A deep learning-based TBA image classification and identification method

Abstract

This invention discloses a deep learning-based TBA image classification and recognition method, belonging to the field of image recognition technology. The method includes performing tissue visual region segmentation and source difference normalization on a full-scale TBA microscopic image to generate a multi-scale tissue image patch sequence. The multi-scale tissue image patch sequence is input into a single-branch residual visual classification network to generate positive / negative visual response maps and intracellular / extracellular spatial response maps. Spatial consistency fusion is performed on the positive / negative visual response maps and intracellular / extracellular spatial response maps to output the TBA image's positive / negative classification result, intracellular / extracellular distribution classification result, and category activation heatmap. This invention focuses on the specific medical examination image object of TBA microscopic images, and strings together tissue region segmentation, source difference normalization, single-branch residual visual classification, spatial response re-projection, and category activation heatmap output into a complete image recognition process. This provides a reproducible data processing path and interpretable spatial basis for TBA image positive / negative classification and intracellular / extracellular distribution recognition.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a TBA image classification and recognition method based on deep learning. Background Technology

[0002] Tissue-based assay (TBA) is a common screening method for anti-neural antibody detection. It typically uses visual information from indirect immunofluorescence microscopy images, such as tissue fluorescence intensity, fluorescence distribution, cell structure texture, and intracellular and extracellular reaction patterns, to provide interpretation criteria for antibodies related to autoimmune neurological diseases. With the increasing sample volume in medical testing institutions, manual image interpretation faces higher demands in terms of consistency, review efficiency, and the interpretation experience of junior staff. Deep learning-based image recognition technology can learn visual features related to positive / negative interpretation and intracellular / extracellular distribution patterns from microscopic images and output image categories or auxiliary interpretation results. Therefore, it is increasingly being used for automatic recognition of medical microscopic images and clinical auxiliary image interpretation scenarios.

[0003] However, TBA microscopic image classification differs from ordinary image classification. Visual interpretation depends not only on fluorescence intensity but also on the integrity of the tissue region, the clarity of cell structure texture, whether the intracellular and extracellular spatial distribution conforms to the interpretation pattern, and whether non-specific background fluorescence interferes with identification. Existing deep learning recognition methods, when applied to TBA images, often learn slide background, blank areas, local reflections, weak texture noise, or differences in sample origin as category-related features, leading to insufficient consistency between positive / negative identification results and real tissue visual evidence. Furthermore, if the model only outputs negative or positive class probabilities, it's difficult to further explain whether its judgment criteria are concentrated on effective tissue structures, cell texture areas, or areas related to intracellular and extracellular distribution, affecting the review and acceptance of identification results by laboratory personnel. In addition, TBA images corresponding to blood samples and cerebrospinal fluid samples may differ in background brightness, tissue fluorescence response, and imaging texture. Without visual feature consistency constraints oriented towards the recognition task, the model's generalization ability in real clinical testing scenarios is easily reduced. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing TBA image classification and recognition methods have the following problems: it is difficult to stably separate tissue regions from background interference; the difference between blood and cerebrospinal fluid sources leads to a shift in visual response distribution; single classification output lacks constraints on the consistency of intracellular and extracellular space and interpretable heatmaps; and how to achieve reproducible recognition of TBA image positivity and intracellular and extracellular distribution.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a TBA image classification and recognition method based on deep learning, comprising performing tissue visual region segmentation and source difference normalization on the full-scale TBA micrograph to generate a multi-scale tissue image block sequence.

[0007] A multi-scale organized image patch sequence is input into a single-branch residual visual classification network to generate positive and negative visual response maps and intracellular and extracellular spatial response maps.

[0008] Spatial consistency fusion is performed on the positive and negative visual response maps and the intracellular and extracellular spatial response maps to output the positive and negative classification results of TBA images, the intracellular and extracellular distribution classification results, and the category activation heatmap.

[0009] As a preferred embodiment of the deep learning-based TBA image classification and recognition method described in this invention, the tissue visual region segmentation of the TBA microscopic full image includes: performing brightness distribution analysis and texture boundary analysis on the TBA microscopic full image to obtain tissue candidate regions. Based on the contour continuity, texture density, and background ratio of the tissue candidate regions, blank visual field regions and non-tissue interference regions are deleted to obtain the tissue visual region. Within the tissue visual region, a tissue contour mask, a cell structure texture map, and a background visual reference region are constructed. The tissue contour mask is used to limit the subsequent image block cropping range, the cell structure texture map is used for subsequent intracellular and extracellular spatial response reprojection, and the background visual reference region is used for subsequent source difference normalization and background response suppression.

[0010] As a preferred embodiment of the deep learning-based TBA image classification and recognition method described in this invention, the source difference normalization includes calculating a background brightness benchmark and a background texture benchmark based on a background visual reference area, and calculating a tissue fluorescence response benchmark based on a tissue visual region. Using the background brightness benchmark, background texture benchmark, and tissue fluorescence response benchmark, brightness normalization, contrast normalization, and texture response normalization are performed on the entire TBA micrograph to obtain a normalized micrograph image. Tissue image blocks at different field-of-view scales are cropped from the normalized micrograph image based on a tissue contour mask. Each tissue image block is mapped to its spatial position in a cell structure texture map, forming a multi-scale tissue image block sequence with spatial position constraints. The multi-scale tissue image block sequence is used for subsequent residual visual feature extraction.

[0011] As a preferred embodiment of the deep learning-based TBA image classification and recognition method described in this invention, the step of inputting a multi-scale tissue image patch sequence into a single-branch residual visual classification network includes, during the training phase, applying Gaussian blur perturbation, brightness perturbation, and contrast perturbation to the training tissue image patches to obtain enhanced tissue image patches. The training tissue image patches and enhanced tissue image patches are then input into the single-branch residual visual classification network. The single-branch residual visual classification network includes a shared residual convolutional path, an anisotropic visual classification path, and an intracellular / intracellular spatial classification path. The shared residual convolutional path extracts shared microscopic visual features from the multi-scale tissue image patch sequence. Both the anisotropic visual classification path and the intracellular / intracellular spatial classification path use the shared microscopic visual features as input and do not construct independent feature extraction paths separately. During the inference phase, the multi-scale tissue image patch sequence corresponding to the full-size microscopic image of the TBA to be identified is input into the trained single-branch residual visual classification network.

[0012] As a preferred embodiment of the deep learning-based TBA image classification and recognition method of the present invention, the generation of positive and negative visual response maps and intracellular / extracellular spatial response maps includes: a shared residual convolution path generating a shared microscopic visual feature map for each tissue image block; a positive and negative visual classification path performing positive fluorescence response localization on the shared microscopic visual feature map to obtain a positive response map for the image block; and an intracellular / extracellular spatial classification path performing intracellular texture response localization and extracellular boundary response localization on the shared microscopic visual feature map to obtain an intracellular response map and an extracellular response map for the image block. Based on the spatial position of each tissue image block in the cell structure texture map, the positive response map, intracellular response map, and extracellular response map of the image block are projected back onto the TBA microscopic full-image coordinates to form the positive and negative visual response map and the intracellular / extracellular spatial response map.

[0013] As a preferred embodiment of the deep learning-based TBA image classification and recognition method of the present invention, the spatial consistency fusion of the positive and negative visual response maps and the intracellular and extracellular spatial response maps includes: forming a positive spatial gating map based on the positive and negative visual response maps; forming an intracellular consistency quantity based on the overlap relationship between the intracellular response region and the cell structure texture region in the intracellular and extracellular spatial response maps; forming an extracellular consistency quantity based on the adjacency relationship between the extracellular response region and the cell boundary texture region in the intracellular and extracellular spatial response maps; forming a background suppression quantity based on the response proportion of the background visual reference region in the positive and negative visual response maps and the intracellular and extracellular spatial response maps; constraining the intracellular consistency quantity and the extracellular consistency quantity using the positive spatial gating map, and weakening the response component located in the background visual reference region using the background suppression quantity to obtain the fused intracellular response quantity and the fused extracellular response quantity; and generating a spatial consistency fusion result based on the spatial competition relationship between the fused intracellular response quantity and the fused extracellular response quantity.

[0014] As a preferred embodiment of the deep learning-based TBA image classification and recognition method of the present invention, the output of the TBA image positive / negative classification result, intracellular / extracellular distribution classification result, and category activation heatmap includes: determining the TBA image positive / negative classification result based on the positive / negative visual response map and the spatial consistency fusion result; determining the intracellular / extracellular distribution classification result based on the fused intracellular response and the fused extracellular response; extracting the category activation weights corresponding to the TBA image positive / negative classification result and the intracellular / extracellular distribution classification result from the shared residual convolution path; superimposing the category activation weights with the shared microscopic visual feature map to obtain the image block category activation map; stitching the image block category activation maps into a category activation heatmap according to the spatial position of each tissue image block in the TBA microscopic full image; and outputting the TBA image positive / negative classification result, intracellular / extracellular distribution classification result, and category activation heatmap.

[0015] As a preferred embodiment of the deep learning-based TBA image classification and recognition system of the present invention, it includes an organization visual normalization module, a residual visual classification module, and a spatial response fusion module.

[0016] The tissue visual normalization module is used to perform tissue visual region segmentation and source difference normalization on the TBA microscopic full image to generate a multi-scale tissue image block sequence.

[0017] The residual visual classification module is used to input multi-scale organized image block sequences into a single-branch residual visual classification network to generate positive and negative visual response maps and intracellular and extracellular spatial response maps.

[0018] The spatial response fusion module is used to perform spatial consistency fusion on the positive and negative visual response maps and the intracellular and extracellular spatial response maps, and output the TBA image positive and negative classification results, intracellular and extracellular distribution classification results, and category activation heatmap.

[0019] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a deep learning-based TBA image classification and recognition method.

[0020] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of a deep learning-based TBA image classification and recognition method.

[0021] The beneficial effects of this invention are: By segmenting the tissue visual region, constructing the background visual reference area, and normalizing the source differences in the TBA microscopic whole image, the separation of tissue region, cell structure texture, and background reference is achieved. This allows subsequent image patch cropping to be limited to the effective tissue range and provides a unified benchmark for the brightness, contrast, and texture response of different sample sources and imaging batches. This achieves the beneficial effect of reducing the interference of blank field of view and source differences on TBA image classification and recognition.

[0022] By inputting multi-scale tissue image patch sequences into a single-branch residual visual classification network and utilizing a shared residual convolution path to simultaneously generate positive and negative visual response maps and intracellular and extracellular spatial response maps, positive and negative fluorescence responses, intracellular texture responses, and extracellular boundary responses are extracted within the same microscopic visual feature space. This allows the model output to be projected back to the TBA microscopic full-image coordinates, thereby achieving the beneficial effect of balancing classification interpretation and spatial localization interpretation.

[0023] By performing positive spatial gating, intracellular-extracellular consistency calculation, background suppression, and spatial competition fusion on positive and negative visual response maps and intracellular-extracellular spatial response maps, a unified output of positive / negative interpretation, intracellular-extracellular distribution judgment, and category activation heatmap is achieved. This allows the final classification result to be jointly constrained by tissue texture location, positive spatial location, and background response ratio, thereby reducing false background triggers and improving the interpretability of clinical review. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 The above is an overall flowchart of a deep learning-based TBA image classification and recognition method provided in Embodiment 1 of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 As an embodiment of the present invention, a deep learning-based TBA image classification and recognition method is provided, comprising: S1. Perform tissue visual region segmentation and source difference normalization on the TBA microscopic full image to generate a multi-scale tissue image block sequence.

[0028] Brightness distribution and texture boundary analysis were performed on the TBA micrographs to obtain candidate tissue regions. Based on the contour continuity, texture density, and background ratio of the candidate tissue regions, blank areas and non-tissue interference areas were removed to obtain the tissue visual regions. Within the tissue visual regions, a tissue contour mask, a cell structure texture map, and a background visual reference region were constructed. The tissue contour mask is used to limit the cropping range of subsequent image patches, the cell structure texture map is used for subsequent intracellular and extracellular spatial response reprojection, and the background visual reference region is used for subsequent source difference normalization and background response suppression.

[0029] Furthermore, the TBA full-screen micrographs are images of tissue sections containing anti-neural antibodies, sourced from indirect immunofluorescence micrographs of blood and cerebrospinal fluid samples. Each TBA full-screen micrograph uses an RGB three-channel image or an equivalent three-channel image derived from a microscopic imaging system as input. The long side of a single image can be uniformly scaled to between 1600 and 4096 pixels, while the short side is scaled proportionally. The source is set to the field-of-view resolution range commonly used in medical testing institutions for microscopic slide reading. This setting preserves tissue structure and fluorescence texture details while avoiding excessive computation caused by directly processing extremely large images.

[0030] Further, the brightness distribution analysis includes converting the TBA micrograph to an image space where the brightness and color channels are separable, and statistically analyzing the mean brightness, variance brightness, and proportion of low-brightness background in the neighborhood of each pixel. Texture boundary analysis involves using local gradient, local variance, and directional texture filter response to characterize tissue boundaries. The local neighborhood window is set to 15×15 pixels to 31×31 pixels, derived from the common pixel scales of cell clusters, tissue edges, and background areas in the microscopic field of view in TBA images. This setting can filter out individual bright spot noise without erasing tissue boundaries.

[0031] Furthermore, considering the uneven background brightness in the TBA micrograph, the proximity of weakly fluorescent tissues to the dark background boundary, and the possibility of isolated noise being misidentified as tissue, this embodiment fuses brightness proximity, texture intensity, boundary response, and background repulsion into a tissue candidate response value to generate tissue candidate regions. The calculation formula is as follows: ; in, Indicates pixel position The candidate response values ​​at the location are derived from the fusion results of brightness distribution analysis and texture boundary analysis, and their values ​​range from 0 to 1 after truncation. This represents the pixel position in the TBA micrograph, derived from the input image coordinate system. The rule for determining this system is to establish a two-dimensional coordinate system with the top left corner of the image as the origin. Indicates pixel position The normalized brightness value is derived from the brightness channel of the input TBA full-size micrograph, and its value ranges from 0 to 1. This represents the initial average background brightness, derived from the intersection statistics of the 10% of pixels with the lowest brightness and a 20-pixel-wide area at the image edge. If the number of intersection pixels is less than 2% of the total number of pixels in the image, then the statistics are taken from the 15% of pixels with the lowest brightness, with a value ranging from 0 to 1. Indicates pixel position The normalized texture density response is derived from the normalized result of the maximum value of the local variance and the directional texture filter response, with a value ranging from 0 to 1. Indicates pixel position The normalized texture boundary response is derived from the fusion of local gradient magnitude and tissue edge direction response, with a value range of 0 to 1. Indicates pixel position The background rejection value is derived from the fusion result of the low saturation ratio and low texture ratio of the background color, and its value ranges from 0 to 1.

[0032] In the above formula, the weights of 0.35, 0.3, 0.25, and 0.1 are the preset engineering values ​​for this embodiment. The contribution order of brightness, texture, and boundary to TBA tissue region localization is set, with brightness and texture as the primary factors, boundary as a secondary factor, and background repulsion used to reduce false triggering in blank fields of view. This setting ensures that the tissue candidate region remains responsive to both weak positive fluorescence and tissue edges, while suppressing background noise. Connectivity regions with a threshold of at least 0.55 are used as initial tissue candidate regions. The threshold of 0.55 is derived from the intermediate separation boundary between the tissue region response and the background region response in the calibration sample, and the value can be selected within the range of 0.5 to 0.6. This setting takes into account both the preservation of weak tissue and the removal of blank background.

[0033] Furthermore, connected component analysis was performed on the candidate tissue regions. Contour continuity was represented by the proportion of effective edge points to the total number of boundary points in the outer boundary of the candidate region. Texture density was represented by the proportion of pixels with a texture response of not less than 0.4 in the candidate region, and background proportion was represented by the proportion of pixels with a background repulsion of not less than 0.6 in the candidate region. The contour continuity threshold was set to 0.35 to 0.5, the texture density threshold was set to 0.2 to 0.35, and the upper limit of the background proportion was set to 0.45 to 0.6. These ranges are derived from the image characteristics of tissue blocks in TBA micrographs, which typically have continuous contours, local texture, and a proportion of non-pure background. This setting removes blank fields of view, bubble edges, slide reflections, and interference from non-tissue particles.

[0034] Furthermore, the tissue contour mask is a binary region map of the same size as the full TBA micrograph. A mask value of 1 indicates that a tissue image block can be cropped, while a mask value of 0 indicates that it will not be used as the center for subsequent image block cropping. The cell structure texture map is a grayscale texture response map corresponding to the tissue visual region, with values ​​ranging from 0 to 1, used to record the distribution of cell bodies, cell clusters, and tissue fibrous textures. The background visual reference area is a region located outside the tissue contour mask with a background repulsion greater than 0.6 and a texture density less than 0.15. Its area is not less than 5% of the total image area; if it is less than 5%, it is supplemented from the region outside the tissue contour mask with the lowest brightness and lowest texture. This setting provides a stable background reference for subsequent source difference normalization and background response suppression.

[0035] Background brightness and texture benchmarks were calculated based on a background visual reference area, and tissue fluorescence response benchmarks were calculated based on tissue visual regions. Using these benchmarks, brightness, contrast, and texture response were normalized on the entire TBA micrograph to obtain a normalized micrograph. Tissue image patches at different field-of-view scales were cropped from the normalized micrograph based on a tissue contour mask. Each tissue image patch was mapped to its spatial location in the cell structure texture map, forming a multi-scale tissue image patch sequence with spatial constraints. This multi-scale tissue image patch sequence was used for subsequent residual visual feature extraction.

[0036] Furthermore, the background brightness benchmark includes the mean and standard deviation of brightness in each color channel of the background visual reference area; the background texture benchmark includes the mean local variance and mean directional texture response of the background visual reference area; and the tissue fluorescence response benchmark includes the channel mean of the pixels with fluorescence intensity between the 60th and 90th percentiles in the tissue visual region. The statistical caliber of the 60th to 90th percentiles is derived from the technical constraint that extremely bright noise, saturated bright spots, and dark backgrounds in TBA images are not suitable for direct use as a reference for tissue fluorescence stability. This setting avoids the normalization result being shifted due to a small number of strong bright spots or a large area of ​​dark background.

[0037] Furthermore, considering the differences in background brightness, fluorescence intensity, and slice texture between blood and cerebrospinal fluid samples, directly inputting them into the model could lead to irrelevant offsets in the feature space for similar antibody images. Therefore, this embodiment uses a background-tissue joint normalization method to generate normalized microscopic images. The calculation formula is as follows: ; in, Represents the pixel position in a normalized micrograph. Color channels The normalized pixel value is derived from the brightness normalization, contrast normalization, and fluorescence response normalization results of the TBA full-image micrograph, and the value ranges from 0 to 1. This indicates the pixel position in the input TBA micrograph. Color channels The original pixel values ​​are derived from the output image of the microscopic imaging system, and the value range is normalized to 0 to 1. This indicates the color channel number, which is derived from the input image channel and can be either the red channel, green channel, or blue channel. Indicates the background visual reference area in the color channel The background brightness benchmark is derived from the average pixel value of the background visual reference area, with a value range of 0 to 1. Indicates the background visual reference area in the color channel The standard deviation of background brightness is derived from the pixel statistics of the background visual reference area and has a value of not less than 0. This represents a stability constant to prevent the denominator from being zero; in this embodiment, it is set to... The source is set to floating-point calculation precision constraint. This setting avoids division by zero errors when the brightness of the background visual reference area is approximately constant. Indicates the visual area of ​​the organization in the color channel The tissue fluorescence response scale is derived from the standard deviation of pixels from the 60th to the 90th percentile in the tissue visual region, and its value is not less than 0. If the value is less than 0.05, it is set to 0.05. Indicates the visual area of ​​the organization in the color channel The tissue fluorescence response baseline is derived from the mean value of the 60th to 90th percentile pixels in the tissue visual region, with a value range of 0 to 1. This function truncates the calculation result to the range of 0 to 1. It originates from the pixel range constraint of digital images and is used to prevent pixel values ​​from going out of bounds after normalization.

[0038] The above normalization calculation results are used to form normalized microscopic images. When the number of pixels in the background visual reference area is insufficient, the standard deviation of a certain channel is too low, or the proportion of saturated pixels exceeds 20%, the background statistics are supplemented by non-organic low-texture regions within adjacent image blocks. If this is still insufficient, the image is marked as a low-background reliable image and its subsequent background suppression weight is reduced. This setting ensures that images from different sample sources and different imaging batches can be brought to a uniform visual scale.

[0039] Furthermore, tissue image patches at different visual scales include local cellular scale image patches, medium tissue scale image patches, and global tissue scale image patches. The original cropping size of local cellular scale image patches is set to 224×224 pixels, medium tissue scale image patches to 448×448 pixels, and global tissue scale image patches to 672×672 pixels. All are resampled to 224×224 pixels before being input into the single-branch residual visual classification network. The cropping stride is set to 50% of the corresponding scale's side length. This setting is derived from the image processing constraint that adjacent image patches need to overlap to reduce the impact of edge truncation. This setting balances local cell texture, tissue distribution background, and computational efficiency.

[0040] Furthermore, each tissue image patch records its top-left corner coordinates, center coordinates, cropping scale, resampling ratio, and corresponding tissue contour mask coverage ratio within the TBA micrograph. Image patches with a tissue contour mask coverage ratio below 0.3 are excluded from the multi-scale tissue image patch sequence; patches with a coverage ratio between 0.3 and 0.5 are marked as boundary image patches; and patches with a coverage ratio above 0.5 are marked as main image patches. This coverage ratio threshold is derived from the tissue region cropping integrity requirement. This setting reduces the risk of false positives / negatives caused by background image patches entering the network, while preserving extracellular boundary responses that may appear at tissue edges.

[0041] Furthermore, the multi-scale tissue image patch sequence with spatial location constraints is arranged in a spatial order from top to bottom and from left to right within the tissue visual region, and sequentially according to local cell scale, intermediate tissue scale, and global tissue scale under the same central coordinate. Each sequence element includes a normalized tissue image patch, scale number, global coordinates, tissue contour mask coverage ratio, and corresponding cell structure texture map slice. This sequence serves as the input for subsequent residual visual feature extraction, enabling the subsequent network not only to receive image content but also to recover the spatial position of each image patch in the TBA micrograph during the response projection.

[0042] It should be noted that this step is not simply performing ordinary cropping or normalization on the microscopic image. Instead, it first distinguishes the tissue visual region, cell structure texture map, and background visual reference region within the full TBA microscopic image. Then, it uses the background brightness benchmark, background texture benchmark, and tissue fluorescence response benchmark to jointly eliminate differences in sample origin. This allows subsequent classification to not rely on a single full-image brightness or a single center cropping result, overcoming the technical problem that visual response shifts are easily generated when blood samples and cerebrospinal fluid samples have differences in the microscopic imaging background.

[0043] S2. Input the multi-scale organized image patch sequence into the single-branch residual visual classification network to generate positive and negative visual response maps and intracellular and extracellular spatial response maps.

[0044] During the training phase, Gaussian blur perturbation, brightness perturbation, and contrast perturbation are applied to the training tissue image patches to obtain enhanced tissue image patches. The training tissue image patches and enhanced tissue image patches are then input into a single-branch residual visual classification network. This network includes a shared residual convolutional path, an anteroposterior visual classification path, and an intracellular / extracellular spatial classification path. The shared residual convolutional path extracts shared microscopic visual features from the multi-scale tissue image patch sequence. Both the anteroposterior visual classification path and the intracellular / extracellular spatial classification path use these shared microscopic visual features as input and do not construct independent feature extraction paths. During the inference phase, the multi-scale tissue image patch sequence corresponding to the TBA microscopic full-image to be identified is input into the trained single-branch residual visual classification network.

[0045] Furthermore, the tissue image patches used for training are derived from a multi-scale tissue image patch sequence obtained through S1 processing. Their labels include positive / negative TBA image labels and intracellular / extracellular distribution labels, formed from medical examination interpretation conclusions or verified clinical examination annotations. The training, validation, and test sets are divided according to patient ID levels, with a ratio of 7:1:2. This ratio is set to prevent data leakage control requirements from allowing different image patches of the same patient to appear simultaneously during the training and testing phases. This configuration allows the model to learn the visual patterns of TBA images, rather than memorizing the local appearance of images from the same patient.

[0046] TBA image labeling includes negative and positive labels. Intracellular and extracellular distribution labels include intracellular distribution labels, extracellular distribution labels, and intracellular-extracellular competitive boundary labels. The label for each training sample is formed independently by at least two qualified reviewers. If the conclusions of two reviewers differ, a third reviewer verifies and determines the final label. Intracellular-extracellular competitive boundary labels participate in loss calculation during the training phase and are output as low-confidence distribution hints during the inference phase.

[0047] Furthermore, considering that slight differences in batches of different microscope equipment, manual focusing, exposure intensity, and staining can cause variations in the clarity, brightness, and contrast of tissue image blocks, this embodiment applies controlled perturbation to the training tissue image blocks. The formula for calculating the enhanced tissue image blocks after perturbation is as follows: ; in, This represents an enhanced tissue image patch, derived from the Gaussian blur perturbation, brightness perturbation, and contrast perturbation results of the training tissue image patch, with a value ranging from 0 to 1. This represents the tissue image patch used for training, which is derived from the multi-scale tissue image patch sequence formed by S1, and its value ranges from 0 to 1. The standard deviation is expressed as The Gaussian blur operator is used to simulate slight defocus or microscopic imaging blur. The standard deviation of Gaussian blur is set to 0.4 to 1.2 pixels in this embodiment. The source of this setting is the range of influence of slight microscopic defocus on the edges of cell texture. This setting enhances the robustness of the model to slight imaging blur without destroying the intracellular and extracellular texture structure. The contrast perturbation coefficient is set to 0.85 to 1.2 in this embodiment. The source of the setting is the contrast fluctuation range under different microscopic exposure and image export conditions. This setting prevents the model from mistakenly using a fixed contrast as the basis for classification. This represents the brightness perturbation offset, which is set to -0.08 to 0.08 in this embodiment. The source is set to the common fluctuation boundary of background brightness and exposure compensation in TBA fluorescence images. This setting improves the model's adaptability to source differences and batch brightness variations. This function truncates the perturbed pixel values ​​to the range of 0 to 1, preventing enhanced pixels from going out of bounds.

[0048] The enhanced tissue image patch and the original training tissue image patch are input into the single-branch residual visual classification network. When the coverage ratio of the enhanced tissue contour mask decreases by more than 10% or the texture density decreases by more than 30%, the enhanced tissue image patch is discarded to avoid excessive perturbation that could damage the TBA tissue structure.

[0049] Furthermore, the single-branch residual visual classification network uses a ResNet34 structure as the basic structure for the shared residual convolutional path, with an input size of 224×224×3. It sequentially includes an initial convolutional layer, a batch normalization layer, a nonlinear activation layer, a pooling layer, and four sets of residual convolutional modules. The shared residual convolutional path outputs shared microscopic visual features. The positive / negative visual classification path includes a global pooling layer and a binary classification output layer, outputting negative and positive probabilities. The intracellular / extracellular spatial classification path includes a spatial response convolutional layer, a global pooling layer, and a distribution classification output layer, outputting intracellular and extracellular distribution probabilities.

[0050] Furthermore, the batch size during the training phase is set to 16 to 64, and the learning rate is set to... to The training epochs were set to 50 to 120 epochs. The weights for positive and negative assignments in the classification loss were set to 0.55 to 0.7, and the weights for intracellular and extracellular distribution assignments were set to 0.3 to 0.45. These parameter ranges are derived from the memory usage, convergence stability, and multi-task label balance requirements of single-branch residual visual classification networks in medical microscopic image block classification tasks. This setting ensures that positive and negative interpretation is the primary task while preserving the spatial discrimination ability of intracellular and extracellular distribution patterns.

[0051] Furthermore, during the inference phase, the TBA microscopic full-image to be identified first generates a multi-scale tissue image patch sequence according to S1, and then inputs each tissue image patch in the sequence into the trained single-branch residual visual classification network. Multiple tissue image patches within the same TBA microscopic full-image are not used individually as the final interpretation result; instead, image patch-level visual responses are first generated, and then back-projected and fused based on the position of the image patch in the full-image coordinates. This setup avoids the occasional strong responses of a single local image patch directly determining the full-image result, ensuring that the final judgment has a spatially consistent basis.

[0052] A shared residual convolution path generates a shared microscopic visual feature map for each tissue image patch. A positive / negative visual classification path locates the positive fluorescence response in the shared microscopic visual feature map, obtaining a positive response map for the image patch. An intracellular / extracellular spatial classification path locates the intracellular texture response and extracellular boundary response in the shared microscopic visual feature map, obtaining intracellular and extracellular response maps for the image patch. Based on the spatial position of each tissue image patch in the cell structure texture map, the positive, intracellular, and extracellular response maps are projected back onto the TBA microscopic full-map coordinates to form a positive / negative visual response map and an intracellular / extracellular spatial response map.

[0053] Furthermore, the shared microscopic visual feature map is a multi-channel feature map output by the last set of residual convolution modules in the shared residual convolution path. Its spatial size is determined by the size of the input tissue image patch and the network downsampling rate. When the input is 224×224×3, the shared microscopic visual feature map can be 7×7×512 or an equivalent size. This shared microscopic visual feature map simultaneously enters both the positive / negative visual classification path and the intracellular / extracellular spatial classification path, without constructing separate independent feature extraction paths. The aim is to express positive fluorescence intensity, tissue texture, and intracellular / extracellular boundary relationships within the same microscopic visual feature space.

[0054] Furthermore, considering that the positive and negative responses, intracellular texture responses, and extracellular boundary responses need to be interpretably located within the same image patch coordinates, this embodiment generates an image patch-level response map based on the category activation weights corresponding to the classification path and the shared microscopic visual feature map. The calculation formula is as follows: ; in, Indicates the first Individual tissue image patches in category Lower, Feature map position The image patch response value at that location is derived from the superposition of the category activation weights and the shared microscopic visual feature map, and the value is a non-negative number. This represents the response category, with values ​​including positive, intracellular, and extracellular categories, derived from the output categories of the positive / negative visual classification path and the intracellular / extracellular spatial classification path. The number represents the tissue image patch, which is derived from the multi-scale tissue image patch sequence formed by S1. The spatial coordinates in the shared microscopic visual feature map are derived from the feature map coordinate system after network downsampling. The channel number represents the shared microscopic visual feature map, with a value ranging from 1 to... . This represents the total number of channels in the shared microscopic visual feature map. In this embodiment, it can be 512 in the last group of residual convolutional modules of ResNet34. The determination rule is the number of output channels of the network structure. Indicates the first Category in tissue image patch The corresponding number Each channel activation weight is derived from the gradient of the corresponding classification path output on the shared microscopic visual feature map channel or the classification layer weight. Indicates the first Shared microscopic visual feature maps of individual tissue image patches in the channel ,Location The feature values ​​at that location are derived from the output of the shared residual convolution path. This represents a non-negative truncation function used to preserve the microscopic visual response that plays a positive role in class determination.

[0055] The formula obtained After being enlarged to the size of the tissue image patch using bilinear interpolation, the images were used as positive response maps, intracellular response maps, and extracellular response maps, respectively. If the maximum and minimum values ​​of the response maps were equal, the response map was set to all zeros to avoid indifferent responses being misinterpreted as valid localization.

[0056] Furthermore, the image patch positive response map is used to locate tissue regions that contribute positive fluorescence visually, the image patch intracellular response map is used to locate regions that overlap with cell body texture, cytoplasmic texture, or texture around the nucleus in the cell structure texture map, and the image patch extracellular response map is used to locate regions adjacent to cell boundary texture, intercellular spaces, or tissue boundaries. Intracellular texture response localization does not rely on a single brightness threshold, but instead utilizes the spatial correspondence between texture channels in the shared microscopic visual feature map and the cell structure texture map to reduce the misidentification of strong background bright spots as intracellular positive regions.

[0057] Furthermore, when projecting the image patch-level response map back to the TBA microscopic full-image coordinates, a mapping from image patch coordinates to full-image coordinates is established based on the top-left corner coordinates, cropping scale, and resampling ratio of the image patch recorded in S1. If multiple image patches are projected to the same full-image pixel location, a weighted average is calculated based on the tissue contour mask coverage ratio, image patch scale weight, and response intensity. The scale weights for local cell scale, intermediate tissue scale, and global tissue scale are set to 0.4, 0.35, and 0.25, respectively. This setting is based on the technical consideration that local cell texture contributes more to intracellular and extracellular localization, while global tissue scale contributes more to regional continuity. This setting balances detail localization and global stability.

[0058] Furthermore, the re-projected positive and negative visual response maps are scaled to the same size as the TBA full-size micrograph or scaled to the same coordinate system at a fixed ratio, with values ​​normalized to 0 to 1. The intracellular and extracellular spatial response maps include intracellular and extracellular response channels, each with values ​​normalized to 0 to 1. The normalization upper and lower limits are derived from the minimum value and the 99th percentile response value of the response map, with the upper limit using the 99th percentile instead of the maximum value to reduce the impact of isolated extrema on the overall response scale. If the 99th percentile is equal to the minimum value, the response channel is set to all zeros. The aforementioned positive and negative visual response maps and intracellular and extracellular spatial response maps serve as inputs for S3 spatial consistency fusion.

[0059] It should be noted that this step does not use ResNet34 as a black-box classifier to directly output a single label. Instead, it simultaneously generates positive and negative visual response maps and intracellular and extracellular spatial response maps based on a single-branch shared residual convolution path, allowing the classification criteria to be projected back into the TBA microscopic full-map coordinates. This process overcomes the problem that independent positive and negative classification paths and independent intracellular and extracellular classification paths may learn inconsistent microscopic features, enabling subsequent fusion to be spatially constrained and output interpretably based on the same shared microscopic visual features.

[0060] S3. Perform spatial consistency fusion on the positive and negative visual response maps and the intracellular and extracellular spatial response maps, and output the TBA image positive and negative classification results, intracellular and extracellular distribution classification results, and category activation heatmap.

[0061] A positive spatial gating map is generated based on the positive and negative visual response maps. Intracellular consistency is determined based on the overlap between intracellular response regions and cell structure texture regions in the intracellular and extracellular spatial response maps. Extracellular consistency is determined based on the adjacency between extracellular response regions and cell boundary texture regions in the intracellular and extracellular spatial response maps. Background suppression is determined based on the response proportion of the background visual reference region in both the positive and negative visual response maps and the intracellular and extracellular spatial response maps. The positive spatial gating map is used to constrain the intracellular and extracellular consistency values, and the background suppression is used to weaken the response components located in the background visual reference region, resulting in fused intracellular and fused extracellular response values. Based on the spatial competition relationship between the fused intracellular and fused extracellular response values, a spatial consistency fusion result is generated.

[0062] Furthermore, the positive spatial gating map is a binary or soft-gated map obtained by converting the positive and negative visual response maps, with values ​​ranging from 0 to 1. When the response value of a pixel in the positive and negative visual response map is not lower than 0.55, that pixel is considered a positive candidate position. When the response value is between 0.45 and 0.55, it is set as a soft-gating transition value in a linear manner. When the response value is lower than 0.45, it is suppressed as a non-positive position. The above gating threshold is derived from the separation boundary between the background response and the tissue positive response after the positive response map of the image patch is normalized. This setting avoids abrupt misjudgment of boundary samples under a single hard threshold.

[0063] Furthermore, considering that intracellular responses only have stable interpretation significance when they coincide with cell structure texture regions, and extracellular responses only have spatial interpretation significance when they are adjacent to cell boundary textures, this embodiment calculates intracellular consistency and extracellular consistency separately, and the calculation formulas are as follows: ; ; in, This represents the intracellular consistency value, derived from the overlap calculation results of positive spatial gating maps, intracellular response maps, and cell structure texture regions, and its value ranges from 0 to 1. The value represents the extracellular uniformity, derived from the overlap calculation results of the positive spatial gating map, the extracellular response map, and the adjacent regions of the cell boundary texture. The value ranges from 0 to 1. This represents the set of pixels covered by the tissue contour mask in the TBA full-image coordinate domain, derived from the tissue contour mask obtained in S1. express The pixel positions within the image are derived from the TBA micrograph coordinate system. Indicates pixel position The positive spatial gating value is derived from the threshold conversion and soft gating processing of the positive and negative visual response maps, and the value ranges from 0 to 1. Indicates pixel position The intracellular response value at the location is derived from the intracellular response channel in the intracellular-intracellular spatial response map formed by S2, and the value ranges from 0 to 1. Indicates pixel position The extracellular response value at the location originates from the extracellular response channel in the intracellular and extracellular spatial response map formed by S2, and its value ranges from 0 to 1. Indicates pixel position The cell structure texture region value at a given location is derived from the texture region after segmentation of the cell structure texture map obtained in S1 using a threshold of 0.4, and takes a value of 0 or 1. The threshold of 0.4 is derived from the medium-intensity boundary of the cell texture response, thus preserving weakly textured cell structures. Indicates pixel position The cell boundary texture adjacency region value is derived from the boundary extraction result of the cell structure texture map, which is then dilated by 3 to 7 pixels to form an adjacency band, and the value is either 0 or 1. The dilation range is derived from the common width of the cell boundary in the microscopic image, and this setting allows for a small spatial deviation between the extracellular response and the boundary texture. This represents a stability constant to prevent the denominator from being zero; in this embodiment, it is set to... The source is set to the stability requirements of floating-point calculations. In the above formula, if empty or If it is less than 1% of the total number of pixels in the organized region of the entire image, then... and All values ​​are set to 0, and the image is marked as a low-positive gating image. This setting prevents unreliable intracellular / extracellular judgments when there is tissue loss or too few positive candidate regions.

[0064] Furthermore, the background suppression amount is jointly formed by the proportion of the background visual reference area's response in the positive and negative visual response maps, intracellular response channels, and extracellular response channels. When the average response in the background visual reference area accounts for more than 0.3% of the average response of the entire tissue visual region, background response contamination is identified, and the background suppression amount is set to a continuous value between 0.2 and 0.6. The proportion threshold of 0.3 stems from the technical constraint that the background region should not bear the main classification response; this setting reduces the impact of non-specific fluorescence, slide reflection, and background noise on the intracellular and extracellular spatial fusion results.

[0065] Furthermore, considering the coupling interference between positive response regions, intracellular and extracellular spatial responses, and background responses, and the potential for misclassification in images with strong background responses when comparing intracellular or extracellular responses alone, this embodiment utilizes positive spatial gating maps and background suppression amounts to generate fused intracellular response amounts, fused extracellular response amounts, and spatial competition amounts. The calculation formulas are as follows: ; ; ; in, This represents the fusion intracellular response, derived from the intracellular consistency after background suppression and the average intracellular response of positive gating, with a value ranging from 0 to 1. This represents the fusion extracellular response, derived from the extracellular uniformity after background suppression and the average extracellular response of positive gating, with a value ranging from 0 to 1. This represents the background suppression amount, derived from the proportion of the background visual reference area's response in the anisotropic visual response map and the intracellular and extracellular spatial response map, with a value ranging from 0 to 0.6. It is set to 0 when the background response proportion is not higher than 0.3, and increases linearly to 0.6 when the background response proportion is higher than 0.3. The average intracellular response under positive spatial gating constraints is derived from... The average value within the tissue contour mask, ranging from 0 to 1. The extracellular mean response under positive spatial gating constraints is derived from... The average value within the tissue contour mask, ranging from 0 to 1. This represents the spatial competition factor, derived from the normalized difference between the fusion intracellular response and the fusion extracellular response, ranging from -1 to 1. A value close to 1 indicates that the intracellular response is dominant, close to -1 indicates that the extracellular response is dominant, and close to 0 indicates that the competition between the two is not significant. This represents a stability constant to prevent the denominator from being zero; in this embodiment, it is set to... The source is set to the stability requirements of floating-point calculation.

[0066] The results of the above formula are used to generate spatial consistency fusion results. When and When all values ​​are below 0.2, the spatial consistency fusion result is labeled as a low-confidence distribution response. Samples with values ​​less than 0.1 are marked as samples representing the intracellular / intracellular competition boundary. The thresholds of 0.2 and 0.1 are derived from the low response boundary and competition boundary settings after response map normalization. This setting avoids forcibly outputting conclusions about intracellular / intracellular distribution when there is insufficient positive or negative evidence.

[0067] Furthermore, the spatial consistency fusion result includes at least the positive spatial gating map, intracellular consistency quantity, extracellular consistency quantity, background suppression quantity, fusion intracellular response quantity, fusion extracellular response quantity, and spatial competition quantity. Among these, the fusion intracellular response quantity and fusion extracellular response quantity serve as direct criteria for judging the intracellular and extracellular distribution classification results, while the positive spatial gating map and background suppression quantity serve as correction criteria for positive and negative classification results. This setup ensures that the final output includes not only category labels but also traceable spatial and background suppression evidence.

[0068] The positive / negative classification results of TBA images are determined based on the fusion results of positive / negative visual response maps and spatial consistency. The intracellular and extracellular distribution classification results are determined based on the fused intracellular and extracellular response quantities. Category activation weights corresponding to the positive / negative classification results and intracellular / extracellular distribution classification results of TBA images are extracted from the shared residual convolution path. These category activation weights are superimposed with the shared microscopic visual feature map to obtain an image patch category activation map. Based on the spatial position of each tissue image patch in the full TBA microscopic image, the image patch category activation maps are stitched together to form a category activation heatmap. The positive / negative classification results, intracellular / extracellular distribution classification results, and category activation heatmap of the TBA images are output.

[0069] Furthermore, the positive / negative classification result of TBA images is jointly determined by the full-image positive response of the positive / negative visual response map, the coverage ratio of the positive spatial gating map, and the spatial consistency fusion result. A positive image is defined as one where the 90th percentile of the positive response within the tissue contour mask is not lower than 0.55, and the coverage ratio of the positive spatial gating map is not lower than 3% of the tissue visual area. A negative image is defined as one where neither of the above conditions is met and the background suppression is not higher than 0.3. If neither of the above positive nor negative criteria is met, a low-confidence positive / negative label is output, and the image is marked as an image awaiting review. Images awaiting review are not considered independent positive / negative categories but rather serve as an auxiliary status to prompt examiners to conduct manual review. The threshold is derived from the technical constraint that positive reactions in TBA images typically present as local or regional tissue fluorescence responses rather than single-pixel noise; this setting reduces the risk of false positives caused by isolated bright spots.

[0070] Furthermore, the classification results for intracellular and extracellular distributions are determined based on the spatial competition relationship between the fusion intracellular and extracellular response values. When the fusion intracellular response value is higher than the fusion extracellular response value and the spatial competition value is not lower than 0.1, the intracellular distribution classification result is output. When the fusion extracellular response value is higher than the fusion intracellular response value and the spatial competition value is not higher than -0.1, the extracellular distribution classification result is output. When the spatial competition value is between -0.1 and 0.1, the intracellular / extracellular competition boundary or the distribution result to be verified is output. This setting ensures that the intracellular / extracellular distribution judgment has a clear spatial competition boundary, rather than relying on a single path probability for direct determination.

[0071] Furthermore, the category activation weights are preferably obtained using a gradient-weighted category activation mapping method, that is, calculating the average gradient of the target category output relative to each channel of the shared microscopic visual feature map, and using this average gradient as the category activation weight of the corresponding channel. In another embodiment, when the classification path includes a global pooling layer and a linear classification layer, the category activation weights can also be obtained from the channel weights corresponding to the target category in the linear classification layer. For the positive or negative classification results of the TBA image, the activation weights corresponding to the positive or negative categories are extracted. For the intracellular or extracellular distribution classification results, the activation weights corresponding to the intracellular or extracellular categories are extracted. The category activation weights are superimposed on the shared microscopic visual feature map to obtain an image patch category activation map, and then the image patch coordinates and scale information recorded in S1 are projected back onto the TBA microscopic full-image coordinates to form a category activation heatmap that can cover the original tissue visual region.

[0072] Furthermore, the category activation heatmap is output in a format that is scaled proportionally to or on the same coordinates as the full TBA micrograph and can be overlaid on the normalized micrograph. Heatmap pixel values ​​are normalized to 0 to 1, where 0 indicates no significant contribution to the current classification result and 1 indicates the highest contribution. The output includes TBA image positive / negative classification results, intracellular / extracellular distribution classification results, category activation heatmap, low-confidence markers, and markers awaiting verification. This output can be integrated into the image reading assistance system of medical testing institutions to provide doctors with auxiliary diagnostic information; however, the final clinical conclusion is still made by qualified laboratory personnel based on clinical data and laboratory procedures.

[0073] It should be noted that this step constrains the final output through positive spatial gating, intracellular and extracellular spatial consistency, background suppression, and category activation heatmaps, rather than relying solely on the single probability score of the deep learning network for interpretation. This processing unifies the positive / negative interpretation, intracellular and extracellular distribution patterns, and interpretable spatial regions of TBA images into the same global coordinate system, overcoming the problems of traditional automatic classification methods that struggle to explain the interpretation criteria, are susceptible to interference from background responses in distribution judgments, and lack intuitive heatmap evidence for physician review.

[0074] Example 2, an embodiment of the present invention, provides a TBA image classification and recognition system based on deep learning, including an organization visual normalization module, a residual visual classification module, and a spatial response fusion module.

[0075] The tissue visual normalization module is used to perform tissue visual region segmentation and source difference normalization on the TBA micrograph, generating a multi-scale tissue image block sequence.

[0076] The residual visual classification module is used to input multi-scale organized image patch sequences into a single-branch residual visual classification network to generate positive and negative visual response maps and intracellular and extracellular spatial response maps.

[0077] The spatial response fusion module is used to perform spatial consistency fusion on the positive and negative visual response maps and the intracellular and extracellular spatial response maps, and outputs the positive and negative classification results of TBA images, the intracellular and extracellular distribution classification results, and the category activation heatmap.

Claims

1. A TBA image classification and recognition method based on deep learning, characterized in that, include: Tissue visual region segmentation and source difference normalization were performed on the TBA microscopic full image to generate a multi-scale tissue image patch sequence; A multi-scale organized image patch sequence is input into a single-branch residual visual classification network to generate positive and negative visual response maps and intracellular and extracellular spatial response maps; Spatial consistency fusion is performed on the positive and negative visual response maps and the intracellular and extracellular spatial response maps to output the positive and negative classification results of TBA images, the intracellular and extracellular distribution classification results, and the category activation heatmap.

2. The TBA image classification and recognition method based on deep learning as described in claim 1, characterized in that: The described method for organizing visual regions in the TBA micrograph includes... Brightness distribution analysis and texture boundary analysis were performed on the full-size TBA micrograph to obtain tissue candidate regions; Based on the contour continuity, texture density, and background ratio of the tissue candidate region, blank visual areas and non-tissue interference areas are removed to obtain the tissue visual region. Construct tissue contour masks, cell structure texture maps, and background visual reference areas within the tissue visual area; Tissue contour mask is used to limit the cropping range of subsequent image blocks, cell structure texture map is used for subsequent intracellular and extracellular spatial response reprojection, and background visual reference area is used for subsequent source difference normalization and background response suppression.

3. The TBA image classification and recognition method based on deep learning as described in claim 2, characterized in that: The source difference normalization includes, The background brightness and texture references are calculated based on the background visual reference area, and the tissue fluorescence response reference is calculated based on the tissue visual region. Using background brightness reference, background texture reference and tissue fluorescence response reference, the TBA micrograph was normalized in brightness, contrast and texture response to obtain a normalized micrograph. Based on the tissue contour mask, tissue image patches at different field-of-view scales are cropped from the normalized microscopic image; Each tissue image block is mapped to its spatial location in the cell structure texture map, forming a multi-scale tissue image block sequence with spatial location constraints. Multi-scale organized image patch sequences are used for subsequent residual visual feature extraction.

4. The TBA image classification and recognition method based on deep learning as described in claim 3, characterized in that: The single-branch residual visual classification network, which uses multi-scale organized image patch sequences as input, includes... During the training phase, Gaussian blur perturbation, brightness perturbation, and contrast perturbation are applied to the training tissue image patches to obtain enhanced tissue image patches. The training tissue image patches and the enhanced tissue image patches are input into the single-branch residual visual classification network; Single-branch residual visual classification networks include shared residual convolutional paths, positive and negative visual classification paths, and intracellular and extracellular spatial classification paths; Shared residual convolutional paths are used to extract shared microscopic visual features from multi-scale tissue image block sequences. Both the yin-yang visual classification path and the intracellular-extracellular spatial classification path use shared microscopic visual features as input, without constructing separate independent feature extraction paths; During the inference phase, the sequence of multi-scale tissue image patches corresponding to the full-scale microscopic image of TBA to be identified is input into the trained single-branch residual visual classification network.

5. The TBA image classification and recognition method based on deep learning as described in claim 4, characterized in that: The generation of positive and negative visual response maps and intracellular and extracellular spatial response maps includes, A shared residual convolution path generates a shared microscopic visual feature map for each tissue image patch; The positive and negative visual classification path locates positive fluorescence response in shared microscopic visual feature maps to obtain positive response maps of image patches; Intracellular and extracellular spatial classification paths are used to locate intracellular texture response and extracellular boundary response of shared microscopic visual feature maps, resulting in intracellular response maps and extracellular response maps of image blocks. Based on the spatial position of each tissue image block in the cell structure texture map, the positive response map, intracellular response map, and extracellular response map of the image block are projected back onto the TBA microscopic full-map coordinates to form positive and negative visual response maps and intracellular and extracellular spatial response maps.

6. The TBA image classification and recognition method based on deep learning as described in claim 5, characterized in that: The spatial consistency fusion of the anisotropic visual response map and the intracellular and extracellular spatial response map includes, A positive spatial gating map is generated based on the positive and negative visual response map; Intracellular consistency is determined by the overlap between intracellular response regions and cell structure texture regions in the intracellular and extracellular spatial response map. The extracellular consistency quantity is formed based on the adjacency relationship between the extracellular response region and the cell boundary texture region in the intracellular and extracellular spatial response map; The background suppression amount is formed based on the response ratio of the background visual reference area in the positive and negative visual response map and the intracellular and extracellular spatial response map; By using positive spatial gating maps to constrain intracellular and extracellular consistency values, and by using background suppression to weaken response components located in the background visual reference region, fused intracellular and extracellular response values ​​are obtained. Based on the spatial competition between the fusion intracellular response and the fusion extracellular response, a spatially consistent fusion result is generated.

7. The TBA image classification and recognition method based on deep learning as described in claim 6, characterized in that: The output TBA image includes positive and negative classification results, intracellular and extracellular distribution classification results, and category activation heatmaps. The TBA image positive / negative classification result is determined based on the positive / negative visual response map and spatial consistency fusion result; The intracellular and extracellular distribution classification results were determined based on the intracellular and extracellular fusion response quantities. Extract category activation weights corresponding to the positive and negative classification results and intracellular and extracellular distribution classification results of TBA images from the shared residual convolution path; The category activation weights are superimposed with the shared microscopic visual feature map to obtain the image patch category activation map; Based on the spatial location of each tissue image patch in the TBA micrograph, the image patch category activation maps are stitched together to form a category activation heatmap; Output TBA image positive / negative classification results, intracellular / extracellular distribution classification results, and category activation heatmap.

8. A TBA image classification and recognition system based on deep learning, employing the TBA image classification and recognition method based on deep learning as described in any one of claims 1 to 7, characterized in that: Includes an organization visual normalization module, a residual visual classification module, and a spatial response fusion module; The tissue visual normalization module is used to perform tissue visual region segmentation and source difference normalization on the TBA microscopic full image to generate a multi-scale tissue image block sequence. The residual visual classification module is used to input multi-scale tissue image block sequences into a single-branch residual visual classification network to generate positive and negative visual response maps and intracellular and extracellular spatial response maps. The spatial response fusion module is used to perform spatial consistency fusion on the positive and negative visual response maps and the intracellular and extracellular spatial response maps, and output the TBA image positive and negative classification results, intracellular and extracellular distribution classification results, and category activation heatmap.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning-based TBA image classification and recognition method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based TBA image classification and recognition method as described in any one of claims 1 to 7.

Images