Intelligent microscope system and auxiliary pathology analysis method based on visual big model

Abstract

The present invention discloses an intelligent microscope system based on a visual macromodel, comprising a microscope and a processor connected to the microscope in communication; the processor comprises an intelligent analysis unit; the intelligent analysis unit comprises a multi-task feature learner constructed based on the visual macromodel; according to the image pair prompted by the task, the intelligent analysis unit controls the microscope to collect microscopic images of the target position, inputs the image pair for the task prompt and the microscopic image of the target position into the multi-task feature learner, and outputs the analysis result of the microscopic image of the target position corresponding to the task. The present invention also discloses an auxiliary pathology analysis method based on the intelligent microscope system. The intelligent microscope system of the present invention improves the efficiency and accuracy of medical image analysis.

Description

Technical Field

[0001] The present invention relates to the field of computer-aided medical technology, and in particular to an intelligent microscope system based on a large visual model and an auxiliary pathology analysis method. Background Art

[0002] Microscopes are indispensable tools for biological and medical analysis. Experts perform cell analysis, lesion detection, biological research, etc. by observing stained slides.

[0003] Current artificial intelligence (AI) technology has pushed the analysis based on microscope images to a digital and intelligent stage. There are roughly three types of AI-assisted microscopic image analysis methods:

[0004] The first is to manually observe samples using an automatic microscope, such as the Chinese patent document with publication number CN113204109A, which discloses a fully automatic microscope for medical microscopic image analysis, including a bottom plate, a focusing mechanism, a scanning camera, a sample horizontal movement mechanism, a sample leveling mechanism, a scanning main light source, a global shooting mechanism, an automatic oiling and waste oil recovery mechanism, etc. Biological and medical analysis based on manual observation is prone to expert fatigue and increases the risk of misanalysis and missed diagnosis.

[0005] The second method is to use a scanner to scan the stained slide into a whole-slice image (WSI), and then use a pre-trained AI model to read, process and analyze it. For example, the Chinese patent with publication number CN116230214A discloses an HCC and VETC auxiliary diagnosis device and equipment, which includes an intelligent microscope device, a processor, a memory and a computer program; the processor implements the slide scanning, processing and analysis steps when executing the computer program. The cost of scanner equipment and digital sample storage is high; training AI models depends on specific microscopic image data sets and complex manual annotations; AI models are generally only effective for specific tasks, and different analysis tasks require different models to be trained, which is inflexible; specific AI models are greatly affected by cross-data styles and have poor generalization.

[0006] The third is to integrate the AI ​​model into the microscope and output the defined content in real time. For example, the processing and analysis methods are integrated into the microscope to display the recognition results of the field of view in real time, such as segmentation, classification, and detection, and then further complete tasks such as driving and generating reports based on the results. For example, the Chinese patent with publication number CN113706438A discloses an image processing method, related devices, equipment, system and storage medium. The existing technology can only handle limited, predefined pathological image analysis tasks, and needs to rely on clear, predefined language tags during the analysis process. It cannot flexibly distinguish pathological results from any form of data, thereby limiting the development of rapid precision medicine. Summary of the invention

[0007] The present invention provides an intelligent microscope system based on a visual macromodel and an auxiliary pathology analysis method, which improves the efficiency and accuracy of medical image analysis.

[0008] The technical solution of the present invention is as follows:

[0009] An intelligent microscope system based on a visual macromodel includes a microscope and a processor in communication with the microscope;

[0010] The processor includes an intelligent analysis unit; the intelligent analysis unit includes a multi-task feature learner built based on a large visual model;

[0011] According to the image pair prompted by the task, the intelligent analysis unit controls the microscope to collect the microscopic image of the target position, inputs the image pair used for task prompts and the microscopic image of the target position into the multi-task feature learner, and outputs the analysis result of the microscopic image of the target position corresponding to the task.

[0012] The tasks include semantic segmentation of cancerous tissue, cell instance segmentation, image color optimization, tissue pseudo-staining, and fluorescence pathology super-resolution image synthesis.

[0013] When the task is semantic segmentation of cancerous tissue, the image pair used for task prompts is the original image and the image with pixel annotations of the cancerous differences in the original image;

[0014] When the task is cell instance segmentation, the image pair used for task prompts is the original image-the image with annotations of different cell instances in the original image;

[0015] When the task is image color optimization, the image pair used for task prompts is the grayscale image of the original image-the original image;

[0016] When the task is tissue pseudo-staining, the image pair used for task prompts is the original image-the stained image of the original image;

[0017] When the task is fluorescence pathology super-resolution image synthesis, the image pair used for task prompts is a clear image of the original image and the synthesized original image.

[0018] Image pairs used for task prompts can be imported from the task image pair dataset, or the microscope can be controlled to collect and annotate them.

[0019] The number of image pairs used for task prompts can be 5-10 pairs.

[0020] The method for constructing the multi-task feature learner includes:

[0021] (i) Collect pathology-related microscopic image data, perform preprocessing and labeling, and construct an image dataset;

[0022] (ii) constructing a multi-task feature learning network model based on the visual big model framework, wherein the multi-task feature learning network model comprises:

[0023] The encoder converts the input image or image sequence into discrete visual tokens and concatenates the discrete visual tokens into a one-dimensional visual token sequence;

[0024] Autoregressive Transformer, which infers the one-dimensional visual tag sequence generated by the encoder to obtain a one-dimensional predicted visual tag sequence;

[0025] The decoder decodes the sequence of predicted visual tags generated by the autoregressive Transformer into a reconstructed image;

[0026] (iii) pre-training the multi-task feature learning network model;

[0027] (iv) Using the image dataset to train the pre-trained multi-task feature learning network model, a multi-task feature learner is obtained.

[0028] Step (i) comprises:

[0029] (i-1) Collect public image data related to pathology and clinical pathology image data and perform preprocessing;

[0030] (i-2) Image data are annotated differently according to different tasks: for the semantic segmentation of cancerous tissues, pathologists annotate the cancerous areas in the image at the pixel level; for the cell instance segmentation task, pathologists annotate different cell instances in the image; for the image color optimization task, the image is converted into a grayscale image to generate an original image-grayscale image pair; for the tissue pseudo-staining task, the unstained image and the stained image are combined into an image pair; for the fluorescence pathology super-resolution image synthesis task, different parts of the same image are collected or images at the same position are collected multiple times to synthesize a clear image, and the synthesized clear image is combined with the original image to form an image pair.

[0031] The image data includes a single-frame image and an image sequence consisting of multiple-frame images.

[0032] The public image data related to pathology comes from 60 open websites and datasets such as Twitter and TCGA; the clinical pathology image data comes from pathology images of different periods, styles, diseases, and tasks from multiple local hospitals. Single-frame images account for 80% of the image dataset, and the rest are image sequences formed by images captured by microscope cameras at different time points.

[0033] These images are processed uniformly and the image size is 256×256. A total of about 200 million image data are collected. The data set data files are divided into training set and test set according to the ratio of 80%:20%.

[0034] Step (iii) includes: pre-training the encoder, decoder and codebook through self-supervised learning; based on the pre-trained encoder, decoder and codebook, reconstructing the codes in the codebook by randomly replacing the codes in the codebook with noise using an autoregressive Transformer, and pre-training the autoregressive Transformer.

[0035] The encoder and decoder are convolutional neural networks.

[0036] Pre-training encoders, decoders, and codebooks through self-supervised learning, including:

[0037] (iii-1) The encoder converts the original image x into a continuous feature ;

[0038] (iii-2) The quantization layer converts continuous features into Converted into a discretized vector z q ;

[0039] (iii-3) The decoder discretizes the vector z q Convert to reconstructed image ;

[0040] (iii-4) Compare the original image x with the reconstructed image Input the discriminator D, which determines the reconstructed image authenticity;

[0041] (iii-5) Repeat steps (iii-1)-(iii-4) and train the encoder, decoder, and codebook simultaneously by minimizing the loss caused by reconstructing the original image, the loss of quantizing continuous features into discrete vectors, and optimizing the loss of the discriminator.

[0042] In step (iv), the AdamW optimizer is used for training, the learning rate is set to 4.5e-06, the momentum parameter is set to (0.9, 0.95), and the model parameters are updated.

[0043] Preferably, the microscope comprises a movable loading platform, an illumination device, an objective lens switching platform, and a microscope camera, wherein the movable loading platform can move along the X, Y, and Z directions, and the objective lens axis is in the Z direction;

[0044] The processor also includes a microscope focusing unit, which controls the movable loading platform and the objective lens conversion table and is used for automatically converting objective lenses of different magnifications and automatically focusing through the movement of the movable loading platform.

[0045] The present invention also provides an auxiliary pathology analysis method based on an intelligent microscope system, comprising:

[0046] Inputting a plurality of sets of image pairs for task prompts into the intelligent analysis unit;

[0047] Control the microscope to scan the slide and obtain a microscopic image of the target location;

[0048] The image pair used for task prompts and the microscopic image of the target position are input into a multi-task feature learner built based on the visual macro model, and the analysis result of the microscopic image of the target position corresponding to the task is output.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] (1) The intelligent microscope system of the present invention can receive and parse instructions through flexibly set task prompts, thereby processing a variety of visual tasks such as semantic segmentation of cancerous tissue, cell instance segmentation, image color optimization, tissue pseudo-staining, and fluorescence pathology super-resolution image synthesis. This improves the efficiency and accuracy of medical image analysis.

[0051] (2) The intelligent microscope system of the present invention integrates automatic control and intelligent analysis into the microscope, saving costs.

[0052] (3) The multi-task feature learner of the present invention learns good features from a large data set. By processing the data set into various image sequences, the feature relationship between images is directly extracted and learned without the need for semantic labels. The multi-task feature learner has strong generalization capabilities and high accuracy in recognition analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The figure is a flow chart of the auxiliary pathological analysis method based on intelligent microscope;

[0054] Figure 2 Pre-training flowchart for CNN encoder, CNN decoder, and codebook;

[0055] Figure 3 Training flowchart for the multi-task feature learner. DETAILED DESCRIPTION

[0056] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0057] like Figure 1 As shown, an auxiliary pathological analysis method based on intelligent microscope comprises the following steps:

[0058] 1.1 Data collection and processing for training a multi-task feature learner for microscopy.

[0059] In order to construct an image dataset, public image data and clinical data related to pathology were collected to form a dataset and visually annotated. The public image data related to pathology came from 60 open websites and datasets such as Twitter and TCGA; the clinical data came from pathology images of different periods, styles, diseases, and tasks in multiple local hospitals. Single-frame images account for 80% of the dataset, and the rest are image sequences formed by images captured by microscope cameras at different time points. These images are uniformly processed and the image size is 256×256. A total of about 200 million image data were collected. The dataset data files are divided into training sets and test sets in a ratio of 80%:20%.

[0060] Depending on the microscope task, the data is labeled in different ways. Specifically, it includes: the task of semantic segmentation of cancerous tissue, in which pathologists perform pixel-level annotation of cancerous areas in the image. The task of cell instance segmentation, in which pathologists label different cell instances in the image. The task of image color optimization: converting the image into a grayscale image and generating a grayscale image-original image pair. The task of tissue pseudo-staining: combining an unstained image with a stained image to form an image pair. The task of fluorescence pathology super-resolution image synthesis: collecting different parts of the same image or collecting images at the same location multiple times, and combining the synthesized clear image with its constituent image pair.

[0061] 1.2 Model:

[0062] Train a multi-task feature learner for smart microscopes, including encoder, decoder and autoregressive transformer. Used for intelligent analysis of microscopes. The encoder and decoder are based on VQGAN, and the autoregressive transformer adopts the transformer structure in the large model LLaMa.

[0063] The specific contents are:

[0064] First, a pathology dataset consisting of a single frame image is used to train a convolutional neural network (CNN) encoder, a convolutional neural network (CNN) decoder, and a codebook through self-supervised learning, such as Figure 2 As shown, the specific implementation process is:

[0065] (1) CNN encoder E converts the input image x into a continuous feature . The specific steps are: the convolution layer used by CNN encoding has a kernel size of 3, a stride of 1, and a padding of 1. The CNN encoder starts with an initial convolution layer, and then loops to add residual blocks and non-local blocks. During this period, downsampling is performed to reduce the spatial resolution of the feature map when the number of channels increases. The attention mechanism is applied at an appropriate resolution to enhance the feature representation. After the cycle ends, residual blocks and non-local blocks are continued to be added to further extract features. Finally, the feature extraction and representation are completed through normalization and activation function x⋅torch.sigmoid(x).

[0066] (2) Input quantization layer to generate discretized vector z q The specific steps are: randomly generate a codebook formed by a finite number of discrete vectors, calculate the square of the Euclidean distance between the input vector z and the vector in the codebook, and use it to find the closest codebook vector. Each encoding position in the code book is used to find the code that is closest to it. The corresponding code book vector is retrieved from the embedding layer using the found code and replaced with a variable with the same dimension to obtain a discretized representation vector z. q .

[0067] (3) The CNN decoder decodes the discretized vector into a reconstructed image . The specific steps are: the convolution layer used by the CNN decoder has a kernel size of 3, a stride of 1, and a padding of 1. The CNN decoder first initializes a starting layer containing a convolution layer, a residual block, and a non-local block to receive the discretized vector and start the image reconstruction process. In subsequent loops, the decoder continues to add residual blocks to refine the feature representation. Except for the first channel of the loop, the decoder performs upsampling each time the number of channels decreases, gradually improving the spatial resolution of the feature map. After the loop ends, the features are adjusted through the normalization layer, and then the activation function x⋅torch.sigmoid(x) is applied to enhance the nonlinear expression, and finally the feature map is converted into a reconstructed image through a convolution layer. .

[0068] (4) The original image and the reconstructed image are sent to the discriminator D to determine their authenticity. The specific steps are: the discriminator consists of multiple convolutional layers, each of which is followed by batch normalization and LeakyReLU activation function to extract image features. The last convolutional layer reduces the number of feature map channels to 1, and the original image and the reconstructed image are judged based on the output. During the training process, the decoder generates a reconstructed image, and the discriminator is responsible for judging whether the generated image is a real image. The two compete with each other and continuously optimize their parameters through adversarial training to make the generated image closer to the real image. By minimizing the loss caused by the reconstruction of the original image, that is, , the loss of quantizing a continuous vector into a discrete vector, i.e., (sg[·] represents the stop gradient operation, and β is the weight factor), and the loss of optimizing the discriminator. These three losses are used to train the encoder, decoder, and codebook simultaneously.

[0069] Based on the CNN encoder and codebook, by adding strong noise through randomly replacing the codes, an autoregressive Transformer is used to reconstruct the code groups in the codebook to improve the generalization ability of the Transformer. The specific implementation process is as follows:

[0070] A sequence is defined for each pixel of the image from left to right and top to bottom. With the sequence, the image can be regarded as a one-dimensional sequence, and the Transfomer can generate the image in the same way as generating a sentence. At the i-th step, the Transformer generates the i-th code according to the previous i - 1 codes. For a given index s < i, the Transformer learns to predict the distribution of possible next indices, i.e., , i.e., .

[0071] (1) The image passes through the trained CNN encoder to generate a continuous feature , and the continuous feature is converted into a discretized vector z through the quantization layer q .

[0072] (2) Flatten z q , randomly replace a part of the codes with randomly generated vectors of the same dimension, and input them into the Transformer.

[0073] (3) The Transformer uses an autoregressive method to infer the next code based on the already predicted codes and outputs the required discrete vector. Reshape the probability logical values (logits) of the predicted codebook vectors and the true values (targets) of the target indices into one-dimensional tensors, calculate the cross-entropy loss between the two, and train the Transformer.

[0074] Load our dataset on the multi-task feature learning neural network and train multi-pathological visual tasks, as Figure 3 shown. The specific implementation process is as follows:

[0075] Before training, add a marker at the beginning and a marker at the end of each image or image sequence, and use sequence concatenation during training to improve efficiency.

[0076] (1) Send the image or image sequence into the pre-trained CNN encoder;

[0077] (2) The CNN encoder converts each input image into 256 discrete tokens, connects the discrete tokens in multiple images into a one-dimensional sequence, and regards it as a visual sentence;

[0078] (3) Input the visual sentence into the autoregressive Transformer model;

[0079] (4) Input the tag sequence generated by Transformer into the CNN decoder;

[0080] (5) The CNN decoder decodes it into an image.

[0081] During training, the AdamW optimizer is used, the learning rate is set to 4.5e-06, and the momentum parameter (betas) is set to (0.9, 0.95) to update the parameters of the Transformer model to minimize the loss function. The Transformer model is trained in the same way on all image sequences, enabling the model to infer the relationship between images from the context rather than from task- or format-specific tags.

[0082] 1.3 Model Testing

[0083] Use the test set to observe the model output effect by giving task prompts.

[0084] The present invention also provides an intelligent microscope system for assisting pathological analysis, including a hardware system and a software system. Hardware system: Place and fix the glass slides in batches on a movable loading platform, connect the microscope to a computer via a USB serial port, and start the microscope system. The movable loading platform can move in XYZ control for positioning and focusing; the lighting system is controllable LED white light and RGB fluorescence for bright field and dark field imaging; the objective lens turret is equipped with multiple objective lenses for controlling the imaging magnification; the camera is used to obtain microscopic images and adjust the images.

[0085] Software system: Intelligent analysis unit and focusing unit of the microscope. The intelligent analysis unit is composed of the above models, and the focusing unit is as follows.

[0086] The focusing unit is used for automatic focusing. The specific method is: first control the microscope to collect the image at the current position and calculate the clarity. Use the Laplace function as the image clarity evaluation index. Control the stage to fine-tune once along the Z axis and calculate the clarity. According to the two clarity changes, adjust the Z axis movement direction and continue to fine-tune until the clarity is maximized.

[0087] Intelligent analysis method: According to the task, import from the test set or control the microscope to collect several sets of image pairs to the analysis unit as task prompts. The image pair includes the input image and the corresponding annotated image. Then control the microscope hardware system to collect the corresponding field of view images. The microscope system receives the field of view transmitted by the camera, starts the analysis unit, and outputs the image analysis results.

[0088] Example

[0089] An example of a smart microscope segmenting cancerous cells in a slide.

[0090] First, 7 sets of cell instance segmentation image pairs are imported from the test set, or the microscope is controlled to collect 7 field of view images and mark the segmentation instances as the system's task prompts; then the microscope is controlled to scan the slide, and the field of view during scanning is used as the input of the microscope system. The microscope performs real-time reasoning on each field of view and outputs the segmentation results of the cancerous cell instances, including the cancer detection box and mark.

[0091] Intelligent microscope for tissue pseudo-staining.

[0092] Import 7 sets of image pairs before and after staining from the test set, or control the microscope to collect 7 sets of field of view images before staining, and collect the corresponding field of view after staining at the same position as the system's task prompt, and then control the microscope to image at the position where pseudo-staining is required. The microscope system infers the field of view and outputs the pseudo-staining result of the slide.

[0093] Smart microscopes optimize fluorescence imaging.

[0094] Five sets of high- and low-quality fluorescence image pairs are imported from the test set as task prompts for the system. The microscope is then controlled to take images at one or more locations that need to be optimized. The microscope system infers the field of view and outputs high-quality fluorescence images.

[0095] Microscope for super-resolution imaging of fluorescence pathology.

[0096] Import 5 sets of high-resolution and low-resolution fluorescence image pairs from the test set, or control the microscope to collect 7 sets of low-resolution fluorescence field images, and collect corresponding high-resolution fluorescence images at the same position as the system's task prompts. Then control the microscope to collect the field of view that needs to be imaged, and the microscope system outputs super-resolution fluorescence imaging results. Record and display the output high-resolution fluorescence images for pathologists to analyze and diagnose.

[0097] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent microscope system based on a visual macro model, characterized in that: including a microscope and a processor in communication with the microscope; The processor includes an intelligent analysis unit; the intelligent analysis unit includes a multi-task feature learner built based on a large visual model; The intelligent analysis unit controls the microscope to collect microscopic images of the target position, inputs the image pair used for task prompts and the microscopic image of the target position into the multi-task feature learner, and outputs the analysis result of the microscopic image of the target position corresponding to the task; Tasks include semantic segmentation of cancerous tissue, cell instance segmentation, image color optimization, tissue pseudo-staining, and fluorescence pathology super-resolution image synthesis; When the task is semantic segmentation of cancerous tissue, the image pair used for task prompts is the original image and the image with pixel annotations for the difference between cancerous tissues in the original image; When the task is cell instance segmentation, the image pair used for task prompts is the original image and the image annotated with different cell instances in the original image; When the task is image color optimization, the image pair used for task prompts is the original image and the grayscale image of the original image; When the task is tissue pseudo-staining, the image pair used for task prompts is the original image and the stained image of the original image; When the task is fluorescence pathology super-resolution image synthesis, the image pair used for task prompts is a clear image of the original image and the synthesized original image.

2. The intelligent microscope system based on the visual big model according to claim 1 is characterized in that: The method for constructing the multi-task feature learner includes: (i) Collect pathology-related microscopic image data, perform preprocessing and labeling, and construct an image dataset; (ii) constructing a multi-task feature learning network model based on the visual big model framework, wherein the multi-task feature learning network model comprises: The encoder converts the input image or image sequence into discrete visual tokens and concatenates the discrete visual tokens into a one-dimensional visual token sequence; Autoregressive Transformer, which infers the one-dimensional visual tag sequence generated by the encoder to obtain a one-dimensional predicted visual tag sequence; The decoder decodes the sequence of predicted visual tags generated by the autoregressive Transformer into a reconstructed image; (iii) pre-training the multi-task feature learning network model; (iv) Using the image dataset to train the pre-trained multi-task feature learning network model, a multi-task feature learner is obtained.

3. The intelligent microscope system based on the visual big model according to claim 2 is characterized in that: Step (i) comprises: (i-1) Collect public image data related to pathology and clinical pathology image data and perform preprocessing; (i-2) Image data are annotated differently according to different tasks: for the semantic segmentation of cancerous tissues, pathologists annotate the cancerous areas in the image at the pixel level; for the cell instance segmentation task, pathologists annotate different cell instances in the image; for the image color optimization task, the image is converted into a grayscale image to generate an original image-grayscale image pair; for the tissue pseudo-staining task, the unstained image and the stained image are combined into an image pair; for the fluorescence pathology super-resolution image synthesis task, different parts of the same image are collected or images at the same position are collected multiple times to synthesize a clear image, and the synthesized clear image is combined with the original image to form an image pair.

4. The intelligent microscope system based on the visual big model according to claim 2 is characterized in that: Step (iii) includes: pre-training the encoder, decoder and codebook through self-supervised learning; based on the pre-trained encoder, decoder and codebook, reconstructing the codes in the codebook by randomly replacing the codes in the codebook with noise using an autoregressive Transformer, and pre-training the autoregressive Transformer.

5. The intelligent microscope system based on the visual big model according to claim 4 is characterized in that: Pre-training encoders, decoders, and codebooks through self-supervised learning, including: (iii-1) The encoder converts the original image x into a continuous feature ; (iii-2) The quantization layer converts continuous features into Converted into a discretized vector z q ; (iii-3) The decoder discretizes the vector z q Convert to reconstructed image ; (iii-4) Compare the original image x with the reconstructed image Input the discriminator D, which determines the reconstructed image authenticity; (iii-5) Repeat steps (iii-1)-(iii-4) and train the encoder, decoder, and codebook simultaneously by minimizing the loss caused by reconstructing the original image, the loss of quantizing continuous features into discrete vectors, and optimizing the loss of the discriminator.

6. The intelligent microscope system based on the visual big model according to claim 1 is characterized in that: The microscope comprises a movable loading platform, an illumination device, an objective lens switching platform, and a microscope camera. The movable loading platform can move along the X, Y, and Z directions, and the objective lens axis is in the Z direction. The processor also includes a microscope focusing unit, which controls the movable loading platform and the objective lens conversion table and is used for automatically converting objective lenses of different magnifications and automatically focusing through the movement of the movable loading platform.

7. An auxiliary pathological analysis method based on the intelligent microscope system according to any one of claims 1 to 6, characterized in that: include: Inputting a plurality of sets of image pairs for task prompts into the intelligent analysis unit; Control the microscope to scan the slide and obtain a microscopic image of the target location; The image pair used for task prompts and the microscopic image of the target position are input into a multi-task feature learner built based on the visual macro model, and the analysis result of the microscopic image of the target position corresponding to the task is output.

Images