observation needs of important intracellular organelles, and it is known as the most suitable super-resolution imaging technology for subcellular structures (such as mitochondria).

[0014] S2: Image preprocessing and annotation for image segmentation model training

[0015] After cropping and filtering the mitochondrial super-resolution fluorescence image obtained from S1, the mitochondrial super-resolution fluorescence image is labeled, including the region of mitochondria in the image and their specific outline.

[0016] S3: Training of Deep Learning-Based Image Segmentation Model

[0017] The image segmentation model was trained using the S2-annotated dataset. The input to the training process was the annotated super-resolution fluorescence image of mitochondria. The learning objective was the spatial distribution and shape contour of mitochondria. The output was a binarized segmented image containing the spatial distribution and shape contour of mitochondria, and a quantifiable pseudo-color image of mitochondria that takes into account both the color information of the original image and the contour information of the segmented image was obtained.

[0018] S4: Image preprocessing and annotation for image classification models

[0019] The binarized segmented images obtained by S3 were filtered and expanded, and mitochondria of different morphologies were classified into corresponding tag folders for labeling.

[0020] S5: Training of Image Classification Model Based on Deep Learning

[0021] The image classification model is trained using the S4-annotated dataset. The input to the training process is the binarized segmented image obtained by S4 preprocessing and S3 annotation. The learning target is the shape category of mitochondria, and the output is the prediction and score of the mitochondrial category.

[0022] S6: Training of Image Classification and Fusion Model Based on Deep Learning

[0023] The image fusion model is trained. The input to the training process is the predicted mitochondrial category and score obtained by the S5 image classification model and the morphological parameter index of the mitochondrial image obtained by the image processing algorithm. The learning target is the shape category of mitochondria, and the output is the predicted mitochondrial category and score.

[0024] That is, the output of S5 is used as one of the inputs to S6 for further learning (secondary learning). Since the feature extraction process using deep learning technology in the CNN-11 classification model is a "black box problem"—the underlying data and code logic is incomprehensible and lacks interpretability—this invention utilizes the XGBoost model to perform multi-indicator morphological analysis of mitochondria, thereby assisting deep learning in related applications. The fitting effect of this classification fusion model is better than using either the CNN-11 model or the XGBoost model alone. The CNN-11 model simulates brain function, representing "emotional thinking," while the XGBoost model uses mathematical statistics, representing "rational analysis." In practical applications, combining "emotional thinking" with "rational analysis" often yields better results.

[0025] S7: Obtain the segmentation and classification model of mitochondrial super-resolution images that has been built and trained based on deep learning.

[0026] Furthermore, the improved model U-RNet+ includes 104 convolutional layers and 16 pooling layers, and the model is divided into a downsampling part in the first half and an upsampling part that follows immediately.

[0027] The downsampling part is a typical convolutional neural network structure, which is repeated 5 times using a combination of two convolutional layers and one pooling layer. After each convolutional layer, the structure of the residual unit is added and improved. An improved residual unit consists of an improved convolutional residual block and two improved identity residual blocks. In the improved convolutional residual block, the convolutional layers with a stride of 2 in the main path and the side path are replaced with a combination of a convolutional layer with a stride of 1 and a pooling layer. In the improved identity residual block, the input layer of the side path is replaced with a convolutional layer with a stride of 1.

[0028] The upsampling part first performs a transposed convolution operation, then reassembles a new feature map by splicing the feature maps of the same dimension of the corresponding left channel, and then uses two convolutional layers to extract features. This structure is repeated 4 times.

[0029] Furthermore, due to the increasing skip connection span in the U-RNet+ model, an attention mechanism (Attention Module) is added to the outermost feature map in the Convolutional Block Attention Module (CBAM), i.e., the spatial and channel dimensions. In the final output layer, the feature map is mapped to the output image.

[0030] The improved CNN-11 model includes 5 convolutional layers, 3 pooling layers, 2 fully connected layers, and 1 softmax regression layer; the order of these layers is as follows: 1. Convolutional layer, 2. Pooling layer, 3. Convolutional layer, 4. Pooling layer, 5. Convolutional layer, 6. Convolutional layer, 7. Convolutional layer, 8. Pooling layer, 9. Fully connected layer, 10. Fully connected layer, 11. Softmax regression layer.

[0031] Convolutional layers are used to reduce image noise and extract image features.

[0032] Pooling layers are used to reduce the dimensionality of features, remove redundant information, and maintain the translation invariance, rotation invariance, and scale invariance of the feature image.

[0033] Fully connected layers are used to integrate the feature space mappings calculated by previous layers into value outputs;

[0034] The Softmax regression layer then classifies the output values;

[0035] The feature vectors used by the image fusion model XGBoost include the predicted mitochondrial category and score obtained through CNN-11 and the morphological parameter indicators of mitochondrial images obtained based on image processing algorithms.

[0036] Furthermore, the conditions for training the improved model U-RNet+ are as follows:

[0037] The programming language is Python, the high-level neural network API used is Keras, and a virtual environment mitounet is configured to implement this task. The main Python toolkits and their corresponding versions used are libtiff=4.2.0, matplotlib=3.0.3, opencv=4.5.4, pandas=1.2.4, numpy=1.19.5, keras-preprocessing=1.1.2, tensorflow=2.5.0, and tensorflow-gpu=2.5.0.

[0038] The hyperparameters are as follows: activation function is ReLU, optimizer is Adam, loss function is binary cross-entropy, batch size is 2, and learning rate is 1×10⁻⁶. -4 The number of iterations is 0.5 × 10. 3 Second-rate;

[0039] The conditions for training the improved CNN-11 model are as follows:

[0040] The programming language is Python, the high-level neural network API used is Tensorflow, and a virtual environment miconn is configured to implement this task. The main Python toolkits and their corresponding versions used are matplotlib=3.3.4, numpy=1.19.0, pandas=1.1.5, opencv=4.5.5, tensorflow=1.9.0, and tensorflow-gpu=1.9.0.

[0041] The hyperparameters are as follows: activation function is Sigmoid, optimizer is Adam, loss function is MeanSquared Error, batch size is 32, and learning rate is 1×10⁻⁶. -5 The number of iterations is 1×10 5 Second-rate.

[0042] The training parameters of both models can be adjusted and updated, but the parameters mentioned above are currently the optimal ones. During model training, the goal is model convergence, and the evaluation metric is the loss function in the improved CNN-11 hyperparameters. When the loss function value on the model training set decreases to a stable level with minimal fluctuations, it indicates that the model has reached convergence, i.e., it is well trained.

[0043] The conditions for implementing the image processing algorithm to obtain mitochondrial morphological parameters are as follows:

[0044] The programming language is Python, and the main Python packages and their corresponding versions used are numpy=1.18.0, matplotlib=3.5.2, opencv=3.4.2, scikit-image=0.19.2, scipy=1.4.1, mkl-random=1.2.2, backport.shutil-which=3.5.2, pandas=1.13.0, and request=2.27.1.

[0045] The training conditions for the image fusion model XGBoost are as follows:

[0046] The programming language is Python, and the machine learning library used is Scikit-Learn. A virtual environment, mitoxgb, was configured to implement this task, with the main Python packages and their corresponding versions being xgboost=1.5.0, matplotlib=3.5.2, numpy=1.21.5, xlwt=1.3.0, pandas=1.4.4, and scikit-learn=1.1.1.

[0047] The specific parameters are: booster = gbtree, eta = 0.3, max_depth = 5, min_child_weight = 3, gamma = 0.1, subsample = 0.7, lambda = 3, objective = reg:gamma, colsample_bytree = 0.7, seed = 0, nthread = 4, and the number of iterations is 5 × 10^6. 3 .

[0048] Furthermore, the specific steps of S1 are as follows:

[0049] Mitochondria within cells are stained and prepared as fluorescent samples (i.e., fluorescent dye staining method). Mitochondria are then imaged using a SIM microscopy device, and at least 1000 super-resolution fluorescent images of mitochondria at a resolution of 2048×2048 are captured.

[0050] Furthermore, step 2) specifically involves:

[0051] S2.1 uses Python programming to batch crop the mitochondrial super-resolution fluorescence images obtained in S1 into several 512×512 resolution image patches according to a sliding window of fixed pixel size (cropping the original image serves two purposes: 1. It increases the total number of training images while reducing the data labeling workload of a single image, which is beneficial for model training; 2. Invalid background information and noise account for a certain proportion in a single image, and cropping and filtering them helps to ensure the quality of training data and the information content of a single image. Using a cropping tool can greatly improve processing efficiency). Image patches with high proportions of invalid background information and noise are filtered out, and the remaining high-quality image patches with morphological representativeness form the dataset for training the image segmentation model, and are divided into training set, validation set and test set in a ratio (4:1:5).

[0052] S2.2 Use Labelme image annotation software to manually annotate the mitochondrial contours of the training and validation sets. After annotation, generate the corresponding number of JSON files; for example, after manually annotating 400 training images and 100 validation images, generate 500 corresponding JSON files.

[0053] S2.3 uses the written test.bat script to extract information from JSON files in batches, including the original image, labeled image, label name, and YAML file representing the labeled content; in the labeled image, black represents the background and white represents the mitochondrial outline of the label.

[0054] S2.4 Use Python programming to extract the original and labeled images one by one in sequence and name them according to specific rules (they can be named in numerical order of 0, 1, 2, 3, 4, 5...), and classify them into the data and label folders respectively.

[0055] Furthermore, the specific steps of S4 are as follows:

[0056] Using Python programming, images with mitochondrial pixel values ​​less than 3% in the binarized segmented images obtained by S3 were screened out. Data augmentation strategies of clockwise rotation and vertical flipping were used to expand the data, and finally, a training set and a test set with expanded data were obtained.

[0057] The annotation process involves naming images of healthy mitochondria that are mostly filamentous and images of diseased mitochondria that are mostly round, and finally classifying them into the Fusion and Fission folders respectively.

[0058] Furthermore, the specific steps of S6 are as follows:

[0059] S6.1 Morphological denoising of mitochondrial segmentation images based on the Opening Area algorithm;

[0060] S6.2 uses image processing algorithms to obtain 102 morphological parameters of mitochondrial images, including: mean mitochondrial area, median area, area standard deviation, mean eccentricity, median eccentricity, eccentricity standard deviation, mean isochoria, median isochoria, isochoria standard deviation, mean Euler number, median Euler number, Euler number standard deviation, mean spread, median spread, spread standard deviation, mean principal axis length, median principal axis length, principal axis length standard deviation, mean secondary axis length, median secondary axis length, secondary axis length standard deviation, mean direction, median direction, direction standard deviation, and mean perimeter. Median perimeter, perimeter standard deviation, mean convexity, median convexity, convexity standard deviation, centroid x-coordinate mean, centroid x-coordinate median, centroid x-coordinate standard deviation, centroid y-coordinate mean, centroid y-coordinate median, centroid y-coordinate standard deviation, distance mean, distance median, distance standard deviation, weighted centroid x-coordinate mean, weighted centroid x-coordinate median, weighted centroid x-coordinate standard deviation, mean weighted centroid y-coordinate, median weighted centroid y-coordinate, weighted centroid y-coordinate standard deviation, mean weighted distance, median weighted distance, weighted distance standard deviation, mean shape factor, median shape factor , morphological factor standard deviation, mean roundness, median roundness, roundness standard deviation, mean branch count, median branch count, branch count standard deviation, mean branch length, median branch length, branch length standard deviation, mean total branch length, median total branch length, total branch length standard deviation, mean median branch length, median median branch length, median branch length standard deviation, branch length standard deviation, mean weighted branch angle, median weighted branch angle, branch angle standard deviation, mean median branch angle, median median branch angle, median branch angle standard deviation, mean branch angle standard deviation, median Branch angle standard deviation, branch angle standard deviation standard deviation, mean total density, median total density, total density standard deviation, mean average density, median mean density, mean density standard deviation, mean median density, median median density, median density standard deviation, kurtosis x, weighted kurtosis x, kurtosis y, weighted kurtosis y, kurtosis squared, weighted kurtosis squared, skewness x, weighted skewness x, skewness y, weighted skewness y, skewness squared, weighted skewness squared, network orientation (degrees), network principal axis (pixels), network secondary axis (pixels), network eccentricity, network effective range, network effective solidity, network fractal dimension;

[0061] S6.3 inputs the mitochondrial morphological parameters obtained above and the results output from S5 into the XGBoost model for analysis to obtain the prediction and score of mitochondrial category.

[0062] Meanwhile, this invention provides a mitochondrial segmentation and classification model based on deep learning, which is unique in that it is built and trained using the above-mentioned method.

[0063] Furthermore, a deep learning-based method for mitochondrial segmentation and classification is characterized by: using a model established by any one of the methods in claims 1-7 for mitochondrial segmentation and classification, inputting the super-resolution fluorescence image of the mitochondria to be processed into the model, and obtaining the prediction and score of the mitochondrial category.

[0064] A computer-readable storage medium and electronic device storing a computer program thereon, characterized in that: when the computer program is executed by a processor, it implements the steps of the above-described method.

[0065] The principle of this invention:

[0066] This invention uses a SIM microscope to capture and acquire a large number of mitochondrial super-resolution fluorescence microscopic images; preprocesses the images using Python programming and labels them using Labelme software, which serves as the dataset for model training; builds and trains an improved U-Net model, U-RNet+, for morphological segmentation of mitochondria in super-resolution fluorescence images, generating corresponding pseudo-color images, and quantifying their fluorescence intensity; based on the segmented images, builds and trains an improved CNN model, CNN-11, and further optimizes the model by fusing it with an XGBoost model for morphological classification of mitochondrial segmentation images.

[0067] Compared with the prior art, the present invention has the following advantages:

[0068] 1. This invention achieves qualitative and quantitative analysis of mitochondrial super-resolution images through a "segmentation-classification" approach combining image segmentation and image classification. The image segmentation part embodies quantitative analysis; deep learning can quantify the morphological contours of mitochondria in each image and the fluorescence intensity of the mitochondrial pseudo-color image. The image classification part embodies both qualitative and quantitative analysis; deep learning can classify mitochondria of different morphologies (round or filamentous) qualitatively and statistically analyze the proportion of round mitochondria to quantify the overall mitochondrial pathology within the cell. Specifically, this invention integrates the U-Net model, an improved residual network, and an attention mechanism to achieve mitochondrial super-resolution image segmentation. It uses Python programming and the U-RNet+ model to generate batches of mitochondrial pseudo-color images that incorporate both the original image's color information and the segmented image's contour information. Furthermore, it employs a method that integrates a CNN model and an XGBoost model to classify mitochondrial segmented images. This method offers a degree of interpretability, and the use of tools during segmentation significantly improves both efficiency and accuracy.

[0069] 2. The technical solution provided by this invention solves a long-standing technical problem that has remained unsolved: This invention improves the segmentation performance of mitochondrial super-resolution image segmentation, demonstrating superior performance in contour information preservation, noise removal, and contour clarity and continuity. It can also correctly segment mitochondria with special shapes (such as rings). This invention can process thousands of images in a single run, with an average processing time of only about 0.6 seconds per image, exhibiting high throughput and speed.

[0070] 3. The segmentation and classification of mitochondrial super-resolution images achieved using U-RNet+ and CNN-11+XGBoost in this invention has high application potential and value. The fitting effect of the CNN-11+XGBoost fusion model is better than that of using the CNN-11 model or XGBoost model alone. The CNN-11 model simulates brain function, representing "intuitive thinking," while the XGBoost model uses mathematical and statistical methods, representing "rational analysis." In practical applications, combining "intuitive thinking" and "rational analysis" often achieves better results.

[0071] 4. The results of predicting and assessing the degree of mitochondrial damage using this invention are in good agreement with the results of relevant biological testing experiments. This indicates that this invention can assist in biological experiments, thereby reducing a large number of tedious cell experiments. Compared with the current internationally leading mitochondrial image segmentation tools, this invention demonstrates advantages in image segmentation quality, throughput, and efficiency. Furthermore, this invention achieves quantitative and qualitative analysis of mitochondria through classification, and provides an evaluation tool for future cutting-edge biomedical applications such as mitochondrial detection, mitochondrial therapy, and mitochondrial implantation.

[0072] 5. This invention can achieve high-throughput, fast and automated image segmentation, and obtain pseudo-color images that take into account both the color information of the original image and the contour information of the segmented image, as well as a quantitative evaluation of their fluorescence intensity. Attached Figure Description

[0073] Figure 1 Flowchart of the deep learning-based mitochondrial image segmentation and classification method provided by the present invention;

[0074] Figure 2 A flowchart illustrating the training image preprocessing and annotation process for the segmentation model provided in this invention.

[0075] Figure 3 This is a diagram of the U-Net model architecture for mitochondrial super-resolution image segmentation provided by the present invention.

[0076] Figure 4 This refers to the improved residual unit in the improved model U-RNet+ provided by this invention;

[0077] Figure 5 The CBAM attention mechanism in the improved model U-RNet+ provided by this invention;

[0078] Figure 6 A comparison of the segmentation results of the improved model U-RNet+ provided by this invention and the basic U-Net model shows that the improved model U-RNet+ is superior to the U-Net model in terms of contour information preservation, noise removal, contour clarity and continuity.

[0079] Figure 7 The present invention provides a pseudo-color image that takes into account both the color information of the original image and the contour information of the segmented image, and a quantitative evaluation of its fluorescence intensity.

[0080] Figure 8 A flowchart illustrating the training image preprocessing and annotation process for the classification model provided in this invention.

[0081] Figure 9 A diagram of the CNN-11 model architecture for mitochondrial segmentation image classification provided by this invention;

[0082] Figure 10 The present invention provides an image processing algorithm for obtaining mitochondrial morphological indicators and a CNN-11 and XGBoost fusion algorithm.

[0083] Figure 11 A comparison between the predicted fitted curve of the drug action dose relationship provided by the present invention and the actual MMP damage;

[0084] Figure 12 This invention provides a comparison between the predicted fitting curves based on the drug action time relationship and the actual MMP damage. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0086] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0087] A deep learning-based segmentation and classification method for mitochondrial super-resolution images, such as... Figure 1 As shown, the process includes Part 1: Dataset Creation Steps; Part 2: Model Building and Training Steps; Part 3: Model Application Steps; and Part 4: Application Evaluation Steps.

[0088] Part 1 Dataset Creation Steps, such as Figure 2 As shown, it specifically includes the following three steps:

[0089] S1: Intracellular mitochondria were stained with fluorescent dyes and other methods to prepare fluorescent samples. The mitochondria were then imaged using a SIM microscopy system, with a total of approximately 1.1 × 10⁻⁶ images collected. 3 A 2048×2048 resolution super-resolution fluorescence image of mitochondria.

[0090] S2: Using Python programming, batch-process mitochondrial super-resolution fluorescence images are cropped into several 512×512 resolution image blocks using a sliding window of fixed pixel size. Invalid background information and image blocks with high noise levels are then removed, leaving only 1×10... 3 The dataset for the image segmentation model consists of high-quality image patches that are morphologically representative, and is divided into a training set of 400 images, a validation set of 100 images, and a test set of 500 images.

[0091] S3: Manually annotate mitochondrial contours on 400 training images and 100 validation images using Labelme image annotation software, totaling over 1×10⁻⁶ annotations. 5 Each mitochondrial is labeled, generating 500 corresponding JSON files. A written script, test.bat, is used to batch extract information from these JSON files, including the original image, labeled image, label name, and a YAML file representing the labeled content. In the labeled image, black represents the background, and white represents the outline of the labeled mitochondria. Python is used to extract the original and labeled images sequentially, name them according to specific rules, and categorize them into the `data` and `label` folders.

[0092] Part 2, model building and training, mainly includes the following four steps:

[0093] S1: Train the improved U-Net model U-RNet+ on the dataset created in Part 1. The input to the training process is image data, the learning target is the spatial distribution and shape contour of mitochondria, and the output is a binarized segmented image containing the spatial distribution and shape contour of mitochondria.

[0094] S2: The binary segmented images obtained through the U-RNet+ model are automatically filtered and augmented using Python programming to achieve data augmentation, and different mitochondria are classified into corresponding tag folders.

[0095] S3: The improved CNN model CNN-11 is trained using the dataset labeled by S2. The input to the training process is the dataset created by S2 (binarized segmentation image results obtained by the image segmentation model). The learning objective is the shape category of mitochondria, and the output is the prediction and score of the mitochondrial category.

[0096] S4: The XGBoost model is used for training. The input to the training process is the predicted mitochondrial category and score obtained by the S3 image classification model and the morphological parameters of the mitochondrial image obtained by the image processing algorithm. The learning objective is the shape category of the mitochondria, and the output is the predicted mitochondrial category and score.

[0097] specific:

[0098] The base model of the improved U-RNet+ image segmentation model in S1 is U-Net, such as... Figure 3 As shown. The first half (i.e. Figure 3 The downsampling process in the left half (middle part) is a typical convolutional neural network structure, repeated 5 times using a combination of 2 convolutional layers (Conv2D) and 1 pooling layer (Maxpooling). The latter half (i.e....) Figure 3 The upsampling process (right half) first involves transposing a convolution to restore features. Then, skip connections are used to concatenate features of the same dimension from the downsampling process. Finally, two typical convolutional layers are used for feature extraction. This "feature restoration - dimension concatenation - feature extraction" structure is repeated four times.

[0099] The transposed convolution serves to reduce the dimensionality of high-dimensional abstract features, restoring the positional information lost during upsampling, supplementing image details, and gradually restoring the feature map size to its initial size. Since upsampling can only restore the image size and not its details, skip connections are used to fuse multi-scale features. Large-scale features from the upsampling process are concatenated with positional features of the same dimension in the corresponding left channel, increasing the information content of that dimension describing the image. This approach fully considers the positional and dimensional relationships of the image, ensuring spatial consistency in the results. In the final output layer, U-Net discards traditional fully connected layers and uses two convolutional layers to map the feature map into a two-dimensional output map and a one-dimensional matrix.

[0100] The improved U-RNet+ image segmentation model in S1 adds an improved residual network after each convolutional layer of the U-Net base network. Because the base U-Net network has relatively shallow layers, the extracted features are relatively simple. Theoretically, to extract deeper features, the number of network layers needs to be increased, but this leads to the problems of vanishing and exploding gradients. The traditional solution to this problem is data regularization, but this method degrades model performance. The emergence of residual networks helps to solve the problems of vanishing and exploding gradients and maintains good performance even after increasing the number of network layers. An improved residual unit consists of an improved convolutional residual block and two improved identity residual blocks. In the improved convolutional residual block, the convolutional layers with a stride of 2 in both the main and side paths are replaced with a combination of convolutional layers and pooling layers with a stride of 1. In the improved identity residual block, the input layer of the side path is replaced with a convolutional layer with a stride of 1, as shown below. Figure 4 As shown in the diagram, the pooling layer is added to increase the translation invariance of image features, enabling higher-level features to have a larger receptive field while preserving the texture features of the image. Furthermore, the pointwise operation of pooling is more conducive to optimization than the weighted sum of convolutional layers.

[0101] The improved U-RNet+ image segmentation model in S1 adds a CBAM (Convolutional Block Attention Module) mechanism, such as... Figure 5 As shown, the skip connection span of the U-RNet+ model gradually increases with the depth of the network layers. Furthermore, since the outermost skip connection uses the most information from its feature map, it is most prone to errors during dimensional concatenation. Therefore, an attention mechanism (Attention Module) is added to the feature map passed to the outermost skip connection of U-RNet+ in the mixed dimensions, namely the spatial and channel dimensions, to filter out important parts from a large amount of information. CBAM combines channel and spatial attention mechanisms, processing the input features using both mechanisms separately. The channel attention mechanism is implemented by performing global average pooling and global max pooling on the input features. The spatial attention mechanism takes the maximum and average values ​​of the input features at each feature point's channel, stacks them, adjusts the number of channels using convolution, and finally multiplies the weights of each feature point of the obtained input features with the original input feature layer.

[0102] The U-RNet+ model in S1 consists of 104 convolutional layers (Conv2D) and 16 pooling layers (Maxpooling). The model comprises a downsampling section in the first half and an upsampling section in the second half. The downsampling section uses a typical convolutional neural network structure, repeating five times with a combination of two convolutional layers (yellow cuboids) with 3×3 kernels, batch normalization, and ReLU activation, and one maxpooling layer (orange cuboid). The five downsampling processes yield feature maps of sizes 256×256×64, 64×64×128, 16×16×256, 4×4×512, and 2×2×1024, respectively. The upsampling process first performs two transposed convolutions (blue cuboids) and an upsampling module based on bilinear interpolation (orange cuboids). Then, it uses skip connections to concatenate feature maps of the same dimension from the left channel, recovering the spatial information lost during downsampling and reconstructing a new feature map (green cuboid). Two convolutional layers are then used for feature extraction, and this structure is repeated four times. The four upsampling processes yield feature maps of sizes 8×8×512, 32×32×256, 128×128×128, and 512×512×64, respectively. In the final output layer, a convolutional layer with a 1×1 kernel, batch normalization, and a sigmoid activation function is used to obtain an output of size 512×512×1. A softmax layer is used to predict the probabilities of the two labels (mitochondria and background) in the output, and the feature map is mapped to the output image.

[0103] In S1, the U-RNet+ model is trained using ReLU as the activation function, Adam as the optimizer, binary cross-entropy as the loss function, a batch size of 2, and a learning rate of 1×10⁻⁶. -4 The number of iterations is 0.5 × 10. 3 Next, save the weight parameters after model training and use them for the test set.

[0104] In S2, a data augmentation strategy involving clockwise rotation and vertical flipping was implemented using Python programming to augment the data, ultimately yielding approximately 7 × 10⁻⁶ data points. 3 The training set includes images of filamentous mitochondria and round mitochondria, named according to specific rules and categorized into the Fusion and Fission folders, respectively. Figure 6 As shown, data augmentation is a commonly used regularization method in image classification tasks. Its purpose is to increase the size and diversity of the dataset, thereby enhancing the generalization ability of the trained model. Such data preprocessing allows the model to learn more about mitochondrial morphological distribution during training, significantly improving model performance and preventing overfitting.

[0105] The improved CNN-11 model in S3 specifically includes 5 convolutional layers, 3 pooling layers, 2 fully connected layers, and 1 softmax regression layer, as follows: Figure 7 As shown. After training, the above 11 layers of the network yield feature maps of sizes 30×30×48, 15×15×48, 8×8×128, 4×4×128, 4×4×192, 4×4×192, 4×4×128, 2×2×128, 2048×1, 1024×1, and 2×1, respectively. In the convolutional layers, the feature map of the previous layer is convolved with the filter in the upper left corner, multiplying the corresponding numbers and then adding them together. The filter slides smoothly until all features have been calculated to form the feature map of that layer. The pooling layer is located between two convolutional layers and helps with feature dimensionality reduction, removing redundant information, and preserving the translation invariance, rotation invariance, and scale invariance of the feature image. It divides the input feature map into different regions, reduces the image resolution of each region through pooling operations, and eliminates information shifts and distortions in the image. The fully connected layer is closely connected to the output feature map of the previous layer, thereby transforming the multi-dimensional output into a one-dimensional vector. It integrates and summarizes the different features extracted by the convolutional and pooling layers, and finally achieves classification through the Softmax regression layer.

[0106] In S3, the CNN-11 model training process uses the Sigmoid activation function, the Adam optimizer, the Mean Squared Error loss function, a batch size of 32, and a learning rate of 1×10⁻⁶. -5 The number of iterations is 1×10 5 Next, save the weight parameters after model training and use them for the test set.

[0107] Part 3, the model application steps, mainly include the following four steps:

[0108] S1: Input the pre-selected 500 test images into the pre-trained improved model U-RNet+. The segmentation results show that the improved model U-RNet+ outperforms the other two models in terms of contour information preservation, noise removal, contour clarity, and continuity. Figure 8 As shown.

[0109] S2: An evaluation system is established for the mitochondrial image segmentation results, using pixel accuracy, specificity, sensitivity, precision, and Dice coefficient as evaluation metrics. These five metrics represent the accuracy of predicting all pixels, the ability to detect background, the ability to detect mitochondria, the accuracy of mitochondrial detection, and the similarity between all predicted results and the actual results, respectively. The segmentation results have only two classes: positive (mitochondria) and negative (background).

[0110] (1)TP: The number of instances that are correctly classified as positive, i.e., the number of instances that are actually positive and are classified as positive by the deep learning model;

[0111] (2)FP: The number of instances that are incorrectly classified as positive, i.e., the number of instances that are actually negative but are classified as positive by the deep learning model;

[0112] (3) FN: The number of instances that were incorrectly classified as negative, i.e., the number of instances that were actually positive but were classified as negative by the deep learning model;

[0113] (4) TN: The number of instances that are correctly classified as negative, i.e., the number of instances that are actually negative and are classified as negative by the deep learning model.

[0114] Determine the classification of each pixel in the segmented image: it belongs to either mitochondria or the background. If both can be correctly classified, then the mitochondria can be correctly segmented. Therefore, the four metrics TP, FP, FN, and TN mentioned above are calculated, and the following five metrics are also calculated:

[0115]

[0116]

[0117]

[0118]

[0119]

[0120] The U-Net base model achieves a pixel accuracy of 95.18%, specificity of 99.66%, recall of 94.58%, precision of 70.97%, and a Dice coefficient of 0.8291. The proposed U-RNet+ segmentation model achieves a pixel accuracy of 97.51%, specificity of 98.06%, recall of 93.36%, precision of 86.45%, and a Dice coefficient of 0.8977. The improved U-RNet+ model has a higher pixel accuracy than U-Net, indicating a higher accuracy in detecting all pixels. Both models exhibit high and similar specificity. While the improved U-RNet+ model has a lower recall than U-Net, indicating lower sensitivity, it also demonstrates stronger anti-interference capabilities and less susceptibility to noise, which is particularly crucial in mitochondrial image segmentation. The improved model U-RNet+ has a higher precision than U-Net, indicating that U-RNet+ has higher segmentation accuracy for mitochondria and produces clearer mitochondrial outlines. The improved model U-RNet+ also has a higher Dice coefficient than U-Net. This metric is the weighted harmonic mean of recall and precision. When recall and precision show discrepancies, the Dice coefficient is used for comprehensive evaluation, demonstrating that the improved model U-RNet+ performs better overall than U-Net.

[0121] S3: Based on the mapping relationship between image fluorescence intensity and RGB, a Python program was written to perform pseudo-color processing on the original image. This pseudo-color image was then multiplied with the segmentation result obtained through U-RNet+, resulting in a pseudo-color image that balances the color information of the original image and the contour information of the segmented image. The fluorescence intensity of the pseudo-color image generated by Python+U-RNet+ was quantitatively evaluated according to a standard ranging from 0% to 100%. Figure 9 As shown. This invention can generate mitochondrial pseudocolor images in batches, and the colors can be adjusted at any time according to the programmed instructions.

[0122] S4: A pre-selected test set of 1000 images (including 500 images of mitochondria treated with CCCP (Carbonyl cyanide 3-chlorophenylhydrazone, the mitochondrial oxidative phosphorylation uncoupling agent) and normal mitochondria) was input into the trained CNN-11 model for testing. Ultimately, in the 500 CCCP-treated mitochondrial segmentation images, the CNN-11 model classified 495 as Fission and 5 as Fusion, achieving a classification accuracy of 99.0% for CCCP-treated mitochondrial segmentation images. In the 500 normal mitochondrial segmentation images, the CNN-11 model classified 497 as Fusion and 3 as Fission, achieving a classification accuracy of 99.4% for normal mitochondrial segmentation images. Because the U-RNet+ model used in the image segmentation process has a deep layer and excellent segmentation performance, the information features of the obtained mitochondrial segmentation images are relatively clear.

[0123] Part 4, the model application steps, mainly includes the following two steps:

[0124] S1: First, morphological denoising of mitochondrial segmentation images is achieved based on the Opening Area algorithm. Second, based on image processing algorithms, 102 morphological parameters of the mitochondrial images are obtained, such as... Figure 10 As shown, XGBoost is an efficient gradient boosting decision tree model. Its idea is to integrate multiple weak classifiers into a strong classifier using a specific method. After U-RNet+ segmentation, the mitochondrial morphology parameters obtained above and the results obtained from CNN-11 are input into the XGBoost model for text analysis and curve fitting.

[0125] S2: Using the method in S1, Fission images in mitochondrial images taken in batches under different experimental conditions were predicted and evaluated to explore the drug dose-response relationship and the drug action time-response relationship, and the results were compared with those of the mitochondrial membrane potential (MMP) test. The results showed a similar increasing trend, such as... Figure 11 and Figure 12 As shown, this invention can predict the degree of mitochondrial damage by using deep learning to assist biological experiments, thereby reducing a large number of tedious cell experiments and providing an evaluation tool for cutting-edge biomedical applications such as mitochondrial therapy, mitochondrial probes, and mitochondrial implantation. Simultaneously, this invention explores the relationship between mitochondrial morphology and drug action time and concentration, and predicts the trend of mitochondrial membrane potential changes under drug action, showing good potential for applications in disease diagnosis and drug efficacy evaluation.

[0126] Flowcharts and / or block diagrams of methods and computer program products according to embodiments of this application describe various aspects of this application. It should be understood that each block of the flowcharts and / or block diagrams, as well as combinations of blocks in the flowcharts and / or block diagrams, can be implemented by computer-readable program instructions.

[0127] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processing unit of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0128] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions that execute on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0129] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0130] Note that, unless otherwise explicitly stated, all features disclosed in this specification (including any appended abstracts and drawings) can be replaced by alternative features for achieving the same, equivalent, or similar purpose. Therefore, unless explicitly stated otherwise, each disclosed feature is merely one example of a set of equivalent or similar features. Where used, "further," "preferably," "even further," and "more preferably" are simple starting points for describing another embodiment based on the foregoing embodiments, the combination of which with the foregoing embodiments constitutes the complete configuration of another embodiment. Any combination of several "further," "preferably," "even further," or "more preferably" arrangements following the same embodiment constitutes yet another embodiment.

[0131] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for establishing a deep learning-based segmentation and classification model of a mitochondrion super-resolution image, characterized in that: The model, from input to output, includes an image segmentation model and an image classification fusion model; The image segmentation model is the improved U-RNet+ model. The improved U-RNet+ model adds an improved residual network after each convolutional layer of the U-Net base network. An improved residual unit consists of an improved convolutional residual block and two improved identity residual blocks. In the improved convolutional residual block, the convolutional layers with a stride of 2 in the main path and the side path are replaced with a combination of convolutional layers with a stride of 1 and pooling layers. In the improved identity residual block, the input layer of the side path is replaced with a convolutional layer with a stride of 1. The CBAM mechanism is added to the feature map passed by the outermost skip connection. The image classification and fusion model includes an image classification model and an image fusion model set sequentially; wherein, the image classification model is the improved CNN-11 model; and the image fusion model is XGBoost. This model is established through the following steps: S1: Obtain training sample data Mitochondrial super-resolution fluorescence images were captured using structured light illumination super-resolution fluorescence microscopy for training image segmentation models. S2: Image preprocessing and annotation for image segmentation model training After cropping and filtering the mitochondrial super-resolution fluorescence image obtained from S1, the mitochondrial super-resolution fluorescence image is labeled, including the region of mitochondria in the image and their specific outline. S3: Training of Deep Learning-Based Image Segmentation Model The image segmentation model was trained using an S2-annotated dataset. The input to the training process was an annotated super-resolution fluorescence image of mitochondria. The learning objective was the spatial distribution and shape contour of mitochondria. The output was a binarized segmented image containing the spatial distribution and shape contour of mitochondria, and a quantifiable pseudo-color image of mitochondria that took into account both the color information of the original image and the contour information of the segmented image was obtained. S4: Image preprocessing and annotation for image classification models The binarized segmented images obtained by S3 were filtered and expanded, and mitochondria of different morphologies were classified into corresponding tag folders for labeling. S5: Training of Image Classification Model Based on Deep Learning The image classification model is trained using the S4-annotated dataset. The input to the training process is the S4 preprocessed and annotated binarized segmented images of S3. The learning objective is the shape category of mitochondria, and the output is the prediction and score of the mitochondrial category. S6: Training of Image Classification and Fusion Model Based on Deep Learning The image fusion model is trained. The input to the training process is the predicted mitochondrial category and score obtained by the S5 image classification model and the morphological parameter index of the mitochondrial image obtained by the image processing algorithm. The learning target is the shape category of mitochondria, and the output is the predicted mitochondrial category and score. S7: Obtain the segmentation and classification model of mitochondrial super-resolution images that has been built and trained based on deep learning.

2. The method for establishing according to claim 1, characterized in that: The improved model U-RNet+ includes 104 convolutional layers and 16 pooling layers, and the model is divided into a downsampling part in the first half and an upsampling part in the second half. The downsampling part is a typical convolutional neural network structure, which is repeated 5 times using a combination of two convolutional layers and one pooling layer. After each convolutional layer, the structure of the residual unit is added and improved. An improved residual unit consists of an improved convolutional residual block and two improved identity residual blocks. In the improved convolutional residual block, the convolutional layers with a stride of 2 in the main path and the side path are replaced with a combination of convolutional layers and pooling layers with a stride of 1. In the improved identity residual block, the input layer of the side path is replaced with a convolutional layer with a stride of 1. The upsampling part first performs a transposed convolution operation, then reassembles a new feature map by splicing the feature maps of the same dimension of the corresponding left channel, and then uses two convolutional layers to extract features. This structure is repeated 4 times. Furthermore, as the skip connection span of the U-RNet+ model gradually increases, attention mechanisms are added to the outermost feature map in the mixing dimension, namely the spatial dimension and the channel dimension; in the final output layer, the feature map is mapped to the output result map; The improved model CNN-11 includes 5 convolutional layers, 3 pooling layers, 2 fully connected layers and 1 Softmax regression layer; Convolutional layers are used to reduce image noise and extract image features. Pooling layers are used to reduce the dimensionality of features, remove redundant information, and maintain the translation invariance, rotation invariance, and scale invariance of the feature image. Fully connected layers are used to integrate the feature space mappings calculated by previous layers into value outputs; The Softmax regression layer then classifies the output values; The feature vectors used by the image fusion model XGBoost include the predicted mitochondrial category and score obtained through CNN-11 and the morphological parameter indicators of mitochondrial images obtained based on image processing algorithms.

3. The establishment method according to claim 1 or 2, characterized in that, The specific steps for S1 are as follows: Mitochondria within cells were stained and prepared as fluorescent samples. The mitochondria were then imaged using a SIM microscopy device, and at least 1,000 super-resolution fluorescent images of mitochondria at a resolution of 2048×2048 were captured.

4. The method of claim 3, wherein, Step 2) specifically involves: S2.1 uses Python programming to batch crop the mitochondrial super-resolution fluorescence images obtained in S1 into several 512×512 resolution image blocks according to a sliding window of fixed pixel size, and filters out image blocks with invalid background information and high noise ratio. The remaining high-quality image blocks with morphological representativeness form a dataset for training the image segmentation model, and are divided into training set, validation set and test set according to proportion. S2.2 Use Labelme image annotation software to manually annotate the mitochondrial contours of the training and validation sets, and generate the corresponding number of JSON files after annotation is completed; S2.3 uses the written test.bat script to extract information from JSON files in batches, including the original image, labeled image, label name, and YAML file representing the labeled content; in the labeled image, black represents the background and white represents the mitochondrial outline of the label. S2.4 Use Python programming to extract the original image and labeled image one by one in sequence, and classify them into the data and label folders respectively.

5. The method for establishing according to claim 4, characterized in that, The specific steps of S4 are as follows: Using Python programming, images with mitochondrial pixel values ​​less than 3% in the binarized segmented images obtained by S3 were screened out. Data augmentation strategies of clockwise rotation and vertical flipping were used to expand the data, and finally, a training set and a test set with expanded data were obtained. The annotation process involves naming images of healthy mitochondria that are mostly filamentous and images of diseased mitochondria that are mostly round, and finally classifying them into the Fusion and Fission folders respectively.

6. The method of claim 5, wherein the step of establishing comprises the step of: The specific steps for S6 are as follows: ​ S6.1 Morphological denoising of mitochondrial segmentation images based on the Opening Area algorithm; S6.2 Based on image processing algorithms, 102 morphological parameters of mitochondrial images are obtained, including mitochondrial number, total area, average area, median area, area standard deviation, average eccentricity, median eccentricity, eccentricity standard deviation, average isodiameter, median isodiameter, isodiameter standard deviation, average Euler number, median Euler number, Euler number standard deviation, average spread, median spread, spread standard deviation, average principal axis length, median principal axis length, principal axis length standard deviation, average secondary axis length, and median... Secondary axis length, standard deviation of secondary axis length, mean direction, median direction, standard deviation of direction, mean perimeter, median perimeter, standard deviation of perimeter, mean convexity, median convexity, standard deviation of convexity, mean of centroid x-coordinate, median of centroid x-coordinate, standard deviation of centroid x-coordinate, mean of centroid y-coordinate, median of centroid y-coordinate, standard deviation of centroid y-coordinate, mean of distance, median of distance, standard deviation of distance, mean of weighted centroid x-coordinate, median of weighted centroid x-coordinate, standard deviation of weighted centroid x-coordinate Mean weighted center y-coordinate, median weighted center y-coordinate, standard deviation of weighted center y-coordinate, mean weighted distance, median weighted distance, standard deviation of weighted distance, mean shape factor, median shape factor, standard deviation of shape factor, mean roundness, median roundness, standard deviation of roundness, mean branch count, median branch count, standard deviation of branch count, mean branch length, median branch length, standard deviation of branch length, mean total branch length, median total branch length, standard deviation of total branch length, mean median branch length, median median branch length, standard deviation of median branch length, standard deviation of branch length, mean weighted branch angle, median weighted branch angle, standard deviation of branch angle, mean median branch angle Degrees, median median branch angle, median branch angle standard deviation, mean branch angle standard deviation, median branch angle standard deviation, branch angle standard deviation, mean total density, median total density, total density standard deviation, mean average density, median mean density, mean density standard deviation, mean median density, median median density, median density standard deviation, kurtosis x, weighted kurtosis x, kurtosis y, weighted kurtosis y, kurtosis squared, weighted kurtosis squared, skewness x, weighted skewness x, skewness y, weighted skewness y, skewness squared, weighted skewness squared, network orientation (degrees), network principal axis (pixels), network secondary axis (pixels), network eccentricity, network effective range, network effective solidity, network fractal dimension; S6.3 The mitochondrial morphological parameters obtained above and the results output from S5 are input into the XGBoost model for analysis to obtain the prediction and score of mitochondrial category.

7. The method for establishing according to claim 6, characterized in that: In step 3), the conditions for implementing the improved model U-RNet+ training are: The programming language is Python, the high-level neural network API used is Keras, and a virtual environment mitounet is configured to implement this task. The main Python toolkits and their corresponding versions used are libtiff=4.2.0, matplotlib=3.0.3, opencv=4.5.4, pandas=1.2.4, numpy=1.19.5, keras-preprocessing=1.1.2, tensorflow=2.5.0, and tensorflow-gpu=2.5.

0. The hyperparameters are specifically: the excitation function is ReLU, the optimizer is set to Adam, the loss function is a binary cross-entropy function, the batch size is set to 2, the learning rate is 1x10 -4 , and the number of iterations is 0.5x10 3 . In step 5), the conditions for improving the CNN-11 model training are as follows: The programming language is Python, the high-level neural network API used is Tensorflow, and a virtual environment miconn is configured to implement this task. The main Python toolkits and their corresponding versions used are matplotlib=3.3.4, numpy=1.19.0, pandas=1.1.5, opencv=4.5.5, tensorflow=1.9.0, and tensorflow-gpu=1.9.

0. The hyperparameters are specifically: the excitation function is Sigmoid, the optimizer is set to Adam, the loss function is Mean SquaredError, the batch size is set to 32, the learning rate is 1 x 10 -5 , and the number of iterations is 1 x 10 5 . In step 6), the conditions for implementing the image processing algorithm to obtain mitochondrial morphological parameter indicators are as follows: The programming language is Python, and the main Python packages and their corresponding versions used are numpy=1.18.0, matplotlib=3.5.2, opencv=3.4.2, scikit-image=0.19.2, scipy=1.4.1, mkl-random=1.2.2, backport.shutil-which=3.5.2, pandas=1.13.0, and request=2.27.

1. The training conditions for the image fusion model XGBoost are as follows: The programming language is Python, and the machine learning library used is Scikit-Learn. A virtual environment, mitoxgb, was configured to implement this task. The main Python toolkits and their corresponding versions used are xgboost=1.5.0, matplotlib=3.5.2, numpy=1.21.5, xlwt=1.3.0, pandas=1.4.4, and scikit-learn=1.1.

1. The parameters are specifically: booster=gbtree, eta=0.3, max_depth=5, min_child_weight=3, gamma=0.1, subsample=0.7, lambda=3, objective=reg:gamma, colsample_bytree=0.7, seed=0, nthread=4, and the number of iterations is 5x10 3 times.

8. A deep learning-based mitochondria segmentation and classification model, characterized in that: It is obtained by using any one of the methods described in claims 1-7.

9. A method for mitochondria segmentation and classification based on deep learning, characterized in that: Mitochondrial segmentation and classification are performed using the model established by any one of claims 1-7. The super-resolution fluorescence image of the mitochondria to be processed is input into the model to obtain the prediction and score of the mitochondrial category.

10. A computer readable storage medium and an electronic device, having stored thereon a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method of claim 9.

Images