Intelligent pathological auxiliary diagnosis system in 5g environment

Abstract

The application relates to a 5G environment intelligent pathological auxiliary diagnosis system and relates to the technical field of pathological section diagnosis, in particular to a 5G environment intelligent pathological auxiliary diagnosis system. The application aims to solve the problem that the generalization performance of an existing AI auxiliary pathological diagnosis model is limited and the system cannot maintain the recognition accuracy of pathological digital images. The 5G environment intelligent pathological auxiliary diagnosis system comprises a CT image acquisition module, a CT image processing module, a deep learning module and a recognition module. The CT image acquisition module is used for acquiring CT images of patients by a CT scanner. The CT image processing module is used for processing the acquired CT images of patients to obtain processed CT images of patients, and the processed CT images of patients are used as a training set. The deep learning module is used for constructing a deep learning model and obtaining a trained deep learning model based on the training set. The recognition module is used for recognizing the processed CT images of patients to be measured based on the trained deep learning model.

Description

Technical Field

[0001] This invention relates to the field of pathological slide diagnosis technology, and more specifically to a smart pathological auxiliary diagnosis system in a 5G environment. Background Technology

[0002] In today's world, with the rapid iteration of medical technology and the widespread acceptance of the concept of precision medicine, the total amount of data in the medical field is increasing exponentially. The continuous breakthroughs in diagnostic technology are also injecting new momentum into the innovation of disease diagnosis and treatment models. As an indispensable core component of the modern medical diagnostic system, digital pathology is gradually breaking through the limitations of traditional pathology, promoting the transformation of pathological diagnosis towards digitalization, intelligence, and precision, and becoming an important bridge connecting basic medical research and clinical diagnosis and treatment practices.

[0003] Digital pathology is an interdisciplinary field integrating digital image processing, computer science, clinical medicine, and pathology. Its core essence lies in utilizing advanced digital imaging technologies and intelligent analysis algorithms to achieve a complete digital upgrade of the pathological diagnosis, research, and teaching process. Compared to traditional pathological diagnosis, the core difference lies in the fundamental shift in "observation and analysis methods": Traditional pathology relies on pathologists manually observing and judging stained tissue sections using optical microscopes to complete disease diagnosis. This model is not only limited by subjective factors such as the doctor's experience level and visual fatigue, but also presents many inconveniences in the preservation, transmission, and sharing of sections. Digital pathology, on the other hand, uses professional digital scanning equipment (such as whole-slide scanners) or high-end digital microscopes to convert traditional tissue sections into high-resolution, high-definition digital images. Then, with the help of computer vision technology and digital image analysis algorithms, it performs refined processing, feature extraction, and quantitative analysis of digital sections, accurately capturing subtle changes in tissue structure, characteristic patterns of lesion areas, and abnormal cell morphology, providing objective, accurate, and traceable quantitative evidence for pathological diagnosis.

[0004] The complete technical workflow of digital pathology combines professionalism and systematic approach, with each step ensuring the quality of the final diagnosis. The first step is the digital acquisition of tissue sections. Using devices such as whole-slice scanners, tissue sections only a few micrometers thick can be scanned to generate digital images containing complete tissue information. The resolution reaches the micrometer level, completely restoring every detail of the tissue section and ensuring consistency between the digital image and the original slice. Next is the processing and analysis of digital images. Computer algorithms are used to preprocess the digital images, including noise reduction, enhancement, and segmentation, eliminating interference from factors such as lighting and instrument errors during imaging. Then, feature extraction algorithms accurately extract key information such as morphological features, texture features, and cell density of the lesion area, enabling precise judgment of the lesion type and severity. Finally, these processed and analyzed digital images and feature data can be widely applied in clinical pathology diagnosis, disease classification and grading, prognostic assessment, and basic medical research, providing scientific support for clinical treatment decisions and offering a standardized and reproducible research platform for pathology research.

[0005] Compared to traditional pathological diagnostic methods, digital pathology offers significant advantages, with its core value manifested in three aspects: "breaking limitations, improving efficiency, and empowering innovation." Firstly, it enables remote and shared diagnosis, breaking down the limitations of time and space. Pathologists can remotely view digital slides uploaded by medical institutions in other locations through a network platform, conducting remote consultations and case discussions. This effectively addresses the issues of resource scarcity and uneven diagnostic levels among pathologists in primary healthcare institutions, allowing more patients to benefit from high-quality pathological diagnostic resources. Secondly, it significantly improves diagnostic accuracy and efficiency. Digital images can be arbitrarily magnified, reduced, and annotated, facilitating the capture of subtle lesions. Simultaneously, computer-aided analysis reduces subjective errors in human judgment, lowering the risk of missed or misdiagnosed cases and drastically shortening the diagnostic cycle, thus gaining valuable time for early intervention and treatment. Thirdly, it provides convenience for case management, medical research, and teaching. By establishing a standardized digital pathology database, standardized storage, rapid retrieval, and batch analysis of cases can be achieved. This facilitates pathologists in accumulating clinical experience and conducting retrospective studies, while also providing authentic and intuitive teaching materials for medical education, promoting the improvement of the quality of pathology talent training. It can be said that digital pathology is not only a technological innovation in the field of pathology, but also an important support for promoting the implementation of precision medicine and the high-quality development of the medical cause. It has an irreplaceable and broad application prospect in clinical medicine and basic scientific research.

[0006] It is worth noting that traditional microscopic pathological diagnosis has long relied on the personal experience and visual assessment of pathologists, and its diagnostic effectiveness is greatly affected by subjective factors. The emergence of digital pathology, however, has not only enabled the digitization, standardization, and sharing of pathological images, but more importantly, it has laid a solid foundation for the application of artificial intelligence (AI) technology in pathological diagnosis, ushering in a new era for the development of intelligent pathology systems. With the rapid development of AI technologies such as deep learning and computer vision, using AI to assist in the diagnosis of pathological slides has become a research hotspot and cutting-edge direction in the medical field. AI models, with their powerful feature extraction and rapid analysis capabilities, can automatically identify lesion areas in pathological slides and quantify the degree of lesions, providing pathologists with efficient auxiliary diagnostic tools and further improving diagnostic efficiency and accuracy. They have shown great application potential, especially in the early screening and diagnosis of common tumors such as lung cancer, breast cancer, and colorectal cancer.

[0007] However, despite significant progress in AI-assisted pathology diagnosis, several bottlenecks remain in practical clinical applications, the most prominent being the limited generalization performance of deep learning models. The core root of this problem lies in the numerous differences in the preparation of pathology slides across different medical institutions: different hospitals use different slide preparation instruments, staining reagents, and staining methods, resulting in variations in slide thickness and staining depth; simultaneously, lighting conditions and scanning parameters during imaging differ due to variations in equipment models and operating procedures. These factors collectively lead to significant "domain differences" in digital pathology images from different medical institutions. This domain difference causes AI models trained on a dataset from one hospital to exhibit significant performance degradation when applied to a new data domain from another hospital, failing to maintain stable diagnostic accuracy. This severely limits the large-scale clinical application of AI-assisted pathology diagnosis and represents a key challenge that requires immediate research breakthroughs in this field. Summary of the Invention

[0008] This invention addresses the problem that existing AI-assisted pathology diagnosis models have limited generalization performance and cannot maintain the accuracy of the system's recognition of digital pathological images, and proposes a smart pathology-assisted diagnosis system in a 5G environment.

[0009] The intelligent pathology-assisted diagnosis system in a 5G environment includes: a CT image acquisition module, a CT image processing module, a deep learning module, and a recognition module;

[0010] The CT image acquisition module is used to acquire CT images of the patient by the CT scanner;

[0011] The CT image processing module is used to process the acquired patient CT images to obtain processed patient CT images, which are then used as a training set.

[0012] The deep learning module is used to build deep learning models and obtain trained deep learning models based on the training set.

[0013] The recognition module is used to identify the processed patient CT images to be tested based on a trained deep learning model.

[0014] Preferably, the CT image processing module is used to process the acquired patient CT images to obtain processed patient CT images, which are then used as a training set; the specific process is as follows:

[0015] Access medical databases from different hospitals or medical institutions, and extract continuous two-dimensional pathological slide data of patients scanned by CT scanners from the medical databases;

[0016] The surface rendering method constructs a three-dimensional CT image from consecutive two-dimensional slices of each patient;

[0017] The 3D CT image is divided into a series of 3D cube images;

[0018] The category label of the center pixel of each 3D cube image is used as the category of the corresponding 3D cube image;

[0019] All 3D cube images were used as the training set.

[0020] Preferably, the deep learning module is used to construct a deep learning model and obtain a trained deep learning model based on the training set; the specific process is as follows:

[0021] 1) Constructing a deep learning model; the specific process is as follows:

[0022] Deep learning models include: convolutional kernel size of The depthwise convolutional layer has a kernel size of 1. The depthwise convolutional layer has a kernel size of 1. The depthwise convolutional layer has a kernel size of 1. The 2D convolutional layer, BN layer, GELU layer, Mamba block, global average pooling layer, linear layer, and softmax activation function layer;

[0023] 2) Obtain a trained deep learning model based on the training set; the specific process is as follows:

[0024] 21) For each 3D cube image in the training set Processing along the horizontal direction yields a one-dimensional sequence. ;

[0025] 22) For each 3D cube image in the training set Processing along the vertical direction yields a one-dimensional sequence. ;

[0026] 23) For each 3D cube image in the training set Processing along the horizontal direction yields a one-dimensional sequence. ;

[0027] 24) For each 3D cube image in the training set Processing along the horizontal direction yields a one-dimensional sequence. ;

[0028] 25) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. ;

[0029] 26) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. ;

[0030] 27) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. ;

[0031] 28) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. ;

[0032] 29) One-dimensional sequence One-dimensional sequence One-dimensional sequence One-dimensional sequence By adding elements one by one, we obtain the feature map. ;

[0033] For feature maps The feature map is obtained by reshaping the image; the feature map is then input into a global average pooling layer, which outputs the feature map. ;

[0034] feature map The input is sequentially processed by a linear layer and a softmax activation function layer. The softmax activation function layer outputs the final predicted probability distribution.

[0035] Preferably, in step 21), each three-dimensional cube image in the training set... Processing along the horizontal direction yields a one-dimensional sequence. The specific process is as follows:

[0036] Each 3D cube image in the training set The input convolution kernel size along the horizontal direction is The deep convolutional layer outputs a feature map. ;

[0037] feature map and After element-wise addition, the input convolution kernel size is along the horizontal direction. The deep convolutional layer outputs a feature map. ;

[0038] feature map , and After element-wise addition, the input convolution kernel size is along the horizontal direction. The deep convolutional layer outputs a feature map. ;

[0039] Image Feature map Feature map Feature map By adding elements one by one, we obtain the feature map. ;

[0040] Feature map along the horizontal axis Flattening the sequence yields a one-dimensional sequence. .

[0041] Preferably, in step 22), each three-dimensional cube image in the training set... Processing along the vertical direction yields a one-dimensional sequence. The specific process is as follows:

[0042] Each 3D cube image in the training set The input convolution kernel size is along the vertical direction. The deep convolutional layer outputs a feature map. ;

[0043] feature map and After element-wise addition, the input convolution kernel size is along the vertical direction. The deep convolutional layer outputs a feature map. ;

[0044] feature map , and After element-wise addition, the input convolution kernel size is along the vertical direction. The deep convolutional layer outputs a feature map. ;

[0045] Image Feature map Feature map Feature map By adding elements one by one, we obtain the feature map. ;

[0046] Feature map along the vertical axis Flattening the sequence yields a one-dimensional sequence. .

[0047] Preferably, in step 23), each three-dimensional cube image in the training set... Processing along the horizontal direction yields a one-dimensional sequence. The specific process is as follows:

[0048] 231) For each 3D cube image in the training set Each channel in the process undergoes feature shifting to form a feature map. The specific process is as follows:

[0049] Feature map In a single channel The characteristics of the row are ,in ; Representation of feature map The size of its height or width;

[0050] The first Characteristics of lines Divided into part 1 and part 2; part 1 represents the first part... Characteristics of lines The segmented head; part 2 indicates the segmented head. Characteristics of lines The split tail section;

[0051] Assign part 1 to the new row feature From the position To the final characteristics;

[0052] Assign part 2 to the new row feature. From position 0 to Features;

[0053] Assign the value 0 to the new row feature. From position 0 to Features;

[0054] Assign the value 0 to the new row feature. From the position To the final characteristics;

[0055] Feature map Each channel in the process undergoes feature shifting to form a feature map. ;

[0056] 232) Feature map based on horizontal axis After processing, a one-dimensional sequence is obtained. ;

[0057] For the obtained one-dimensional sequence The former One and after The features are cropped to obtain the cropped one-dimensional sequence. .

[0058] Preferably, in step 24), each three-dimensional cube image in the training set... Processing along the horizontal direction yields a one-dimensional sequence. The specific process is as follows:

[0059] 241) Regarding the feature map Transpose the horizontal and vertical dimensions of the spatial dimension;

[0060] Each channel in the transposed feature map is feature-shifted to form a new feature map. ;

[0061] 242) For feature maps Dimension inversion is performed, and the feature map after dimension inversion is processed based on the vertical axis to obtain a one-dimensional sequence. ;

[0062] For the obtained one-dimensional sequence The former One and after The features are cropped to obtain the cropped one-dimensional sequence. .

[0063] Preferably, in step 25), the one-dimensional sequence After processing, a one-dimensional sequence is obtained. The specific process is as follows:

[0064] 251) For one-dimensional sequences Perform reshape deformation, and then input the reshaped sequence sequentially into the convolution kernel of size [size missing]. The 2D convolutional layer, BN layer, and GELU layer output feature maps. ;

[0065] 252) Feature map Input Mamba block, output feature map ;

[0066] 253) Regarding feature maps Perform feature reversal to obtain the feature map after feature reversal. ;

[0067] 254) Feature map Input Mamba block, output feature map ;

[0068] 258) Feature map and feature map By adding elements one by one, a one-dimensional sequence is obtained. .

[0069] Preferably, in step 252), the feature map Input Mamba block, output feature map The process is as follows:

[0070] feature map Perform two-dimensional pointwise convolution operations sequentially. Shape reshaping operation By swapping the channel dimension and the sequence dimension, we obtain sequence 1.

[0071] Perform on sequence 1 Generate the reversed input sequence 2;

[0072] Sequence 1 will be processed sequentially. Two-dimensional pointwise convolution operation Interchange the channel dimension and the sequence dimension to obtain sequence 3;

[0073] Sequence 2 passes through in sequence Two-dimensional pointwise convolution operation is performed to obtain sequence 4;

[0074] The feature map is obtained by concatenating sequences 3 and 4. ;

[0075] This indicates the Mamba framework based on SSM.

[0076] Preferably, the loss function of the deep learning model is obtained as follows:

[0077] 1) Extract the pseudo-labels from the output of the softmax activation function layer along the category dimension. The maximum value and the former Category position index corresponding to the maximum value ;

[0078] 2) The weights are obtained after passing through the softmax activation function layer. ;

[0079] 2) Based on weight get Supervision items with topk weighted cross-entropy ; indicates as:

[0080]

[0081] in, Indicates the first preprocessed step The first image corresponds to the Supervision items of topk weighted cross-entropy, Indicates the first preprocessed step The topk weights corresponding to the images are as follows: Each weight value Indicates the first preprocessed step The topk index corresponding to the image is the first One index value; express Categories ; express Middle Count , ; This indicates the total number of preprocessed images; Indicates the first preprocessed step Zhang Image ; Indicates the total number of categories; express Middle count, ; Indicates an indicator function, hour Take 1, hour Set to 0;

[0082] 3) Based on Supervision items with topk weighted cross-entropy Calculate loss ; indicates as:

[0083]

[0084] in, Indicates the first The supervising term for the topk weighted cross-entropy; the superscript T indicates the transpose; This represents the preprocessed image input into the neural network, and the classification pseudo-label output by the neural network.

[0085] The beneficial effects of this invention are as follows:

[0086] This invention utilizes surface rendering to construct three-dimensional images from continuous two-dimensional CT slices, restoring the true spatial morphology, infiltration range, and tissue relationships of lesions. This overcomes the limitation of two-dimensional slices in representing spatial structure, significantly improving the accuracy of benign / malignant tumor identification and tumor subtyping. The three-dimensional volume data is divided into regular three-dimensional cubic images, with the center pixel category used as a label, forming a standardized and clearly labeled training set. This solves the problem of inconsistent data formats, sizes, and storage across different hospitals, providing a standardized data foundation for cross-institutional model training. Large-volume CT data is split into small cubic input units, reducing the computational requirements for model training while ensuring that each sample focuses on the local lesion area, improving feature learning efficiency and training stability. Data from multiple hospital medical databases can be directly accessed and processed uniformly, breaking down data silos and allowing the model to be trained on data from multiple institutions and devices. This alleviates domain differences at the data level and improves model generalization performance. Using center pixel labels instead of full-image annotation simplifies the annotation process, reduces annotation errors, and facilitates the rapid construction of a large-scale, high-quality digital pathology training library, supporting the clinical deployment of the system.

[0087] This invention constructs a deep learning model comprising multi-scale deep convolution, Batch Normalization (BN) layers, GELU layers, Mamba blocks, global average pooling layers, linear layers, and Softmax activation function layers. Employing a four-way parallel feature extraction, sequence processing, and feature fusion process, it can fully extract multi-directional, multi-scale spatial features from 3D CT pathological images. The Mamba blocks efficiently capture long-range spatial dependencies, enhancing the model's learning and representation capabilities for lesion features. Simultaneously, through multi-path feature complementarity and fusion, it effectively reduces the impact of differences in pathological image domains between different medical institutions on model performance, enhancing model generalization and diagnostic stability. Ultimately, it achieves accurate classification and identification of normal tissues, benign lesions, and malignant lesions, providing high-performance algorithm support for intelligent pathological auxiliary diagnosis in a 5G environment.

[0088] This invention uses , , Three types of convolutional kernels capture tissue textures, cell arrangements, and lesion extension features of varying lengths in the horizontal direction, covering horizontal structural information from minute to macroscopic. The original image and the outputs of each layer are superimposed element-wise to construct a residual structure, preserving key details at the lower levels and avoiding the loss of subtle lesion features in deep convolutions. The original image and multi-scale outputs are then fused to obtain a clean and complete horizontally specific feature map, highlighting the patterns of lesions in the horizontal direction.

[0089] This invention also uses , , Convolution is used to extract information on tissue thickness, cell layering, and lesion invasion depth in the vertical direction, supplementing vertical structures not covered in the horizontal direction. Residual superposition preserves subtle lesions in the vertical direction, avoids gradient vanishing, and improves sensitivity to minor vertical anomalies. This results in a vertically specific feature map that focuses on the morphology of tissues and lesions in the vertical direction, resulting in more focused features and lower noise.

[0090] This invention simulates the minute shifts in pathological images during scanning, staining, and slice thickness by performing channel-level feature shifts on three-dimensional cubic images. This allows the model to learn core lesion features that remain in place and scale, reducing the interference of imaging detail differences on recognition. It expands the effective receptive field without increasing the convolution kernel size or computational power, using feature shifting to obtain a larger receptive field, better capturing long-range spatial correlations of cells or tissues, and improving the detection capability of minute lesions. Redundant noise is eliminated by pruning the front and back ends of one-dimensional sequences, removing invalid or noisy features at the edges, retaining core lesion information, reducing interference, and improving feature purity and training stability. It also adapts to multi-institutional data, significantly mitigating domain shifts caused by differences in image style and quality between different hospitals, allowing the model to maintain high accuracy on cross-institutional data.

[0091] This invention first performs a spatial transpose in the horizontal or vertical direction, then performs feature shifting and dimension reversal, extracting features from all dimensions including vertical, horizontal, diagonal, and oblique, overcoming the limitation of conventional convolution which only extracts features along a fixed direction. It is adapted to CT 3D pathological data, enabling the model to understand the spatial distribution and connectivity of lesions in the in vivo data, improving the accuracy of benign / malignant diagnosis and tumor subtyping. The combination of multiple spatial transformations significantly reduces image distribution shifts caused by differences in equipment and processes, significantly improving generalization performance and supporting large-scale clinical deployment. Dimension reversal combined with end-point cropping compresses feature length while retaining key information, accelerating inference speed and reducing computational requirements, making it suitable for real-time assisted diagnosis in 5G environments. (State Intellectual Property Office)

[0092] This invention performs dimensionality reshaping, pointwise convolution, normalization, and activation processing on one-dimensional sequences to achieve feature dimension regularization, normalization, and nonlinear activation, ensuring a uniform distribution of features from different directions and institutions. Furthermore, it employs bidirectional Mamba and reverse Mamba sequence modeling to fully exploit long-range spatial dependencies in pathological images, enhancing effective lesion features and suppressing noise interference. Simultaneously, it standardizes and purifies sequence features, significantly improving the model's adaptability to differences in cross-institutional pathological image domains, enhancing feature expression capabilities and diagnostic robustness, and laying a high-quality feature foundation for subsequent multi-path feature fusion and accurate classification.

[0093] This implementation method, based on the SSM's Mamba framework, efficiently captures the spatial correlation between distant tissues and lesions, solving the problem that traditional CNNs can only see local areas and cannot clearly see large-scale infiltrative lesions. Forward Mamba is used to extract forward spatial features from sequence 1, and reverse Mamba is used to extract inverse spatial features from sequence 2. After bidirectional stitching, the lesion features are more complete, the boundaries are clearer, and the classification is more accurate. Through pointwise convolution, reshape, and permute dimensional transformation, the features are adapted to the 3D CT volumetric data format, resulting in a more regular feature distribution and more stable training. Reinforcement learning utilizes the essential structural features of the lesion, mitigating image style differences caused by staining, scanning, and slice thickness, significantly improving the model's diagnostic stability on data from different hospitals. Bidirectional feature complementarity and superposition automatically suppress scanning noise and artifact interference, outputting high signal-to-noise ratio feature maps.

[0094] This implementation extracts and weights the top-k maximum predicted probabilities, allowing the model to focus more on high-confidence, high-value diagnostic categories and improving classification reliability. A weighted cross-entropy supervision term filters low-confidence noise labels, reducing training interference from labeling errors and image differences, resulting in a more accurate and stable model. On datasets from multiple hospitals, with multiple devices, and multiple staining styles, adaptive alignment of data distribution significantly improves model generalization performance and avoids sharp performance drops across different domains. A smoother loss function and more stable gradients prevent overfitting and accelerate convergence, allowing the model to maintain high accuracy even with limited labeled data. It enhances the ability to distinguish between benign, malignant, and normal tissues, reducing missed diagnoses and misdiagnoses, and assisting pathologists in making more reliable judgments.

[0095] This invention, through a deep learning module technology solution, effectively overcomes the bottleneck of insufficient generalization performance in intelligent pathology auxiliary diagnostic systems under 5G environments. Its implementation will bring multi-dimensional positive effects, comprehensively promoting the clinical application and high-quality development of digital pathology and intelligent pathology auxiliary diagnostic systems under 5G environments. From a clinical diagnostic perspective, the implementation of this technology solution can significantly improve the diagnostic stability and accuracy of intelligent pathology auxiliary diagnostic systems under 5G environments in different medical institutions and data domains, greatly reducing the risk of missed diagnoses and misdiagnoses caused by domain differences. This allows intelligent pathology auxiliary diagnostic systems under 5G environments to truly become a "powerful assistant" for pathologists, further reducing their workload. In particular, it can provide precise auxiliary diagnostic support for primary healthcare institutions, narrowing the gap in pathology diagnostic levels between hospitals of different regions and levels, and promoting the equitable distribution of high-quality pathology diagnostic resources.

[0096] The intelligent pathology-assisted diagnostic system in the 5G environment of this invention will promote the standardization of pathological slide preparation, digital imaging, and model training, facilitate the interoperability and sharing of pathological data among different medical institutions, break down data silos, lay the foundation for establishing national and regional digital pathology databases, and further empower scientific research and innovation in pathology. On the one hand, standardized data and a high-performance intelligent pathology-assisted diagnostic system in the 5G environment can accelerate basic research and clinical translation related to pathology, help explore the pathogenesis and lesion characteristics of diseases in depth, and provide a more scientific basis for early screening, accurate classification, and prognostic assessment of diseases. On the other hand, it can promote the large-scale application of intelligent pathology systems, drive the transformation of pathological diagnosis from "human-led" to "human-machine collaboration," improve the efficiency and quality of the entire pathology diagnosis industry, and help the deep implementation of the concept of precision medicine.

[0097] The efficiency improvements of this invention can shorten the pathological diagnosis cycle, allowing patients to obtain diagnostic results more quickly. This buys valuable time for early intervention and treatment, improving patients' treatment outcomes and quality of life, while reducing the medical burden caused by repeated examinations and misdiagnosis. This invention can reduce labor costs and waste of consumables in the pathological diagnosis process, optimize the allocation of medical resources, promote refined management in the medical industry, and drive the development of related industries such as digital pathology equipment. It injects new momentum into the medical technology industry, achieving a win-win situation for pathology, the medical industry, and patients.

[0098] This invention includes a CT image acquisition module for acquiring CT images of patients using a CT scanner; a CT image processing module for processing the acquired patient CT images to obtain processed patient CT images, which serve as a training set; a deep learning module for constructing a deep learning model, obtaining a trained deep learning model based on the training set; and a recognition module for recognizing the processed patient CT images to be tested based on the trained deep learning model. This effectively overcomes domain differences caused by different hospitals' slide preparation, staining, and scanning equipment, ensuring the model remains stable and accurate across institutional data, supporting large-scale clinical implementation. It reduces missed diagnoses and misdiagnoses due to data differences, assists pathologists in accurately identifying lesions, and reduces subjective human error. It provides reliable auxiliary diagnosis for primary hospitals, narrowing the gap in pathological diagnosis levels between hospitals of different regions and levels, and promoting the equalization of high-quality resources. It promotes unified standards for slide preparation, digital imaging, and model training, breaking down data silos and laying the foundation for establishing a national / regional digital pathology database. It provides faster diagnostic results, allowing more time for early treatment, improving treatment outcomes and quality of life, and reducing the medical burden caused by repeated examinations and misdiagnoses. Reducing labor costs and waste of consumables will drive the development of the digital pathology equipment industry and achieve a win-win situation for pathology, the medical industry, and patients.

[0099] In summary, the intelligent pathology-assisted diagnosis system of this invention in a 5G environment exhibits strong generalization performance and high recognition accuracy for different pathological images, solving the problem that existing AI-assisted pathology diagnosis models have limited generalization performance and cannot maintain the system's recognition accuracy for digital pathological images. Attached Figure Description

[0100] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0101] Specific Implementation Method 1: The intelligent pathology auxiliary diagnosis system in this 5G environment includes: a CT image acquisition module, a CT image processing module, a deep learning module, and a recognition module;

[0102] The CT image acquisition module is used to acquire CT images of the patient by the CT scanner; the CT images are CT images of the patient's lungs.

[0103] The CT image processing module is used to process the acquired patient CT images to obtain processed patient CT images, which are then used as a training set.

[0104] The deep learning module is used to build deep learning models and obtain trained deep learning models based on the training set.

[0105] The recognition module is used to recognize the processed patient CT images to be tested based on a trained deep learning model; the CT images are lung CT images of the patient.

[0106] CT images: Digital grayscale tomographic images, which distinguish tissue density by black, white and gray tones. They can be quantitatively assessed by CT values ​​(unit: HU). Bones appear white (high density, CT value +1000), gases appear black (low density, CT value -1000), and soft tissues appear in different shades of gray. They are usually axial tomographic images, which can clearly show minute lesions in the body (millimeter level). There is no tissue image overlap, which is convenient for the localization and feature analysis of pathological lesions.

[0107] The category label for a CT image is the pathology or tissue category label of the center pixel of the 3D cube image.

[0108] The pathological or tissue category of the center pixel in a 3D cube image includes: normal tissue, benign lesion, malignant lesion / tumor, background / non-target area; / represents or;

[0109] Normal tissues such as normal lungs, liver, kidneys, and soft tissues;

[0110] Benign lesions such as inflammation, nodules, and hyperplasia;

[0111] Malignant lesions / tumors such as adenocarcinoma, squamous cell carcinoma, small cell carcinoma, metastatic carcinoma, etc.

[0112] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the CT image processing module is used to process the acquired patient CT images to obtain processed patient CT images, which are then used as a training set; the specific process is as follows:

[0113] Access medical databases from different hospitals or medical institutions, and extract continuous two-dimensional pathological slide data of patients scanned by CT scanners (Computed Tomography) from the medical databases;

[0114] The surface rendering method is used to construct a three-dimensional CT image from consecutive two-dimensional slices of each patient;

[0115] The standard format for exporting from hospital CT / PACS is 2D for a single image and 3D for the entire set.

[0116] The 3D CT image is divided into a series of 3D cube images;

[0117] The category label of the center pixel of each 3D cube image is used as the category of the corresponding 3D cube image;

[0118] All 3D cube images were used as the training set.

[0119] Each 3D cube image is used as input to the model.

[0120] This invention utilizes surface rendering to construct three-dimensional images from continuous two-dimensional CT slices, restoring the true spatial morphology, infiltration range, and tissue relationships of lesions. This overcomes the limitation of two-dimensional slices in representing spatial structure, significantly improving the accuracy of benign / malignant tumor identification and tumor subtyping. The three-dimensional volume data is divided into regular three-dimensional cubic images, with the center pixel category used as a label, forming a standardized and clearly labeled training set. This solves the problem of inconsistent data formats, sizes, and storage across different hospitals, providing a standardized data foundation for cross-institutional model training. Large-volume CT data is split into small cubic input units, reducing the computational requirements for model training while ensuring that each sample focuses on the local lesion area, improving feature learning efficiency and training stability. Data from multiple hospital medical databases can be directly accessed and processed uniformly, breaking down data silos and allowing the model to be trained on data from multiple institutions and devices. This alleviates domain differences at the data level and improves model generalization performance. Using center pixel labels instead of full-image annotation simplifies the annotation process, reduces annotation errors, and facilitates the rapid construction of a large-scale, high-quality digital pathology training library, supporting the clinical deployment of the system.

[0121] The other steps and parameters are the same as in Specific Implementation Method 1.

[0122] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that the deep learning module is used to construct a deep learning model and obtain a trained deep learning model based on the training set; the specific process is as follows:

[0123] 1) Constructing a deep learning model; the specific process is as follows:

[0124] Deep learning models include: convolutional kernel size of The depthwise convolutional layer has a kernel size of 1. The depthwise convolutional layer has a kernel size of 1. The depthwise convolutional layer has a kernel size of 1. The 2D convolutional layer, BN layer, GELU layer, Mamba block, global average pooling layer, linear layer, and softmax activation function layer;

[0125] 2) Obtain a trained deep learning model based on the training set; the specific process is as follows:

[0126] 21) For each 3D cube image in the training set Processing along the horizontal direction yields a one-dimensional sequence. ;

[0127] 22) For each 3D cube image in the training set Processing along the vertical direction yields a one-dimensional sequence. ;

[0128] 23) For each 3D cube image in the training set Processing along the horizontal direction yields a one-dimensional sequence. ;

[0129] 24) For each 3D cube image in the training set Processing along the horizontal direction yields a one-dimensional sequence. ;

[0130] 25) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. ;

[0131] 26) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. The process is the same as in 25);

[0132] 27) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. The process is the same as in 25);

[0133] 28) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. The process is the same as in 25);

[0134] 29) One-dimensional sequence One-dimensional sequence One-dimensional sequence One-dimensional sequence By adding elements one by one, we obtain the feature map. ;

[0135] For feature maps The feature map is obtained by reshaping the image; the feature map is then input into a global average pooling layer, which outputs the feature map. ;

[0136] feature map The input is sequentially processed into a linear layer and a softmax activation function layer. The softmax activation function layer outputs the final predicted probability distribution.

[0137] This implementation constructs a deep learning model comprising multi-scale deep convolutional layers, BN layers, GELU layers, Mamba blocks, global average pooling layers, linear layers, and Softmax activation function layers. Employing a four-way parallel feature extraction, sequence processing, and feature fusion process, it can fully extract multi-directional, multi-scale spatial features from 3D CT pathological images. The Mamba blocks efficiently capture long-range spatial dependencies, enhancing the model's learning and representation capabilities for lesion features. Simultaneously, through multi-path feature complementarity and fusion, it effectively reduces the impact of differences in pathological image domains between different medical institutions on model performance, enhancing model generalization and diagnostic stability. Ultimately, it achieves accurate classification and identification of normal tissues, benign lesions, and malignant lesions, providing high-performance algorithm support for intelligent pathology-assisted diagnosis in a 5G environment.

[0138] Other steps and parameters are the same as in specific implementation method one or two.

[0139] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that, in step 21), each three-dimensional cube image in the training set... Processing along the horizontal direction yields a one-dimensional sequence. The specific process is as follows:

[0140] Each 3D cube image in the training set The input convolution kernel size along the horizontal direction is The deep convolutional layer outputs a feature map. ;

[0141] feature map and After element-wise addition, the input convolution kernel size is along the horizontal direction. The deep convolutional layer outputs a feature map. ;

[0142] feature map , and After element-wise addition, the input convolution kernel size is along the horizontal direction. The deep convolutional layer outputs a feature map. ;

[0143] Image Feature map Feature map Feature map By adding elements one by one, we obtain the feature map. ;

[0144] Feature map along the horizontal axis Flattening the sequence yields a one-dimensional sequence. .

[0145] This invention uses , , Three types of convolutional kernels capture tissue textures, cell arrangements, and lesion extension features of varying lengths in the horizontal direction, covering horizontal structural information from minute to macroscopic. The original image and the outputs of each layer are superimposed element-wise to construct a residual structure, preserving key details at the lower levels and avoiding the loss of subtle lesion features in deep convolutions. The original image and multi-scale outputs are then fused to obtain a clean and complete horizontally specific feature map, highlighting the patterns of lesions in the horizontal direction.

[0146] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0147] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that, in step 22), each three-dimensional cube image in the training set... Processing along the vertical direction yields a one-dimensional sequence. The specific process is as follows:

[0148] Each 3D cube image in the training set The input convolution kernel size is along the vertical direction. The deep convolutional layer outputs a feature map. ;

[0149] feature map and After element-wise addition, the input convolution kernel size is along the vertical direction. The deep convolutional layer outputs a feature map. ;

[0150] feature map , and After element-wise addition, the input convolution kernel size is along the vertical direction. The deep convolutional layer outputs a feature map. ;

[0151] Image Feature map Feature map Feature map By adding elements one by one, we obtain the feature map. ;

[0152] Feature map along the vertical axis Flattening the sequence yields a one-dimensional sequence. .

[0153] This invention also uses , , Convolution is used to extract information on tissue thickness, cell layering, and lesion invasion depth in the vertical direction, supplementing vertical structures not covered in the horizontal direction. Residual superposition preserves subtle lesions in the vertical direction, avoids gradient vanishing, and improves sensitivity to minor vertical anomalies. This results in a vertically specific feature map that focuses on the morphology of tissues and lesions in the vertical direction, resulting in more focused features and lower noise.

[0154] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0155] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that, in step 23), each three-dimensional cube image in the training set... Processing along the horizontal direction yields a one-dimensional sequence. The specific process is as follows:

[0156] 231) For each 3D cube image in the training set Each channel in the process undergoes feature shifting to form a feature map. The specific process is as follows:

[0157] Feature map In a single channel The characteristics of the row are ,in ; Representation of feature map The size of its height or width;

[0158] The first Characteristics of lines Divided into part 1 and part 2; part 1 represents the first part... Characteristics of lines The segmented head; part 2 indicates the segmented head. Characteristics of lines The split tail section;

[0159] Assign part 1 to the new row feature From the position To the final characteristics;

[0160] Assign part 2 to the new row feature. From position 0 to Features;

[0161] Assign the value 0 to the new row feature. From position 0 to Features;

[0162] Assign the value 0 to the new row feature. From the position To the final characteristics;

[0163] Feature map Each channel in the process undergoes feature shifting to form a feature map. ;

[0164] 232) Feature map based on horizontal axis The process is the same as in 21), resulting in a one-dimensional sequence. ;

[0165] For the obtained one-dimensional sequence The former One and after The features are cropped to obtain the cropped one-dimensional sequence. .

[0166] By performing channel-level feature shifting on 3D cubic images, the model simulates minute shifts in pathological images during scanning, staining, and slice thickness. This allows the model to learn core lesion features that remain invariant in position and scale, reducing the interference of imaging detail differences on recognition. Expanding the effective receptive field without increasing the convolution kernel or computational power, the model utilizes feature shifting to obtain a larger receptive field, better capturing long-range spatial correlations of cells or tissues and improving the detection capability of minute lesions. Redundant noise is removed by pruning the front and back ends of the one-dimensional sequence, removing invalid or noisy edge features, retaining core lesion information, reducing interference, and improving feature purity and training stability. Adapting to multi-institutional data significantly alleviates domain shift caused by differences in image style and quality across different hospitals, allowing the model to maintain high accuracy on cross-institutional data.

[0167] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0168] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One to Six in that, in step 24), each three-dimensional cube image in the training set... Processing along the horizontal direction yields a one-dimensional sequence. The specific process is as follows:

[0169] 241) Regarding the feature map Transpose the horizontal and vertical dimensions of the spatial dimension;

[0170] Each channel in the transposed feature map is feature-shifted to form a new feature map. ;

[0171] 242) For feature maps Perform dimensionality reversal, then process the dimensionally reversed feature map based on the vertical axis, as described in section 22), to obtain a one-dimensional sequence. ;

[0172] For the obtained one-dimensional sequence The former One and after The features are cropped to obtain the cropped one-dimensional sequence. .

[0173] First, spatial transpose is performed in the horizontal or vertical direction, followed by feature shifting and dimension reversal. This extracts features from all dimensions, including vertical, horizontal, diagonal, and oblique, overcoming the limitation of conventional convolution which only extracts features along a fixed direction. It adapts to CT 3D pathological data, enabling the model to understand the spatial distribution and connectivity of lesions in in vivo data, improving the accuracy of benign / malignant diagnosis and tumor subtyping. Multiple spatial transformations significantly reduce image distribution shifts caused by differences in equipment and processes, significantly improving generalization performance and supporting large-scale clinical deployment. Dimension reversal combined with end-point cropping compresses feature length while preserving key information, accelerating inference speed and reducing computational requirements, adapting to real-time assisted diagnosis in 5G environments. (State Intellectual Property Office)

[0174] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0175] Specific Implementation Method Eight: This implementation method differs from one of Specific Implementation Methods One to Seven in that, in step 25), the one-dimensional sequence... After processing, a one-dimensional sequence is obtained. The specific process is as follows:

[0176] 251) For one-dimensional sequences Perform reshape deformation, and then input the reshaped sequence sequentially into the convolution kernel of size [size missing]. The 2D convolutional layer, BN layer, and GELU layer output feature maps. ;

[0177] 252) Feature map Input Mamba block, output feature map ;

[0178] 253) Regarding feature maps Perform feature reversal to obtain the feature map after feature reversal. ;

[0179] 254) Feature map Input Mamba block, output feature map The process is the same as 252);

[0180] 258) Feature map and feature map By adding elements one by one, a one-dimensional sequence is obtained. .

[0181] This implementation method performs dimensionality reshaping, pointwise convolution, normalization, and activation processing on one-dimensional sequences to achieve feature dimension regularization, normalization, and nonlinear activation, ensuring a uniform distribution of features from different directions and institutions. Furthermore, it employs bidirectional Mamba and reverse Mamba sequence modeling to fully exploit long-range spatial dependencies in pathological images, enhance effective lesion features, and suppress noise interference. Simultaneously, it standardizes and purifies sequence features, significantly improving the model's adaptability to differences in cross-institutional pathological image domains, enhancing feature expression capabilities and diagnostic robustness, and laying a high-quality feature foundation for subsequent multi-path feature fusion and accurate classification.

[0182] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0183] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One to Eight in that, in step 252), the feature map... Input Mamba block, output feature map The process is as follows:

[0184] feature map Perform two-dimensional pointwise convolution operations sequentially. Shape reshaping operation Interchange the channel dimension and sequence dimension (e.g.) ( It is the channel dimension. (It is the sequence dimension) becomes , ( It is the channel dimension. (This is the sequence dimension), resulting in sequence 1;

[0185] Perform on sequence 1 Generate the reversed input sequence 2;

[0186] Sequence 1 will be processed sequentially. Two-dimensional pointwise convolution operation Interchange the channel dimension and sequence dimension (e.g.) ( It is the channel dimension. (It is the sequence dimension) becomes , ( It is the channel dimension. (This is the sequence dimension) to obtain sequence 3;

[0187] Sequence 2 passes through in sequence Two-dimensional pointwise convolution operation is performed to obtain sequence 4;

[0188] The feature map is obtained by concatenating sequences 3 and 4. ;

[0189] This indicates a Mamba framework based on SSM, used to capture long-range spatial dependencies.

[0190] This implementation method, based on the SSM's Mamba framework, efficiently captures the spatial correlation between distant tissues and lesions, solving the problem that traditional CNNs can only see local areas and cannot clearly see large-scale infiltrative lesions. Forward Mamba is used to extract forward spatial features from sequence 1, and reverse Mamba is used to extract inverse spatial features from sequence 2. After bidirectional stitching, the lesion features are more complete, the boundaries are clearer, and the classification is more accurate. Through pointwise convolution, reshape, and permute dimensional transformation, the features are adapted to the 3D CT volumetric data format, resulting in a more regular feature distribution and more stable training. Reinforcement learning utilizes the essential structural features of the lesion, mitigating image style differences caused by staining, scanning, and slice thickness, significantly improving the model's diagnostic stability on data from different hospitals. Bidirectional feature complementarity and superposition automatically suppress scanning noise and artifact interference, outputting high signal-to-noise ratio feature maps.

[0191] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0192] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One through Ninety-Eight in that the loss function acquisition process of the deep learning model is as follows:

[0193] 1) Extract the pseudo-labels from the output of the softmax activation function layer along the category dimension. The maximum value and the former Category position index corresponding to the maximum value ;

[0194] 2) The weights are obtained after passing through the softmax activation function layer. ;

[0195] 2) Based on weight get Supervision items with topk weighted cross-entropy ; indicates as:

[0196]

[0197] in, Indicates the first preprocessed step The first image corresponds to the Supervision items of topk weighted cross-entropy, Indicates the first preprocessed step The topk weights corresponding to the images are as follows: Each weight value Indicates the first preprocessed step The topk index corresponding to the image is the first One index value; express Categories ; express Middle Count , ; This indicates the total number of preprocessed images; Indicates the first preprocessed step Zhang Image ; Indicates the total number of categories; express Middle count, ; Indicates an indicator function, hour Take 1, hour Set to 0;

[0198] 3) Based on Supervision items with topk weighted cross-entropy Calculate loss ; indicates as:

[0199]

[0200] in, Indicates the first The supervising term for the topk weighted cross-entropy; the superscript T indicates the transpose; This represents the preprocessed image input into the neural network, and the classification pseudo-label output by the neural network.

[0201] This implementation extracts and weights the top-k maximum predicted probabilities, allowing the model to focus more on high-confidence, high-value diagnostic categories and improving classification reliability. A weighted cross-entropy supervision term filters low-confidence noise labels, reducing training interference from labeling errors and image differences, resulting in a more accurate and stable model. On datasets from multiple hospitals, with multiple devices, and multiple staining styles, adaptive alignment of data distribution significantly improves model generalization performance and avoids sharp performance drops across different domains. A smoother loss function and more stable gradients prevent overfitting and accelerate convergence, allowing the model to maintain high accuracy even with limited labeled data. It enhances the ability to distinguish between benign, malignant, and normal tissues, reducing missed diagnoses and misdiagnoses, and assisting pathologists in making more reliable judgments.

[0202] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0203] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

A smart pathology-assisted diagnostic system under 1.5G environment, characterized by: The system includes: a CT image acquisition module, a CT image processing module, a deep learning module, and a recognition module; The CT image acquisition module is used to acquire CT images of the patient by the CT scanner; The CT image processing module is used to process the acquired patient CT images to obtain processed patient CT images, which are then used as a training set. The deep learning module is used to build deep learning models and obtain trained deep learning models based on the training set. The recognition module is used to identify the processed patient CT images to be tested based on a trained deep learning model.

2. The intelligent pathology-assisted diagnostic system in a 5G environment according to claim 1, characterized in that: The CT image processing module is used to process the acquired patient CT images to obtain processed patient CT images, which are then used as a training set. The specific process is as follows: Access medical databases from different hospitals or medical institutions, and extract continuous two-dimensional pathological slide data of patients scanned by CT scanners from the medical databases; The surface rendering method constructs a three-dimensional CT image from consecutive two-dimensional slices of each patient; The 3D CT image is divided into a series of 3D cube images; The category label of the center pixel of each 3D cube image is used as the category of the corresponding 3D cube image; All 3D cube images were used as the training set.

3. The intelligent pathology-assisted diagnostic system in a 5G environment according to claim 2, characterized in that: The deep learning module is used to construct a deep learning model and obtain a trained deep learning model based on the training set; the specific process is as follows: 1) Construct a deep learning model; The specific process is as follows: Deep learning models include: convolutional kernel size of The depthwise convolutional layer has a kernel size of 1. The depthwise convolutional layer has a kernel size of 1. The depthwise convolutional layer has a kernel size of 1. The 2D convolutional layer, BN layer, GELU layer, Mamba block, global average pooling layer, linear layer, and softmax activation function layer; 2) Obtain a trained deep learning model based on the training set; the specific process is as follows: 21) For each 3D cube image in the training set Processing along the horizontal direction yields a one-dimensional sequence. ; 22) For each 3D cube image in the training set Processing along the vertical direction yields a one-dimensional sequence. ; 23) For each 3D cube image in the training set Processing along the horizontal direction yields a one-dimensional sequence. ; 24) For each 3D cube image in the training set Processing along the horizontal direction yields a one-dimensional sequence. ; 25) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. ; 26) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. ; 27) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. ; 28) For one-dimensional sequences After processing, a one-dimensional sequence is obtained. ; 29) One-dimensional sequence One-dimensional sequence One-dimensional sequence One-dimensional sequence By adding elements one by one, we obtain the feature map. ; For feature maps The feature map is obtained by reshaping the image; the feature map is then input into a global average pooling layer, which outputs the feature map. ; feature map The input is sequentially processed by a linear layer and a softmax activation function layer. The softmax activation function layer outputs the final predicted probability distribution.

4. The intelligent pathology-assisted diagnostic system in a 5G environment according to claim 3, characterized in that: In step 21), each three-dimensional cube image in the training set... Processing along the horizontal direction yields a one-dimensional sequence. ; The specific process is as follows: Each 3D cube image in the training set The input convolution kernel size along the horizontal direction is The deep convolutional layer outputs a feature map. ; feature map and After element-wise addition, the input convolution kernel size is along the horizontal direction. The deep convolutional layer outputs a feature map. ; feature map , and After element-wise addition, the input convolution kernel size is along the horizontal direction. The deep convolutional layer outputs a feature map. ; Image Feature map Feature map Feature map By adding elements one by one, we obtain the feature map. ; Feature map along the horizontal axis Flattening the sequence yields a one-dimensional sequence. .

5. The intelligent pathology-assisted diagnostic system in a 5G environment according to claim 4, characterized in that: In step 22), each three-dimensional cube image in the training set Processing along the vertical direction yields a one-dimensional sequence. ; The specific process is as follows: Each 3D cube image in the training set The input convolution kernel size is along the vertical direction. The deep convolutional layer outputs a feature map. ; feature map and After element-wise addition, the input convolution kernel size is along the vertical direction. The deep convolutional layer outputs a feature map. ; feature map , and After element-wise addition, the input convolution kernel size is along the vertical direction. The deep convolutional layer outputs a feature map. ; Image Feature map Feature map Feature map By adding elements one by one, we obtain the feature map. ; Feature map along the vertical axis Flattening the sequence yields a one-dimensional sequence. .

6. The intelligent pathology-assisted diagnostic system in a 5G environment according to claim 5, characterized in that: In step 23), each three-dimensional cube image in the training set Processing along the horizontal direction yields a one-dimensional sequence. The specific process is as follows: 231) For each 3D cube image in the training set Each channel in the process undergoes feature shifting to form a feature map. The specific process is as follows: Feature map In a single channel The characteristics of the row are ,in ; Representation of feature map The size of its height or width; The first Characteristics of lines Divided into part 1 and part 2; part 1 represents the first part... Characteristics of lines The segmented head; part 2 indicates the segmented head. Characteristics of lines The split tail section; Assign part 1 to the new row feature From the position To the final characteristics; Assign part 2 to the new row feature. From position 0 to Features; Assign the value 0 to the new row feature. From position 0 to Features; Assign the value 0 to the new row feature. From the position To the final characteristics; Feature map Each channel in the process undergoes feature shifting to form a feature map. ; 232) Feature map based on horizontal axis After processing, a one-dimensional sequence is obtained. ; For the obtained one-dimensional sequence The former One and after The features are cropped to obtain the cropped one-dimensional sequence. .

7. The intelligent pathology-assisted diagnostic system in a 5G environment according to claim 6, characterized in that: In step 24), each three-dimensional cube image in the training set will be... Processing along the horizontal direction yields a one-dimensional sequence. The specific process is as follows: 241) Regarding the feature map Transpose the horizontal and vertical dimensions of the spatial dimension; Each channel in the transposed feature map is feature-shifted to form a new feature map. ; 242) For feature maps Dimension inversion is performed, and the feature map after dimension inversion is processed based on the vertical axis to obtain a one-dimensional sequence. ; For the obtained one-dimensional sequence The former One and after The features are cropped to obtain the cropped one-dimensional sequence. .

8. The intelligent pathology-assisted diagnostic system in a 5G environment according to claim 7, characterized in that: In 25), the one-dimensional sequence After processing, a one-dimensional sequence is obtained. The specific process is as follows: 251) For one-dimensional sequences Perform reshape deformation, and then input the reshaped sequence sequentially into the convolution kernel of size [size missing]. The 2D convolutional layer, BN layer, and GELU layer output feature maps. ; 252) Feature map Input Mamba block, output feature map ; 253) Regarding feature maps Perform feature reversal to obtain the feature map after feature reversal. ; 254) Feature map Input Mamba block, output feature map ; 258) Feature map and feature map By adding elements one by one, a one-dimensional sequence is obtained. .

9. The intelligent pathology-assisted diagnostic system in a 5G environment according to claim 8, characterized in that: The feature map in 252) Input Mamba block, output feature map The process is as follows: feature map Perform two-dimensional pointwise convolution operations sequentially. Shape reshaping operation By swapping the channel dimension and the sequence dimension, we obtain sequence 1. Perform on sequence 1 Generate the reversed input sequence 2; Sequence 1 will be processed sequentially. Two-dimensional pointwise convolution operation Interchange the channel dimension and the sequence dimension to obtain sequence 3; Sequence 2 passes through in sequence Two-dimensional pointwise convolution operation is performed to obtain sequence 4; The feature map is obtained by concatenating sequences 3 and 4. ; This indicates the Mamba framework based on SSM.

10. The intelligent pathology-assisted diagnostic system in a 5G environment according to claim 9, characterized in that: The process of obtaining the loss function of the deep learning model is as follows: 1) Extract the pseudo-labels from the output of the softmax activation function layer along the category dimension. The maximum value and the former Category position index corresponding to the maximum value ; 2) The weights are obtained after passing through the softmax activation function layer. ; 2) Based on weight get Supervision items with topk weighted cross-entropy ; indicates as: in, Indicates the first preprocessed step The first image corresponds to the Supervision items of topk weighted cross-entropy, Indicates the first preprocessed step The topk weights corresponding to the images are as follows: Each weight value Indicates the first preprocessed step The topk index corresponding to the image is the first One index value; express Categories ; express Middle Count , ; This indicates the total number of preprocessed images; Indicates the first preprocessed step Zhang Image ; Indicates the total number of categories; express Middle count, ; Indicates an indicator function, hour Take 1, hour Set to 0; 3) Based on Supervision items with topk weighted cross-entropy Calculate loss ; indicates as: in, Indicates the first The supervising term for the topk weighted cross-entropy; the superscript T indicates the transpose; This represents the preprocessed image input into the neural network, and the classification pseudo-label output by the neural network.

Images