Multi-photon microscopic image automatic segmentation and regional quantitative feature extraction method based on weak supervision pseudo-label and application of multi-photon microscopic image automatic segmentation and regional quantitative feature extraction method

Abstract

The invention discloses a multi-photon microscopic image automatic segmentation and regional quantitative feature extraction method based on a weak supervision pseudo label and application thereof. The method is based on lung cancer Hamp; and E, performing multi-photon imaging on the dyed paraffin section, and collecting SHG and TPEF multi-channel high-resolution images. A classification network with ResNet101 as a trunk is adopted, and pixel-level multi-class pseudo labels are generated. And inputting the pseudo tag as a supervision signal into a U-Net segmentation network to complete automatic segmentation of tumor, tumor-related interstitial substance, tumor necrosis and lymphocyte multi-category regions. Based on the segmentation result, multiple quantitative algorithms such as skeletonization, principal axis analysis, branch detection, area and density calculation, fractal and curvature analysis and the like are further integrated, and automatic extraction and standardized statistics of multi-dimensional morphology and spatial distribution characteristics of the target area are achieved. According to the method, the pixel-level manual labeling requirement is greatly reduced, the segmentation precision and the analysis automation level are improved, and the method is suitable for tumor microenvironment, matrix research, molecular pathology auxiliary diagnosis and other scenes.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing, artificial intelligence, and biomedical engineering technology, and in particular to an automatic segmentation and quantitative feature extraction method for lung cancer multiphoton microscopy images based on weakly supervised pseudo-labels, and its application. Background Technology

[0002] Lung cancer, a major threat in the global disease spectrum, maintains a persistently high incidence and mortality rate, seriously threatening human health and placing enormous pressure on public health. Histopathological imaging is a core tool for the precise diagnosis and classification of lung cancer, providing physicians with crucial information at the cellular and tissue structural levels, serving as an important basis for clinical decision-making. In recent years, multiphoton microscopy techniques (such as second harmonic generation (SHG) and two-photon excited fluorescence (TPEF)) have demonstrated significant advantages in the early diagnosis of tumors such as lung cancer and in the analysis of tissue microenvironment due to their high spatial resolution and label-free imaging, providing important technical support for the multi-level observation of tumor pathological characteristics.

[0003] While multiphoton imaging has enriched the methods for acquiring tissue structures, its practical application in lung cancer pathology diagnosis still faces numerous challenges. On the one hand, the tissue morphology in multiphoton microscopy images differs significantly from traditional HE-stained images. Current clinical pathologists have limited experience interpreting multiphoton images, resulting in a high barrier to objective interpretation of image results. On the other hand, automated analysis methods for multiphoton microscopy images are still in the exploratory stage. Some studies have attempted to apply deep learning segmentation algorithms, but due to the lack of large-scale, pixel-level manually labeled data, truly fully supervised deep learning-based segmentation models are difficult to scale up. In practice, most methods rely on a limited amount of manually coarsely labeled data, combined with traditional image processing, weakly supervised, or semi-supervised algorithms for segmentation. This not only limits the improvement of model performance but also increases the annotation burden on experts, becoming a significant bottleneck hindering the high-throughput automated analysis and clinical application of multiphoton imaging.

[0004] Furthermore, current analysis of tissue features such as collagen fibers in multiphoton images primarily relies on visual observation and semi-quantitative subjective judgment by pathologists. For example, parameters such as collagen distribution, orientation, and density are often assessed through visual scoring or manual delineation. While this approach has some practical guiding significance, its inherent subjectivity means that evaluation results may differ among different observers, lacking a unified objective standard and making it difficult to guarantee the reproducibility and consistency of the analysis results. Especially in cases of complex tissue structures and significant variations in signal quality, accurate and objective quantitative analysis of features such as collagen remains a practical challenge.

[0005] Weakly supervised segmentation methods, by introducing coarse-grained, low-cost annotation techniques combined with the model's self-learning capabilities, have brought new breakthroughs to the automated and intelligent analysis of medical images. These methods can effectively reduce the burden of manual annotation and improve segmentation efficiency and the applicability of results. However, most current weakly supervised segmentation techniques are mainly applied to routine HE-stained pathological images or medical images such as CT / MRI, and their applicability to complex tissue segmentation and quantitative analysis in diseases such as lung cancer using multiphoton microscopy remains insufficient.

[0006] In summary, there is an urgent need for a novel method suitable for multiphoton microscopy images of lung cancer that requires no pixel-level annotation, can automatically segment multiple tissue regions, and accurately and in batches extract quantitative features of target areas such as collagen, in order to promote the in-depth application of multiphoton imaging technology in the pathological diagnosis and scientific research of lung cancer. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide an automatic segmentation and quantitative feature extraction method for multiphoton microscopy images based on weakly supervised pseudo-labels and its application. This method significantly reduces the need for pixel-level manual annotation, improves segmentation accuracy and the level of automation in analysis, and is applicable to scenarios such as tumor microenvironment, matrix research and molecular pathology auxiliary diagnosis.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a method for automatic segmentation and quantitative feature extraction of multiphoton microscopy images based on weakly supervised pseudo-labels, comprising the following steps: Step 1: Use paraffin-embedded tissue sections from lung cancer and perform H&E staining according to standard procedures; Step 2: Use a multiphoton microscope to obtain SHG and TPEF two-channel images; Step 3: Divide the original SHG and TPEF images into 8×8 non-overlapping patches. Each patch contains 512×512 pixels. Two pathology experts use the open-source annotation tool Labelme to annotate the regions of each patch and store them in JSON format. Step 4: Input the SHG and TPEF channel images and multi-class classification labels into the pseudo-label generation network. The pseudo-label generation network adopts the ResNet101 network structure and uses five categories of labels, namely tumor, tumor-associated stroma, tumor necrosis, lymphocytes and background, as supervision to perform feature learning and multi-class classification training. Step 5: After the classification network is trained, based on its output feature map, Grad-CAM and DropoutAttention mechanisms are used to jointly generate saliency heatmaps for the five classes. Through probability thresholding and spatial consistency strategies, the heatmaps are transformed into pixel-level multi-class pseudo-labels. Step 6: The generated multi-class pseudo-labels are used as pixel-level supervision signals and input into the U-Net segmentation network for training. The U-Net network backbone is a 5-layer encoder-decoder structure with an initial number of 32 channels, which doubles with each layer. The convolutional kernel size is 3×3, the activation function is ReLU, and BatchNorm regularization is used. An SE channel attention module is introduced after each encoding layer. The compression ratio is 16, and the weights are learned from the fully connected layers. The training process uses the cross-entropy loss function. Step 7: Use the trained segmentation model to infer all multiphoton microscopy images and output the segmentation mask for each image; Step 8: For each segmented target region, the Zhang-Suen skeletonization algorithm is used to extract the main line, the PCA method is used to calculate the direction and length of the main axis, the 8-neighbor pixel connection detection method is used to count the number of branch points, the area is calculated by multiplying the total number of mask pixels by the area of ​​a single pixel, the density is the ratio of target pixels per unit area, the fractal dimension is calculated by the box-counting method, the curvature is calculated by the chord length method based on the coordinates of the three points of the skeleton sampling points, and the region width is calculated by taking the average width value after distance transformation and calculating the standard deviation. Step 9: Output all the above morphological and spatial distribution features in CSV format, containing the area, density, principal axis orientation, principal axis length, branching degree, fractal dimension, mean and extreme values ​​of curvature, and mean and standard deviation of width for each region.

[0009] In a preferred embodiment, the pseudo-label generation network specifically includes: an input layer receiving a fused image from two channels, SHG and TPEF; a backbone using a ResNet101 structure, and an output layer using a five-class softmax, with classes including tumor, tumor-associated stroma, tumor necrosis, lymphocytes, and background; Grad-CAM and Dropout Attention modules acting on the feature maps of each layer of the backbone network, generating heatmaps independently for each class, and finally obtaining pixel-level five-class pseudo-labels through fusion and thresholding strategies.

[0010] In a preferred embodiment, the U-Net segmentation network specifically includes: inputs are two-channel images and corresponding pixel-level pseudo-labels; each layer of the encoder consists of two sets of 3×3 convolutions, ReLU activation, BatchNorm, and 2×2 max pooling, with the number of channels being 32, 64, 128, 256, and 512 respectively; each layer of the decoder consists of transposed convolutions, 3×3 convolutions, ReLU activation, and BatchNorm, with the number of channels being 512, 256, 128, 64, and 32 respectively; skip connections use feature concatenation; an SE channel attention module is added after each convolution output of the encoder, with a compression ratio of 16 and weights activated by sigmoid; the output layer is a five-class softmax output per pixel, corresponding to tumor, tumor-associated stroma, tumor necrosis, lymphocytes, and background segmentation masks.

[0011] In a preferred embodiment, in step 8, skeletonization uses the Zhang-Suen algorithm to process all target regions, each containing at least 100 pixels; PCA principal axis analysis takes the skeleton pixels of the target region as input and outputs the principal axis direction angle and principal axis length; branch point detection is based on the 8-neighborhood of skeleton pixels, and pixels with more than 2 neighbor connections are counted as branch points; region area = number of pixels × 0.306916μm², density = target pixels / total pixels of patch; fractal dimension uses the box-counting method, with box side lengths of 2, 4, 8, 16, 32, and 64 pixels, counted independently at each scale, and the slope is taken from the regression results; curvature is sampled at equal intervals of 5 pixels with a skeleton point step size, and calculated using the three-point method, outputting the mean and maximum curvature in units of 1 / μm; width is obtained through Euclidean distance transformation, where twice the distance from the skeleton to the boundary is the local width, and the mean and standard deviation are output.

[0012] In a preferred embodiment, the post-processing of the segmentation mask includes: applying a morphological closing operation of 3×3 structuring elements to all segmentation masks to eliminate small gaps, followed by applying a 3×3 opening operation to remove small spots; removing all noise patches smaller than 1% of the main target area through area thresholding; filling small holes inside the main area using 8-neighborhood connected component analysis; and optimizing the mask edges using density-guided filtering with a filtering window size of 5×5 pixels.

[0013] In a preferred embodiment, the thickness of the paraffin-embedded tissue section for lung cancer is specifically 4 μm.

[0014] This invention also provides an application of a method for automatic segmentation and quantitative feature extraction of multiphoton microscopy images based on weakly supervised pseudo-labels. This method is applied to the preparation of auxiliary diagnostic products for lung cancer pathology.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention significantly reduces the reliance on pixel-level manual annotation by generating multi-class pseudo-labels based on ResNet101 features, Grad-CAM, and DropoutAttention mechanisms, thereby reducing labor costs and improving data annotation efficiency.

[0016] 2) This invention drives U-Net segmentation training through multi-class pseudo-labels and integrates multiple quantitative image processing algorithms to achieve high-precision automatic segmentation of tumors and their microenvironment in multiple tissue regions and automatic extraction of multi-dimensional quantitative features. This greatly improves the objectivity, repeatability and standardization of the analysis, and provides strong data support and scientific basis for tumor microenvironment and related clinical auxiliary diagnosis. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of the automatic segmentation and quantitative feature extraction method for multiphoton microscopy images according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the pseudo-label generation and segmentation network structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the automatic extraction and analysis of quantitative features of tissue regions such as collagen in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0022] A method for automatic segmentation and quantitative feature extraction of multiphoton microscopy images based on weakly supervised pseudo-labels, referenced Figure 1-3 This includes the following steps: Step 1: Surgically removed lung cancer tissue was fixed in 10% neutral formalin at room temperature for 12-24 hours and then embedded in paraffin. After embedding, tissue sections with a thickness of 4 μm were cut using a microtome. The sections were stained with Hematoxylin-Eosin (H&E) according to standard procedures, including dewaxing, gradient alcohol infusion, hematoxylin staining, hydrochloric acid-ethanol differentiation, eosin counterstaining, gradient alcohol dehydration, xylene clearing, and neutral resin mounting, finally obtaining clear and usable H&E stained tissue samples.

[0023] Step 2: Place the prepared H&E-stained sections under a Zeiss LSM880 multiphoton microscope using a 20× objective lens. Set the laser excitation wavelength to 810 nm. Set the SHG and TPEF signal acquisition channels to acquire high-resolution images of 4096×4096 pixels, with a single pixel physical size of 0.554 μm × 0.554 μm. SHG signals reflect matrix structures such as collagen, while TPEF signals reflect endogenous fluorescent components in the tissue. Acquire SHG and TPEF images of at least three regions of interest for each sample to ensure sample representativeness and data richness.

[0024] Step 3: Divide each acquired SHG and TPEF dual-channel raw image into 8×8 non-overlapping patches, with each patch being 512×512 pixels, ensuring that the segmented regions are easy to label and have high spatial resolution. The segmentation process is automated using a custom Python script, with uniform numbering and location identification for easy data traceability and processing later.

[0025] Step 4: Two experienced pathologists used the open-source annotation tool Labelme to annotate each patch, including categories such as tumor, tumor-associated stroma, tumor necrosis, lymphocytes, and background. The annotation results were saved in JSON format.

[0026] Step 5: Perform quality control and remove patches with low signal strength or containing artifacts. Ultimately, each sample should have at least 150 valid patches participating in subsequent training.

[0027] Step 6: Using the image-enhanced SHG and TPEF dual-channel patches and their multi-class labels as input, construct a neural network with ResNet101 as the backbone structure for multi-class classification training of tumor, tumor-associated stroma, tumor necrosis, lymphocytes, and background. The network structure is as follows: Input layer: receives 2-channel (SHG+TPEF) patches, size 512×512; Main structure: adopts standard ResNet101, consisting of 5 stages, each stage containing multiple Bottleneck residual modules, all convolutional kernels are 3×3 in size, using BatchNorm and ReLU activation; Output layer: after global average pooling, connected to a fully connected layer, softmax outputs the five-class probabilities. The training parameters are Adam optimizer, initial learning rate 0.001, batch size 16, multi-class cross-entropy loss, training for 100 epochs, and using a validation set early stopping strategy.

[0028] Step 7: After training, for each valid patch, generate pseudo-labels as follows: Keep the Dropout layer in ResNet101 enabled, perform 20 consecutive forward inferences, collecting features from the last convolutional layer each time, and generate heatmaps for each of the five categories using the Grad-CAM algorithm. For each category, average the 20 heatmaps at the pixel level. For each pixel, compare its response in the five category heatmaps. If the maximum response value is greater than 0.6, the pixel is assigned to the corresponding category; if it is less than 0.6, the final label is determined by majority voting from surrounding pixels, achieving spatial consistency correction. Finally, generate five pixel-level pseudo-labels (tumor, tumor-associated stroma, tumor necrosis, lymphocytes, background) of size 512×512 for each patch, which will be used for subsequent segmentation network training.

[0029] Step 7: Use the original image and pixel-level pseudo-labels as the training set for the U-Net segmentation network. The U-Net segmentation network mainly consists of: an input layer that receives SHG and TPEF dual-channel patches and corresponding pixel-level pseudo-labels; an encoder with a 5-layer structure, each layer containing two sets of 3×3 convolutions (channel numbers of 32, 64, 128, 256, and 512), ReLU activation, BatchNorm normalization, and 2×2 max pooling; after the output of each convolution layer, a Squeeze-and-Excitation (SE) channel attention mechanism is introduced, with a compression ratio of 16, and the weights are obtained by sigmoid activation; a decoder with 5 layers, upsampling using transposed convolutions (stride 2), with the number of channels in each layer being 512, 256, 128, 64, and 32 respectively, each layer containing two sets of 3×3 convolutions, ReLU activation, and BatchNorm, with skip connections using feature concatenation; and an output layer with a 1×1 convolution + softmax activation, outputting a 512×512×4 five-class probability distribution map. The loss function is a weighted combination of pixel-level multi-class cross-entropy and Dice loss, with a weight ratio of 1:1. The Adam optimizer is used with an initial learning rate of 0.001, a batch size of 8, and 100 training epochs. Data augmentation (random flipping, rotation, pruning, etc.) is employed during training to improve generalization performance.

[0030] Step 8: Using the finally trained model, automatically infer and predict the segments of all SHG and TPEF images, and output the segmentation mask for each segment. All segmentation masks are stitched together according to the original image number and position to form a complete segmentation mask image, achieving automatic segmentation of the entire image.

[0031] Step 9: Perform morphological closing operations on the automatically segmented mask using 3×3 structuring elements to fill small gaps, followed by 3×3 opening operations to remove small noise spots. Area thresholding is used to remove stray patches smaller than 1% of the main region area, and 8-neighborhood connected component analysis is used to fill holes within the main region. The mask edges are further optimized using density-guided filtering with a 5×5 pixel filter window to enhance the continuity and accuracy of region boundaries.

[0032] Step 10: For each segmented target region, the Zhang-Suen skeletonization algorithm is used to obtain the backbone, with at least 100 skeleton pixels. Based on the skeleton pixels, the PCA algorithm is used to extract the principal axis direction (0~180 degrees, accuracy 0.1 degrees) and principal axis length (unit μm). An 8-neighbor pixel connection detection method is used; pixels with more than 2 connections in the skeleton are considered branch points, and the number of branch points is output as the branch degree. The area is equal to the number of mask pixels in the region × 0.306916μm² (physical area of ​​a single pixel), and the density is the number of pixels in the target region / the total number of pixels in the patch. The fractal dimension is calculated using the box-counting method, with box side lengths of 2, 4, 8, 16, 32, and 64 pixels. Multi-scale counting results are subjected to linear regression, and the slope is the fractal dimension. Curvature is calculated using a three-point method with equidistant sampling points (5-pixel stride) on the skeleton, and the mean and maximum curvature values ​​(unit 1 / μm) are output. The region width is transformed using Euclidean distance, with the local width calculated as twice the distance from the skeleton to the boundary. The mean and standard deviation of the width of the entire region are then calculated. Quantitative characteristics of all regions, including area, density, principal axis direction, principal axis length, branching degree, fractal dimension, mean and extreme values ​​of curvature, and mean and standard deviation of width, are output in a CSV file and saved along with the segmentation mask image for subsequent data analysis, scientific research statistics, and auxiliary medical judgment.

[0033] The equipment selection, parameter configuration, and algorithm details in the specific embodiments of this invention can be appropriately adjusted according to different tissue types, imaging instruments, and research objectives. For example, the patch partitioning method, the number of layers in the segmentation network structure, the weights of the loss function, post-processing operations, and feature extraction algorithms can all be flexibly changed, and those skilled in the art can make equivalent transformations and optimizations without departing from the core ideas of this invention.

Claims

1. A method for automatic segmentation and quantitative feature extraction of regions in multiphoton microscopy images based on weakly supervised pseudo-labels, characterized in that, Includes the following steps: Step 1: Use paraffin-embedded tissue sections from lung cancer and perform H&E staining according to standard procedures; Step 2: Use a multiphoton microscope to obtain SHG and TPEF two-channel images; Step 3: Divide the original SHG and TPEF images into 8×8 non-overlapping patches. Each patch contains 512×512 pixels. For each patch, two pathology experts use the open-source annotation tool Labelme to annotate the region and store it in JSON format. Step 4: Input the SHG and TPEF channel images and multi-class classification labels into the pseudo-label generation network. The pseudo-label generation network adopts the ResNet101 network structure and uses five categories of labels, namely tumor, tumor-associated stroma, tumor necrosis, lymphocytes and background, as supervision to perform feature learning and multi-class classification training. Step 5: After the classification network is trained, based on its output feature map, Grad-CAM and DropoutAttention mechanisms are used to jointly generate saliency heatmaps for the five classes. Through probability thresholding and spatial consistency strategies, the heatmaps are transformed into pixel-level multi-class pseudo-labels. Step 6: The generated multi-class pseudo-labels are used as pixel-level supervision signals and input into the U-Net segmentation network for training. The U-Net network backbone is a 5-layer encoder-decoder structure with an initial number of 32 channels, which doubles with each layer. The convolutional kernel size is 3×3, the activation function is ReLU, and BatchNorm regularization is used. An SE channel attention module is introduced after each encoding layer. The compression ratio is 16, and the weights are learned from the fully connected layers. The training process uses the cross-entropy loss function. Step 7: Use the trained segmentation model to infer all multiphoton microscopy images and output the segmentation mask for each image; Step 8: For each segmented target region, the Zhang-Suen skeletonization algorithm is used to extract the main line, the PCA method is used to calculate the direction and length of the main axis, the 8-neighbor pixel connection detection method is used to count the number of branch points, the area is calculated by multiplying the total number of mask pixels by the area of ​​a single pixel, the density is the ratio of target pixels per unit area, the fractal dimension is calculated by the box-counting method, the curvature is calculated by the chord length method based on the coordinates of the three points of the skeleton sampling points, and the region width is calculated by taking the average width value after distance transformation and calculating the standard deviation. Step 9: Output all the above morphological and spatial distribution features in CSV format, containing the area, density, principal axis orientation, principal axis length, branching degree, fractal dimension, mean and extreme values ​​of curvature, and mean and standard deviation of width for each region.

2. The method for automatic segmentation and quantitative feature extraction of multiphoton microscopy images based on weakly supervised pseudo-labels according to claim 1, characterized in that, The pseudo-label generation network specifically includes: an input layer that receives a fused image from two channels of SHG and TPEF; a backbone using a ResNet101 structure, and an output layer of five-class softmax, with categories including tumor, tumor-associated stroma, tumor necrosis, lymphocytes, and background; Grad-CAM and Dropout Attention modules acting on the feature maps of each layer of the backbone network to generate heatmaps independently for each class, and finally obtaining pixel-level five-class pseudo-labels through fusion and thresholding strategies.

3. The method for automatic segmentation and quantitative feature extraction of multiphoton microscopy images based on weakly supervised pseudo-labels according to claim 1, characterized in that, The U-Net segmentation network specifically includes: input is a two-channel image and corresponding pixel-level pseudo-labels; each layer of the encoder consists of two sets of 3×3 convolutions, ReLU activation, BatchNorm, and 2×2 max pooling, with the number of channels being 32, 64, 128, 256, and 512 respectively; each layer of the decoder consists of transposed convolutions, 3×3 convolutions, ReLU activation, and BatchNorm, with the number of channels being 512, 256, 128, 64, and 32 respectively; skip connections use feature concatenation; an SE channel attention module is added after each convolution output of the encoder, with a compression ratio of 16 and weights activated by sigmoid; the output layer is a five-class softmax output per pixel, corresponding to tumor, tumor-associated stroma, tumor necrosis, lymphocytes, and background segmentation masks.

4. The method for automatic segmentation and quantitative feature extraction of multiphoton microscopy images based on weakly supervised pseudo-labels according to claim 1, characterized in that, In step 8, skeletonization uses the Zhang-Suen algorithm to process all target regions, with each target region containing at least 100 pixels; PCA principal axis analysis takes the skeleton pixels of the target region as input and outputs the principal axis direction angle and principal axis length; branch point detection is based on the 8-neighborhood of skeleton pixels, and pixels with more than 2 neighbor connections are counted as branch points; region area = number of pixels × 0.306916μm², density = target pixels / total pixels of patch; fractal dimension uses the box-counting method, with box side lengths of 2, 4, 8, 16, 32, and 64 pixels, counted independently at each scale, and the slope is taken from the regression results; curvature is sampled at equal intervals of 5 pixels with a skeleton point step size, and calculated using the three-point method, outputting the mean and maximum curvature values ​​in units of 1 / μm; The width is obtained by Euclidean distance transformation. The local width is obtained by doubling the distance from the skeleton to the boundary. The mean and standard deviation are output.

5. The method for automatic segmentation and quantitative feature extraction of multiphoton microscopy images based on weakly supervised pseudo-labels according to claim 1, characterized in that, The post-processing of the segmentation mask includes: applying morphological closing operations with 3×3 structuring elements to all segmentation masks to eliminate small gaps, followed by applying 3×3 opening operations to remove small spots; removing all noise patches smaller than 1% of the main target area through area thresholding; filling small holes inside the main area using 8-neighborhood connected component analysis; and optimizing the mask edges using density-guided filtering with a filtering window size of 5×5 pixels.

6. The method for automatic segmentation and quantitative feature extraction of multiphoton microscopy images based on weakly supervised pseudo-labels according to claim 1, characterized in that, The thickness of the paraffin-embedded tissue section for lung cancer is specifically 4 μm.

7. An application of a method for automatic segmentation and quantitative feature extraction of multiphoton microscopy images based on weakly supervised pseudo-labels, characterized in that, The method for automatic segmentation and quantitative feature extraction of multiphoton microscopy images based on weakly supervised pseudo-labels as described in any one of claims 1-6 is applied to the preparation of lung cancer pathology auxiliary diagnostic products.

Images