Intelligent analysis device for pathological image of testicular nucleoprotein cancer
By employing techniques such as adaptive Kalman filtering and Transformer cross-modal attention mechanism, the problems of signal non-stationarity and insufficient modeling of multi-source heterogeneous signals in testicular nucleoprotein carcinoma detection have been solved. This has enabled high-precision, fully automated pathological image analysis, adapting to the complexity and individual variability of biomedical signals, and improving the accuracy and reliability of detection.
Patent Information
- Application Number
- CN202511648982.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-02-13
AI Technical Summary
Existing pathological image analysis systems face challenges in detecting testicular nucleoprotein carcinoma, including issues such as non-stationarity of biological signals, non-linearity, multi-scale, low signal-to-noise ratio, high physiological noise interference, and insufficient deep correlation modeling of multi-source heterogeneous biological signals. These problems result in low detection accuracy and slow speed.
Adaptive Kalman filtering is combined with preprocessing methods such as empirical mode decomposition, multivariate scattering correction, RMA algorithm and nonlocal mean filtering. It is combined with multi-scale wavelet packet decomposition, mass spectrum peak identification quantization and deep learning feature encoder. Knowledge graph and Transformer cross-modal attention mechanism are introduced to construct a deep forest cascade classification model to realize multi-dimensional and multi-scale feature extraction and deep-level association modeling.
It significantly improves the accuracy and reliability of testicular nucleoprotein cancer detection, automates the entire process from image acquisition to diagnosis, adapts to the complexity and individual variability of biomedical signals, and provides a system-level solution for early and accurate diagnosis.
Smart Images
Figure CN121521858A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of pathological image analysis, and particularly relates to a testicular nuclear protein cancer pathological image intelligent analysis device. BACKGROUND
[0002] As a malignant tumor with specific biological characteristics, the early accurate detection and diagnosis of testicular nuclear protein cancer has significant clinical implications for improving patient treatment outcomes and survival rates. With the continuous progress of medical imaging, biomarker analysis, and various physiological signal monitoring technologies worldwide, the application of advanced signal processing and pattern recognition technologies to disease detection, especially the integration of multi-source heterogeneous signals to achieve more comprehensive and accurate pathological state evaluation, has gradually become a hot topic in the cross-research of biomedical engineering and clinical medicine. The development of such technologies aims to extract key information from complex and variable biological signals, thereby providing objective and quantitative diagnostic evidence for clinicians.
[0003] Meanwhile, testicular nuclear protein cancer pathological analysis is also used to assist pathologists in diagnosing and analyzing testicular tumors, especially those that may involve abnormal nuclear proteins, such as seminomas and embryonal carcinomas. Testicular nuclear protein cancer is a highly malignant tumor originating from germ cells, and its pathological diagnosis relies on the comprehensive analysis of cell morphology, staining conditions, and nuclear protein expression characteristics under a microscope. Traditional pathological diagnosis methods mainly rely on pathologists observing and manually judging tissue sections under a microscope, which is not only time-consuming and subjective, but also easily affected by experience level and observation angle.
[0004] With the rise of digital pathology and artificial intelligence technologies, the automatic recognition and analysis of pathological images have become an important direction in clinical diagnosis, especially deep learning models such as convolutional neural networks (CNN) and visual Transformers (ViT) have shown excellent performance in medical image fields. However, existing pathological intelligent analysis systems are mostly aimed at common tumors such as breast cancer, lung cancer, and cervical cancer, and there is insufficient research on the feature recognition of testicular nuclear protein cancer.
[0005] The pathological images of such tumors have obvious heterogeneity: significant nuclear atypia, complex staining distribution, and high background tissue complexity. Conventional image recognition algorithms have difficulty accurately distinguishing between cancerous and normal tissue regions, especially in immunohistochemical staining images, where nuclear protein signal strength varies, leading to insufficient model stability.
[0006] In addition, pathological section data is large, with a single image resolution of hundreds of millions of pixels, requiring higher computational resources and algorithm optimization. However, existing systems often lack efficient multi-scale feature extraction mechanisms, resulting in slow analysis speed and unstable classification accuracy.
[0007] Therefore, an intelligent device capable of high-precision analysis of testicular nuclear protein cancer pathological characteristics is urgently needed to realize the automation of the whole process from image acquisition, preprocessing, feature extraction to intelligent diagnosis, thereby assisting pathologists to quickly, objectively and accurately complete the diagnosis task
[0008] To this end, we designed an intelligent testicular nuclear protein cancer pathological image analysis device. SUMMARY
[0009] The purpose of the present application is to solve the technical contradictions in the detection of testicular nuclear protein cancer in the prior art, such as non-stationary, nonlinear, multi-scale, low signal-to-noise ratio, high physiological noise interference, and insufficient deep correlation modeling of multi-source heterogeneous biological signals, etc. The present application provides a testicular nuclear protein cancer detection system that fuses signal pattern recognition. The present application aims to optimize the biological signal preprocessing process, enhance the deep fusion capability of multi-source heterogeneous biological signals, and construct a pattern recognition model with high precision and strong robustness, thereby improving the accuracy and reliability of testicular nuclear protein cancer detection, and providing objective and quantitative technical support for clinical diagnosis. An intelligent testicular nuclear protein cancer pathological image analysis device is proposed.
[0010] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:
[0011] An intelligent testicular nuclear protein cancer pathological image analysis device, comprising an operating table and a microscope analyzer arranged on the operating table, an observation glass window is arranged below the operating table, an observation port is formed on the operating table, a glass slide placing base is arranged on the observation port, a cover glass is arranged on the glass slide placing base, a turnover mechanical hand assembly for overturning the glass slide placing base is arranged below the operating table, and the turnover mechanical hand assembly reverses the glass slide placing base on the observation glass window.
[0012] Preferably, a partition is arranged on the side of the observation glass window facing the operating table, the observation glass window is divided into multiple placement cavities by the partition, and multiple microscope display screens are arranged on the placement cavities.
[0013] Preferably, a placing frame is arranged on the observation glass window, and the placing frame is connected to the operating table through a connecting rod.
[0014] Preferably, a fixed shaft is arranged on the operating table, the fixed shaft is arranged in pairs, and the fixed shafts are symmetrically arranged on both sides of the observation port, an elastic pressing piece for pressing the cover glass is rotatably arranged on the fixed shaft.
[0015] Preferably, a first rotating shaft is further arranged at the bottom of the operating table, a rotating base is further arranged on the first rotating shaft, and the rotating base moves on the first rotating shaft through a penetrating sliding hole.
[0016] Preferably, the operation base is arranged on one side of the rotating base, and the operation base horizontally slides on the rotating base through a limiting rod, and the rotating base is internally provided with a driving mechanism.
[0017] Preferably, the operation base is arranged on one side of the rotating base, and the operation base horizontally slides on the rotating base through a limiting rod, and the rotating base is internally provided with a driving mechanism.
[0018] Preferably, the operation base is arranged on one side of the rotating base, and the operation base horizontally slides on the rotating base through a limiting rod, and the rotating base is internally provided with a driving mechanism.
[0019] Preferably, the operation base is arranged on one side of the rotating base, and the operation base horizontally slides on the rotating base through a limiting rod, and the rotating base is internally provided with a driving mechanism.
[0020] Preferably, the operation base is arranged on one side of the rotating base, and the operation base horizontally slides on the rotating base through a limiting rod, and the rotating base is internally provided with a driving mechanism.
[0021] The beneficial effects of the present application are:
[0022] 1、The cover glass and the glass slide placing base are used to place corresponding testicular nuclear protein cancer tissues, and then the mechanical hand assembly is used to place the glass slide placing base upside down on the observation glass window, so that the glass slide placing base can be fixed and supported, and conditions for microscope analysis are provided.
[0023] 2、The image acquisition module is used for high-resolution acquisition of testicular nuclear protein cancer tissue section images, and the microscopic display screen is used for displaying the testicular nuclear protein cancer tissue section images.
[0024] 3、The adaptive Kalman filter is combined with the experience mode decomposition, the multivariate scattering correction, the RMA algorithm, and the non-local mean filter to realize the preprocessing method of biological characteristics perception, so that high noise and complex artifacts in biological signals can be effectively inhibited, key biological information can be retained, and the problem of poor adaptability of the prior art preprocessing to biological signals is overcome; secondly, the multi-scale wavelet packet decomposition, mass spectrum peak identification quantization, differential gene analysis, radiomics, and deep learning feature encoder are used to extract multi-dimensional and multi-scale biological key features from various modal biological signals, so that the comprehensiveness and delicacy of information are ensured.
[0025] 4、The application initiatively introduces a hierarchical fusion strategy of a graph neural network and a Transformer cross-modal attention mechanism based on a knowledge graph, realizes effective modeling and utilization of deep and nonlinear correlations between different biological modalities, which significantly surpasses the limitation of simple splicing or shallow correlation of the prior art at the feature level, so as to mine key biological signal paths specific to TNPC from complex biological networks; finally, the application constructs a cascade classification model based on a deep forest, combines a Monte Carlo Dropout method to realize uncertainty quantification, and combines a SHAP algorithm to realize explainability analysis, thereby significantly improving the precision, robustness and clinical usability of TNPC diagnosis.
[0026] 5、The system can adapt to the inherent high complexity and individual variability of biomedical signals, and provides a new system-level solution for early, accurate and reliable diagnosis of TNPC. The application realizes biological background perception, multi-scale information mining and deep semantic fusion in the whole process from biological signal acquisition to final diagnosis output, and the inherent logical rigor and technical innovation make it have significant non-obviousness. BRIEF DESCRIPTION OF DRAWINGS
[0027] Figure 1 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application;
[0028] Figure 2 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application; Figure 1 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application;
[0029] Figure 3 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application;
[0030] Figure 4 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application; Figure 3 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application;
[0031] Figure 5 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application;
[0032] Figure 6 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application;
[0033] Figure 7 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application;
[0034] Figure 8 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application;
[0035] Figure 9 A structure schematic diagram of a testicular nuclear protein cancer pathological image intelligent analysis device is provided for the application;
[0036] Figure 10 For testicular tissue metastasis section structure micrograph.
[0037] In the figure: 1, operation platform; 2, connecting rod; 3, placing frame; 4, observation glass window; 5, partition; 6, glass slide placing base; 7, cover glass; 8, fixed shaft; 9, elastic pressing piece; 10, microscope analyzer; 11, first rotating shaft; 12, rotating base; 13, through sliding hole; 14, limiting rod; 15, operation base; 16, turnover table; 17, second rotating shaft; 18, electric suction cup. DETAILED DESCRIPTION
[0038] Reference Figures 1-10 A testicular nuclear protein cancer pathological image intelligent analysis device, comprising an operation platform 1 and a microscope analyzer 10 arranged on the operation platform 1, which is composed of a hardware system and a software system, and the two work cooperatively.
[0039] The microscope analyzer 10 is used for observing the section of testicular tissue, and the microscope analyzer 10 comprises a microscope main body, an objective lens group, an adjustable light source system and a digital camera. The microscope analyzer 10 is used for analyzing the corresponding section of testicular tissue and observing in the corresponding tissue section. The objective lens group is a zoom lens group with a magnification of 20x-40x, and the aperture is adjustable. The light source is an LED adjustable multi-band light source, which supports multi-channel imaging in the range of 400-700 nm. The digital camera is a 50 million pixel CMOS sensor with 12-bit color depth. The microscope main body, the objective lens group, the adjustable light source system and the digital camera are all prior art and will not be described in detail here.
[0040] The hardware system and the software system are also arranged on the operation platform 1. The hardware system mainly realizes the collection, storage and preliminary processing of pathological images, and the software system is responsible for image preprocessing, artificial intelligence analysis, feature visualization and result output. The two communicate through a data bus and a local area network to realize task parallelism and efficient cooperation. It should be noted that the combination of the hardware system and the software system includes six unit modules, namely, an image collection module, an image preprocessing module, an artificial intelligence analysis module, a feature visualization module, a result output module and a model training and updating module. The data flow is sequentially transmitted from the image collection module to the image preprocessing module, the artificial intelligence analysis module, the feature visualization module and the result output module, and is periodically optimized through the model training and updating module.
[0041] The source of the above-mentioned testicular nuclear protein cancer tissue section image is obtained by observation of the microscope analyzer 10. The specific observation method is as follows:
[0042] The operating table 1 is provided with a fixed shaft 8, the fixed shaft 8 is provided with two sets and is symmetrically arranged on both sides of the observation port, the elastic pressing plate 9 for pressing the cover glass 7 is rotatably arranged on the fixed shaft 8, the observation glass window 4 is arranged below the operating table 1, the observation port is arranged on the operating table 1, and the cover glass placing base 6 is arranged on the observation port, the cover glass 7 is arranged on the cover glass placing base 6, therefore, the operator can rotate and push the elastic pressing plate 9 on the fixed shaft 8, so that the elastic pressing plate 9 is pressed on the cover glass placing base 6 in the observation port, wherein the cover glass 7 is arranged on the cover glass placing base 6, therefore, the elastic pressing plate 9 can be pressed on the cover glass placing base 6 with the cover glass 7, further, the cover glass 7 and the cover glass placing base 6 carry the corresponding testicular nuclear protein cancer tissue.
[0043] The operating table 1 is provided with a fixed shaft 8, the fixed shaft 8 is provided with two sets and is symmetrically arranged on both sides of the observation port, the elastic pressing plate 9 for pressing the cover glass 7 is rotatably arranged on the fixed shaft 8, the observation glass window 4 is arranged below the operating table 1, the observation port is arranged on the operating table 1, and the cover glass placing base 6 is arranged on the observation port, the cover glass 7 is arranged on the cover glass placing base 6, therefore, the elastic pressing plate 9 can be pressed on the cover glass placing base 6 with the cover glass 7, further, the cover glass 7 and the cover glass placing base 6 carry the corresponding testicular nuclear protein cancer tissue.
[0044] The observation glass window 4 is provided with a placing frame 3, and the placing frame 3 is connected with the operating table 1 through the connecting rod 2, and the first rotating shaft 11 is further arranged at the bottom of the operating table 1, the rotating base 12 is further arranged on the first rotating shaft 11, the operating base 15 is arranged on one side of the rotating base 12, the rotating base 12 moves on the first rotating shaft 11 through the penetrating sliding hole 13, and the driving mechanism is arranged in the rotating base 12, wherein the rotating base 12 rotates and moves up and down on the first rotating shaft 11 under the driving force of the external driving mechanism, wherein the rotation can change the position of the operating base 15 between the operating table 1 and the observation glass window 4; and the up-and-down movement of the rotating base 12 can drive the cover glass placing base 6 to move up and down, thereby providing driving force for the vertical movement of the cover glass placing base 6.
[0045] The operating base 15 horizontally slides on the rotating base 12 through the limiting rod 14, the operating base 15 is provided with a turnover mechanism, the second rotating shaft 17 is rotatably arranged on the operating base 15, the turnover mechanism drives the second rotating shaft 17 to rotate, the turnover base 16 is arranged on the second rotating shaft 17, and the electric suction cup 18 is arranged on the turnover base 16, the turnover mechanism comprises a driving motor, a first driving gear is fixed on the output end of the driving motor in a same axis, and a second driving gear is fixed on the second rotating shaft 17 in a same axis, therefore, the driving motor is started to drive the second rotating shaft 17 to turn over 180°, so that the turnover base 16 with the cover glass placing base 6 rotates, that is, the originally upward cover glass placing base 6 turns over downward, thereby realizing the orientation change and switching of the cover glass placing base 6.
[0046] The observation glass window 4 is provided with a partition 5 on the side facing the operation table 1, and the partition 5 divides the observation glass window 4 into multiple placing cavities, and multiple microscopic display screens are arranged on the placing cavities, so that the inverted glass slide placing base 6 with the cover glass 7 after observation can be placed in the placing cavity, wherein the observation glass window 4 is provided with a pressing device corresponding to the placing cavity, and the pressing device includes a clamping piece for pressing the inverted glass slide placing base 6, so as to avoid the glass slide placing base 6 with the cover glass 7 from being separated.
[0047] Further, the image acquisition module is a multi-spectral microscopic image acquisition system for high-resolution acquisition of testicular nuclear protein cancer tissue section images, and the microscopic display screen is used to display the testicular nuclear protein cancer tissue section images.
[0048] That is, four testicular nuclear protein cancer tissue section images of Figures 6-10 , that is, standard images of four periods of testicular tissue abnormal hyperplasia period, testicular tissue in situ cancer stage, testicular tissue infiltration period and testicular tissue metastasis period.
[0049] The standard images corresponding to the four periods need to be observed by relying on the scanning control equipment on the observation glass window 4 to perform scanning reading and writing.
[0050] 1. The scanning path of the scanning control adopts a snake algorithm to ensure the integrity of the section scanning. The system control software monitors the movement accuracy through a position feedback sensor, and the error is controlled within ±0.3 μm;
[0051] 2. After acquisition, each image block is integrated into a full section image through a splicing algorithm based on SIFT feature matching, and the image size can reach 100000×100000 pixels.
[0052] The image preprocessing includes the following steps:
[0053] 1. Denoising: using an adaptive non-local mean filtering algorithm to remove random noise;
[0054] 2. Color normalization: using Reinhard algorithm to unify the image color distribution to a standard reference section;
[0055] 3. White balance and brightness compensation: improving the consistency of image brightness through histogram matching;
[0056] 4. Block processing: dividing large-size images into 256×256 or 512×512 image blocks for model input.
[0057] The existing signal pattern recognition technology faces multiple deep challenges when applied to testicular nuclear protein cancer detection: first, there is a lack of pre-processing and feature extraction methods designed specifically for biomedical signal characteristics that can effectively suppress noise and accurately extract key biological information; second, in the fusion of multi-source heterogeneous biological signals, the deep biological correlation between different modal data cannot be effectively modeled and utilized; third, the generalization ability and robustness of existing pattern recognition models are significantly insufficient in complex biomedical environments, making it difficult to meet the stringent requirements of high precision and high reliability for early diagnosis of testicular nuclear protein cancer.
[0058] Therefore, the integrated original whole slice image is standardized and enhanced by the image preprocessing module. In order to suppress the noise in the image recognition process, an adaptive non-local mean filter algorithm based on local statistical characteristics is used.
[0059] That is, I'(x) =\frac{1}{C(x)}\sum_{y\in \Omega} e^{-\frac{||I(x)-I(y)||^2}{h^2}} I(y);
[0060] Where I'(x) is the filtered pixel, h is the smoothing parameter, and \Omega is the local neighborhood.
[0061] The adaptive Kalman filter combined with empirical mode decomposition, multivariate scatter correction, RMA algorithm and non-local mean filter can effectively suppress high noise and complex artifacts in biological signals while preserving key biological information, overcoming the problem of insufficient adaptability of existing pre-processing methods for biological signals. Secondly, the present application extracts multi-dimensional and multi-scale biological key features from various modal biological signals through multi-scale wavelet packet decomposition, mass spectrum peak identification and quantification, differential gene analysis, and radiomics and deep learning feature encoder methods, ensuring the comprehensiveness and fineness of the information.
[0062] Then color normalization, due to significant color difference between different slice staining batches, Reinhard algorithm is used to realize color space mapping, so that the input image is consistent with the reference slice statistical characteristics in Lab space:
[0063] L' = \frac{\sigma_{L_{ref}}}{\sigma_L}(L - \mu_L) + \mu_{L_{ref}};
[0064] a' = \frac{\sigma_{a_{ref}}}{\sigma_a}(a - \mu_a) + \mu_{a_{ref}}, \quad b' = -\frac{\sigma_{b_{ref}}}{\sigma_b}(b - \mu_b) + \mu_{b_{ref}};
[0065] Then, the system automatically detects the effective region according to the tissue distribution of the pathological section, and eliminates the background region. The Otsu threshold segmentation method is used to detect the tissue boundary, and then the 256*256 or 512*512 size is used for block resampling as the neural network input.
[0066] The hierarchical fusion strategy based on the knowledge graph-based graph neural network and the Transformer cross-modal attention mechanism is introduced, which realizes effective modeling and utilization of deep and nonlinear correlations between different biological modalities, which significantly surpasses the limitations of simple splicing or shallow correlation of existing technologies at the feature level, so as to mine key biological signal pathways specific to TNPC from complex biological networks; finally, the present application constructs a cascade classification model based on deep forest, and combines the Monte Carlo Dropout method to realize uncertainty quantification, and the SHAP algorithm for explainability analysis, which significantly improves the precision, robustness and clinical usability of TNPC diagnosis, so that the system can adapt to the inherent high complexity and individual variability of biomedical signals, and provides a new system-level solution for early, accurate and reliable diagnosis of TNPC.
[0067] The whole device includes hardware system and software system two parts:
[0068] 1、The hardware system includes multispectral microscopic scanner, image acquisition card, GPU computing unit, data storage server and display terminal;
[0069] 2、The software system includes image preprocessing program, AI recognition model, result report generation system and user interaction interface.
[0070] The whole device carries out data interaction through local area network or cloud server, can be independently deployed in hospital pathology department or laboratory, finally image enhancement, executes random rotation, mirror image, brightness disturbance and Gaussian blur enhancement operation to each sample, to improve the generalization ability of the model.
[0071] The artificial intelligence analysis module is the core part of the present application, which is responsible for extracting pathological features from image data and classifying and determining, and the artificial intelligence analysis module includes the following three units: convolution feature extraction unit, Transformer feature fusion unit and classification and determination unit.
[0072] The convolution feature extraction unit adopts an improved ResNet-50 convolutional neural network, the network structure contains 50 layers of convolutional layers, the gradient disappearance problem is solved by using residual connection, the input is an image block with a size of 512*512*3, and the output is a 2048-dimensional feature vector, and the model adopts the following optimization strategies:
[0073] 1. Convolution kernel size: 3*3, step 2;
[0074] 2. Activation function: ReLU6;
[0075] 3. Regularization: Batch Normalization + Dropout (0.3);
[0076] 4. Weight initialization: He initialization;
[0077] 5. Training optimizer: Adam (learning rate 1e-4)
[0078] Further, the Transformer feature fusion unit, in order to make up for the lack of global modeling ability of CNN, the application introduces a visual Transformer (ViT) in the feature fusion layer, the ViT divides the convolution output feature into a fixed-length patch sequence, and inputs into a multi-head self-attention mechanism:
[0079] Attention (Q, K, V) = Softmax(\frac{QK^T}{\sqrt{d_k}})V;
[0080] The mechanism captures the semantic association between long-distance regions and can identify the spatial dependence relationship of the lesion area. The number of Transformer layers is 6, and each layer contains 8 attention heads.
[0081] Finally, the classification and discrimination unit outputs three types of results after feature fusion, using a fully connected layer + Softmax classifier:
[0082] 1. C1: normal tissue;
[0083] 2. C2: suspicious lesion;
[0084] 3. Diagnosed cancer.
[0085] Therefore, the model output contains a probability distribution vector P = [p1, p2, p3], and the highest value corresponds to the final judgment category.
[0086] Further,
[0087] The system is improved in interpretability by the feature visualization module, the Grad-CAM algorithm is introduced in the output stage, and the heat visualization of the lesion area is realized.
[0088] The specific steps are as follows:
[0089] 1. Obtain the classification layer gradient and the feature map;
[0090] 2. Weighted average of the feature map:
[0091] L_{Grad-CAM}^{c} = ReLU\left(\sum_k -\alpha_k^c A^k\right);
[0092] Where A^k is the kth feature map, and \alpha_k^c is the weight coefficient.
[0093] 3. Superimpose the generated heat map on the original image to display the key area determined by AI.
[0094] The system can also perform morphological segmentation according to the heat intensity to extract the high expression area of nuclear protein and provide intuitive reference for doctors. Generally, Grad-CAM algorithm is used to generate heat map to display the AI attention area; combined with morphological operation to segment the high expression cell group of nuclear protein, the lesion distribution visualization image is generated.
[0095] Meanwhile, according to the analysis of the AI analysis module, the AI analysis module includes:
[0096] 1. Convolution feature extraction unit (CNN): ResNet50 structure is used to extract cell-level features;
[0097] 2. Transformer feature fusion unit (ViT): used for global feature modeling and cross-region dependency analysis;
[0098] 3. Classification and discrimination unit: Softmax layer is used to output three types of results: normal tissue, suspicious lesion and confirmed lesion.
[0099] The model is trained by testicular nuclear protein immunohistochemical section samples, using Adam optimizer and cross-entropy loss function.
[0100] Image preprocessing includes the following steps:
[0101] 1. Denoising: using adaptive non-local mean filtering algorithm to remove random noise;
[0102] 2. Color normalization: using Reinhard algorithm to unify the image color distribution to the standard reference section;
[0103] 3. White balance and brightness compensation: improve the consistency of image brightness through histogram matching;
[0104] 4. Block processing: divide large size images into 256x256 or 512x512 image blocks for model input.
[0105] According to the result output module, the result statistics, report generation and database interaction are realized
[0106] The output content includes:
[0107] 1. Classification result and confidence;
[0108] 2. Lesion area and proportion;
[0109] 3. Nuclear protein signal intensity distribution;
[0110] 4. Risk score (based on logistic regression model calculation);
[0111] 5. Lesion heat map and distribution map.
[0112] The device has an automatic learning function, which can fine-tune the model according to new case sample data. Transfer learning and data augmentation (rotation, mirroring, brightness disturbance) are used to improve the generalization performance.
[0113] The working principle of the application is as follows:
[0114] An observation port is formed in the operation table 1, and a slide placing base 6 is arranged on the observation port. The slide placing base 6 has a cover glass 7, so that the operator rotates and pushes the elastic pressing piece 9 on the fixed shaft 8, so that the elastic pressing piece 9 is pressed on the slide placing base 6 in the observation port. The slide placing base 6 has a cover glass 7, so that the elastic pressing piece 9 can be pressed on the slide placing base 6 with the cover glass 7. Further, the cover glass 7 and the slide placing base 6 have corresponding testicular nuclear protein cancer tissues, and then the mechanical hand assembly is turned over to invert the slide placing base 6 on the observation glass window 4. This can provide fixation and support for the slide placing base 6, and also provide conditions for the microscope analyzer 10 to observe and analyze;
[0115] The rotating base 12 rotates and moves up and down on the first rotating shaft 11 under the driving force of the external drive mechanism. The rotation can change the position of the operating base 15 between the operating table 1 and the observation glass window 4. The up and down movement of the rotating base 12 can drive the slide placement base 6 to move up and down, providing driving force for the vertical movement of the slide placement base 6. Turning on the drive motor can drive the second rotating shaft 17 to rotate 180°, so that the flipping table 16 carrying the slide placement base 6 rotates, that is, the originally upward-facing slide placement base 6 flips to face downward, realizing the orientation change of the slide placement base 6.
[0116] The image acquisition module is a multispectral microscopic image acquisition system used to acquire high-resolution images of testicular nucleocarcinoma tissue sections, with a microscopic display screen used to display the collected testicular nucleocarcinoma tissue section images.
[0117] That will appear Figures 6-10 Standard images of four types of testicular nucleoprotein carcinoma tissue sections: abnormal proliferative phase of testicular tissue, in situ carcinoma of the testis, invasive phase of testicular tissue, and metastatic phase of testicular tissue.
[0118] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A smart analysis device for pathological images of testicular nucleoprotein carcinoma, comprising an operating table (1) and a microscope analyzer (10) disposed on the operating table (1), characterized in that, An observation glass window (4) is provided below the operating table (1). An observation port is provided on the operating table (1), and a slide placement base (6) is provided on the observation port. A cover glass (7) is provided on the slide placement base (6). A flipping robot assembly for flipping the slide placement base (6) is provided below the operating table (1), and the flipping robot assembly flips the slide placement base (6) upside down on the observation glass window (4).
2. The intelligent analysis device for pathological images of testicular nucleoprotein carcinoma according to claim 1, characterized in that, The observation glass window (4) is provided with a partition (5) on the side facing the operating table (1), and the partition (5) divides the observation glass window (4) into multiple placement chambers, and multiple microscopic display screens are provided on the placement chambers.
3. The intelligent analysis device for pathological images of testicular nucleoprotein carcinoma according to claim 1, characterized in that, The observation glass window (4) is equipped with a placement frame (3), and the placement frame (3) is connected to the operating table (1) through a connecting rod (2).
4. The intelligent analysis device for pathological images of testicular nucleoprotein carcinoma according to claim 1, characterized in that, The operating table (1) is equipped with a fixed shaft (8), which is set in two pairs and symmetrically on both sides of the observation port. An elastic pressure plate (9) that presses the cover glass (7) is rotatably set on the fixed shaft (8).
5. The intelligent analysis device for pathological images of testicular nucleoprotein carcinoma according to claim 1, characterized in that, The bottom of the operating table (1) is also provided with a first rotating shaft (11), and a rotating base (12) is also provided on the first rotating shaft (11). The rotating base (12) moves on the first rotating shaft (11) through the through sliding hole (13).
6. The intelligent analysis device for pathological images of testicular nucleoprotein carcinoma according to claim 5, characterized in that, An operating platform (15) is provided on one side of the rotating base (12), and the operating platform (15) slides horizontally on the rotating base (12) through the limiting rod (14). A driving mechanism is provided inside the rotating base (12).
7. The intelligent analysis device for pathological images of testicular nucleoprotein carcinoma according to claim 6, characterized in that, The operating base (15) is provided with a flipping mechanism, and a second rotating shaft (17) rotates on the operating base (15), and the flipping mechanism drives the second rotating shaft (17) to rotate.
8. The intelligent analysis device for pathological images of testicular nucleoprotein carcinoma according to claim 7, characterized in that, The flipping mechanism includes a drive motor, the output end of which is coaxially fixed with a first drive gear, and the second rotating shaft (17) is coaxially fixed with a second drive gear.
9. The intelligent analysis device for pathological images of testicular nucleoprotein carcinoma according to claim 7, characterized in that, A tilting table (16) is provided on the second rotating shaft (17), and an electric suction cup (18) is provided on the tilting table (16).
10. The intelligent analysis device for pathological images of testicular nucleoprotein carcinoma according to claim 2, characterized in that, The observation glass window (4) is equipped with a clamping device corresponding to the placement cavity. The clamping device includes a clip.