An early screening system for cervical cancer based on cell morphology modeling

Abstract

This invention belongs to the field of medical image-assisted diagnosis technology, specifically a cervical cancer early screening system based on cell morphology modeling. It includes an image acquisition and standardized preprocessing unit, an adaptive boundary continuity construction module, a microscopic high-frequency deformation feature deconstruction module, a biophysical property dynamic modeling module, a multidimensional anomalousness evaluation engine, and an intelligent screening result feedback center. It actively evolves and generates closed contours through a physical potential field model, quantifies the high-frequency deformation features of the nuclear membrane using multi-scale curvature flow decomposition, and performs dynamic response analysis based on finite element analysis to invert physical rigidity. This application achieves modeling from apparent pixels to deep mechanisms, effectively distinguishing between physiological compression deformation and malignant morphological anomalousness, significantly improving the system's screening sensitivity and diagnostic robustness.

Description

Technical Field

[0001] This invention belongs to the field of medical image-assisted diagnosis technology, specifically a cervical cancer early screening system based on cell morphology modeling. Background Technology

[0002] Cervical cancer is a major malignant tumor threatening women's health. Early screening and accurate diagnosis are key to reducing mortality. Liquid-based thin-layer cytology is the core of initial screening. Computer-aided diagnostic systems use automated algorithms to quantify cell nuclear morphology parameters, replacing manual observation to improve screening throughput and reduce subjective bias. This technology has been widely adopted in clinical settings. Current mainstream computer-aided screening models simplify cervical cells into a static set of pixels, extract cell nuclear boundaries through edge detection and use macroscopic geometric descriptors to represent morphology. Although they have certain classification efficiency in high-quality samples, they have profound logical limitations and cannot meet the stringent requirements of sensitivity and stability for early screening. Existing technologies have low tolerance for non-ideal imaging quality, are prone to segmentation distortion due to uneven staining and noise interference, and cannot capture early signs of malignancy such as microscopic deformation of the nuclear membrane. At the same time, they ignore the biomechanical properties of the cell nucleus, making it difficult to distinguish between physiological deformation and pathological atypia, which can easily lead to missed diagnoses and misdiagnoses. This has become a key technical bottleneck restricting the accurate early screening of cervical cancer.

[0003] Therefore, the present invention provides an early cervical cancer screening system based on cell morphology modeling. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0005] The technical solution adopted by the present invention to solve its technical problem is: the cervical cancer early screening system based on cell morphology modeling of the present invention includes: an image acquisition and standardization preprocessing unit, used to capture the original cervical cell image and perform quality enhancement processing, and output a digital tensor with a standardized grayscale distribution; The adaptive boundary continuity building module is used to build a physical potential field model based on the principle of energy minimization, and to generate a closed outline of the cell nucleus with physical continuity by guiding the convergence of the evolution curve. The micro-high frequency deformation feature deconstruction module is used to perform multi-scale curvature flow decomposition on the closed contour of the cell nucleus to dissociate the micro-high frequency deformation components of the nuclear membrane boundary. The biophysical property dynamics modeling module is used to convert the closed contour of the cell nucleus into a virtual mesh model and obtain the biomechanical characteristics of the cell nucleus by executing elastic dynamics simulation operators; A multidimensional heterogeneity evaluation engine is used to construct a multi-branch feature fusion neural network to perform integrated evaluation on the microscopic high-frequency deformation components, the biomechanical features, and the geometric features of the cell nucleus closure contour, and output the lesion grading results. The intelligent screening result feedback center is used to receive the lesion grading results and generate a visual diagnostic atlas with pathological morphology annotations.

[0006] Preferably, the image acquisition and standardization preprocessing unit serves as the system's signal input, responsible for digitally capturing and enhancing the quality of the original cervical cell smear image. Addressing the common issues of uneven staining and nonlinear noise generated by the scanner in clinical samples, this unit executes an adaptive light field compensation algorithm. This algorithm constructs a dynamic gain function by calculating the local mean and variance distribution within the pixel spatial domain, nonlinearly stretching contrast-deficient regions. Simultaneously, it utilizes a nonlocal mean filtering operator to fully preserve the original gradient features of the cell nucleus edges while suppressing background noise. The preprocessed image signal is transformed into a digital tensor with a standardized grayscale distribution, providing a physically consistent data foundation for subsequent accurate segmentation.

[0007] Preferably, the adaptive boundary continuity construction module is the key to resolving the contradiction between edge breaks and topological discontinuities. This module no longer relies solely on local pixel gradient operators, but introduces a physical potential field model based on the principle of energy minimization. By constructing a gradient vector flow field in pixel space, the module transforms the process of identifying the cell nucleus boundary into an active shape evolution process. In the evolution logic, the system defines internal constraint energy and external potential energy function. The internal constraint energy is used to simulate the physical tension of the nuclear membrane, forcing the contour curve to maintain second-order geometric continuity. Thus, when encountering areas with extremely light staining or severe background interference, the curvature extrapolation logic is used to automatically complete the physical breakpoints. The external potential energy function guides the evolution curve to converge towards the boundary of the real anatomical structure. This modeling method ensures that the extracted cell nucleus contour has physical closure and smoothness, completely eliminating shape descriptor distortion caused by imaging quality defects.

[0008] Preferably, the microscopic high-frequency deformation feature deconstruction module solves the smoothing effect of traditional geometric features on early malignant signs by executing a multi-scale curvature flow decomposition operator. This module performs spectral decomposition on the closed contour output by the adaptive boundary continuity construction module, dividing it into macroscopic low-frequency geometric components and microscopic high-frequency deformation components. The macroscopic low-frequency geometric components are used to calculate conventional indicators such as area, perimeter, and aspect ratio; while the microscopic high-frequency deformation components focus on the nonlinear fluctuations of the nuclear membrane boundary within a very small spatial scale. By defining a high-frequency deformation operator, the system accurately captures and quantifies the serrated protrusions, fine wrinkles, and deep notches on the nuclear membrane surface. This hierarchical feature deconstruction mechanism ensures that weak physical signs in early cancer are not filtered out in the global averaging calculation, greatly improving the system's accuracy in capturing latent atypical lesions.

[0009] Preferably, the biophysical property dynamic modeling module introduces an elastic dynamic simulation operator based on finite element analysis, which realizes the quantitative evaluation of the biomechanical properties of the cell nucleus. The module transforms the static image contour into a virtual mesh model with physical mass, damping and elastic modulus, performs stress inversion on the curvature distribution of the nuclear membrane surface, and calculates the equivalent stiffness coefficient and surface tension distribution of the nuclear membrane in the current state. Preferably, in the simulation environment, the system applies a fictitious fluid shear force load to the model and monitors the dynamic response path of its morphological evolution. Healthy cell nuclei exhibit high elastic recovery ability in the simulation, while abnormal cell nuclei, due to the collapse of their internal matrix structure, exhibit significant structural brittleness characteristics in their physical model. This physical dimension modeling enables the system to distinguish between physiological compression deformation and pathological morphological abnormalities at the mechanistic level, providing definite physical support for the screening conclusions.

[0010] Preferably, the multidimensional heterogeneity evaluation engine executes a deep learning-based artificial intelligence classification algorithm, which integrates geometric features, micro-deformation features and physical property features by constructing a multi-branch feature fusion neural network. The neural network architecture of the engine includes an input layer, a multi-layer residual convolutional feature extraction layer, an attention mechanism layer, a feature fusion layer and a classification output layer. During model construction, the input layer receives a 4096-dimensional heterogeneous feature vector generated by the aforementioned modules. The high-frequency feature gradient of micro-deformation is transmitted through the deep residual connection mechanism to prevent information loss in deep mapping. The attention mechanism layer automatically learns and assigns higher weight coefficients to the micro-deformation components that reflect early cancer during training to ensure that the system is highly focused on weak signals.

[0011] Preferably, the training steps of the artificial intelligence model of the multidimensional heterogeneity evaluation engine include: first, constructing a database of tens of thousands of cervical cell images confirmed by the gold standard of pathology, and performing data balancing processing to ensure that the proportion of lesion samples at each level is scientific; The image is input into the system, and multi-dimensional physical semantic features are extracted by the feature deconstruction module and the physical property simulation module. Loss function optimization is performed by using the cross-entropy loss function in conjunction with the L2 regularization term, and the backpropagation algorithm is executed using the Adam optimizer to dynamically adjust the weight parameters of the neural network. During training, the system sets the initial learning rate to 0.0001 and executes a periodic learning rate decay strategy to ensure the global stability of the model during convergence.

[0012] Preferably, the intelligent screening result feedback center receives the classification probability distribution output by the multidimensional heterogeneity evaluation engine. The system classifies the judgment results into normal, low-grade lesion, high-grade lesion, and suspected malignancy levels through a preset risk threshold matrix. The feedback center can generate a visual diagnostic atlas, highlighting the microscopic deformation areas and abnormal physical stress areas that trigger the judgment logic on the original image. This traceable feedback mechanism eliminates the black box effect of artificial intelligence algorithms and provides pathologists with intuitive auxiliary diagnostic evidence.

[0013] The system operation logic of this invention ensures deep coupling and signal interaction between various modules. The clean image signal output by the standardized preprocessing unit flows to the boundary construction module. The generated continuous contour signal is input in parallel to the deformation deconstruction module and the physical modeling module. The deformation deconstruction module uses the Laplacian operator to strip out the high-frequency geometric features of the nuclear membrane, while the physical modeling module calculates the mechanical stability index of the cell nucleus through stiffness matrix calculation. These two sets of features with deep pathological semantics are finally fused in the evaluation engine and transformed into the final screening instructions through the nonlinear mapping of the neural network. This end-to-end process processing scheme makes each calculation step have clear pathological and physical significance.

[0014] The beneficial effects of this invention are as follows: 1. The cervical cancer early screening system based on cell morphology modeling described in this invention has an adaptive boundary continuity construction module that introduces physical potential fields and energy constraints to ensure that the cell nucleus outline always maintains the true topological morphology under uneven staining or background noise interference. This improvement in the accuracy of the underlying segmentation eliminates the risk of geometric descriptor distortion and provides a high-reliability geometric benchmark for subsequent quantitative analysis.

[0015] 2. The cervical cancer early screening system based on cell morphology modeling described in this invention uses a microscopic high-frequency deformation feature decomposition module to independently quantify the serrated protrusions and wrinkles that are smoothed out in traditional macroscopic parameters through spectral decomposition technology. This in-depth mining of microscopic pathological signals of the nuclear membrane significantly improves the system's capture rate of early in situ carcinoma and makes up for the sensory blind spots of existing systems in early screening.

[0016] 3. The cervical cancer early screening system based on cell morphology modeling described in this invention introduces the quantification of physical properties based on dynamic simulation, filling the gap in the biomechanical dimension of auxiliary diagnostic systems. The biophysical property dynamic modeling module realizes the quantitative evaluation of cell nuclear rigidity and structural stability by simulating elastic response. This enables the system to have the essential ability to distinguish between physiological compression deformation and pathological aberrations, greatly reducing the misdiagnosis rate under conditions of large-area cell overlap.

[0017] 4. The cervical cancer early screening system based on cell morphology modeling described in this invention achieves intelligent integration of heterogeneous features by integrating a multi-branch fusion neural network with an attention mechanism. The multidimensional heterogeneity evaluation engine can automatically learn the nonlinear correlation between geometry, texture, deformation and physical parameters, and weight key indicators. This mechanism- and data-driven modeling paradigm ensures the generalization ability and diagnostic consistency of the screening system under different pathology centers and different types of smears.

[0018] 5. The cervical cancer early screening system based on cell morphology modeling described in this invention greatly improves the transparency and clinical effectiveness of the screening process. The intelligent screening result feedback center generates a visual diagnostic atlas, transforming the complex algorithm judgment logic into intuitive pathological morphological annotations. This not only shortens the review time for pathologists, but also provides traceable and quantifiable morphological evidence for precision medicine. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a structural block diagram of an early cervical cancer screening system based on cell morphology modeling, as described in this invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0022] like Figure 1As shown in the embodiment of the present invention, a cervical cancer early screening system based on cell morphology modeling achieves full-dimensional modeling of cervical exfoliated cells from the apparent pixel distribution to the deep physiological mechanism by constructing an intelligent screening system that deeply couples physical continuity constraints and dynamic mechanical response analysis. In the overall architecture of the system of this invention, the image acquisition and standardization preprocessing unit serves as the physical starting point for signal input, and is tightly logically coupled with the adaptive boundary continuity construction module, the microscopic high-frequency deformation feature deconstruction module, the biophysical property dynamics modeling module, the multidimensional anisotropy evaluation engine, and the intelligent screening result feedback center through a high-speed data bus. During system operation, the raw image signals acquired by the photoelectric sensor are enhanced in energy and optimized in signal-to-noise ratio by the preprocessing unit; The boundary construction module extracts a physically meaningful closed contour, which is then input in parallel to the deformation deconstruction module and the dynamic modeling module to derive microscopic morphological features and biomechanical features, respectively. Finally, the anomaly evaluation engine performs artificial intelligence classification and judgment, and the feedback center outputs visual diagnostic evidence.

[0023] During the execution of the image acquisition and standardization preprocessing unit, the cervical cell smear is physically captured using a high-resolution scanning device. The light signal is converted into discrete level signals and encapsulated into a high-dimensional digital tensor. In response to the local brightness bias caused by uneven application of staining solution commonly found in clinical samples, as well as the photoelectric conversion thermal noise inherent in the optical imaging system, this unit executes an adaptive light field compensation algorithm. The control logic of the algorithm is as follows: the system traverses the entire pixel space domain through a sliding window and calculates the mean and variance distribution characteristics of the pixels in the current window in real time. Based on these local statistics, the system constructs a non-linear dynamic gain function. This function can perform differentiated grayscale mapping for shadow areas or overexposed areas with extremely low contrast, realize non-linear stretching of contrast-deficient areas, and significantly enhance the boundary between the cell nucleus and cytoplasm in the grayscale space. Furthermore, in order to eliminate random noise without damaging the microscopic texture of the cell nuclear membrane, the system calls the nonlocal mean filter operator; The mechanism of this operator is as follows: it not only refers to the neighborhood information of the current pixel, but also finds pixel blocks with similar topological structures in the global search window, assigns corresponding weight coefficients by calculating the Euclidean distance between the blocks, and performs weighted averaging. This processing method can effectively suppress background noise and completely preserve the original gradient features of the cell nucleus edges, ensuring that the preprocessed image signal has a high degree of physical consistency.

[0024] The adaptive boundary continuity construction module is logically constructed based on the physical potential field model of the energy minimization principle. After receiving the normalized image tensor, the module constructs a gradient vector flow field in the pixel space. This flow field radiates the gradient information of the cell nucleus edge from the local to the surrounding area by solving the partial differential diffusion equation, forming an attractive potential energy that can guide the convergence of the evolution curve. In this physical potential field model, the system defines a joint energy functional consisting of internal constraint energy and external potential energy function. The internal constraint energy is used to simulate the physical tension and bending stiffness of the nuclear membrane. It maintains second-order continuity geometrically by forcing the profile curve through first-order and second-order derivative terms. When the system encounters areas with very faint nuclear membrane defects or physical breakpoints caused by leukocyte obstruction during evolution, the curvature extrapolation logic driven by internal constraints automatically fills in the physical breakpoints based on the slope change trend of the existing boundaries, ensuring that the generated contours are closed in the topological structure. The external potential function uses the gradient vector of the image to guide the evolution curve to accurately lock onto the boundary of the real anatomical structure. This active shape evolution modeling method ensures that the extracted cell nuclear contours have clear physical meaning, completely eliminates the shape descriptor distortion caused by imaging quality defects, and provides a robust geometric topological benchmark for subsequent refined analysis.

[0025] The microscopic high-frequency deformation feature deconstruction module achieves deep deconstruction of microscopic pathological signals of cell nuclear membrane by executing multi-scale curvature flow decomposition operator. The system projects the closed contour output by the adaptive boundary continuity construction module onto the complex plane, regards it as a periodic spatial frequency signal, and uses the Laplace-Beltramian operator to perform spectral decomposition on the signal. During the decomposition process, the system divides it into macroscopic low-frequency geometric components and microscopic high-frequency deformation components. The macroscopic low-frequency geometric components represent the basic physical envelope of the cell nucleus and are used to calculate macroscopic morphological indicators such as area, perimeter, and major-minor axis ratio, characterizing the volume expansion and out-of-roundness of the cell nucleus during the malignant transformation process. The microscopic high-frequency deformation components are stripped of the interference of the basic shape and focus on the nonlinear fluctuation characteristics of the nuclear membrane boundary in a very small spatial scale. The system extracts and quantifies the frequency and amplitude of the serrated protrusions, fine wrinkles, and deep notches on the nuclear membrane surface by defining high-frequency deformation operators. The advantage of this hierarchical feature deconstruction mechanism is that it can capture early signs of malignancy that are easily overlooked in global averaging calculations. This sensitivity to microscopic pathological signals enables the system to effectively lock onto latent atypical cells in the early stages of disease.

[0026] The biophysical property dynamic modeling module is the key to the system’s mechanistic diagnosis. It introduces an elastic dynamic simulation operator based on finite element analysis to transform static visual features into dynamic mechanical parameters. This module performs Delaunay triangulation on closed contours and transforms them into virtual mesh models with physical mass, damping and elastic modulus. The system performs stress inversion on the curvature distribution of the nuclear membrane surface and calculates the equivalent stiffness coefficient and surface tension distribution of the nuclear membrane under the current physiological state using the generalized tensor form of Hooke's law. In the digital simulation environment, the system applies a preset fictitious fluid shear force load to the mesh model, which simulates the physical compression of cells during shedding and slide preparation. The system monitors the dynamic response path of the physical model's morphological evolution in real time. Healthy cell nuclei exhibit high elastic recovery ability after the simulated load is removed, and their displacement field shows a smooth linear distribution. In contrast, due to the high aggregation of internal chromatin and metabolic disorders of nucleoskeletal proteins, the physical model of defective cell nuclei exhibits significant structural fragility characteristics under pressure, namely, nonlinear local collapse or stress concentration. This physical dimension modeling enables the system to effectively distinguish between physiological compression deformation caused by slide preparation conditions and morphological aberration caused by pathological malignancy from the physiological mechanism level, greatly reducing the system's misdiagnosis rate under conditions of large-area cell overlap.

[0027] The multidimensional heterogeneity evaluation engine, as the decision-making center of the system, executes an artificial intelligence classification algorithm based on deep learning. Its core is a multi-branch feature fusion neural network. The neural network architecture is designed to achieve collaborative processing of heterogeneous features. The input layer receives a 4096-dimensional heterogeneity feature vector jointly generated by the preprocessing, boundary construction, deformation deconstruction, and physical modeling modules. These vectors are divided into geometric feature branches, micro-deformation energy spectrum branches, and biomechanical parameter branches according to their attributes. The feature extraction layer adopts a deep residual connection mechanism. By introducing an identity mapping function, it ensures that the high-frequency feature gradients of micro-deformation can be stably transmitted between extremely deep network layers, preventing information loss. In the middle of the neural network, the system embeds an attention mechanism layer. This layer automatically identifies the contribution of different features to cancer grading through a weight matrix generated by self-learning, and assigns higher weight coefficients to the micro-deformation components that reflect early cancer grading, ensuring that the diagnostic logic is highly focused on weak signals. The feature fusion layer concatenates the deep features of each branch into tensors and maps them to the Softmax classification output layer through a fully connected layer to give the final lesion grading probability distribution.

[0028] The training steps of the AI ​​model for the multidimensional heterogeneity evaluation engine are fundamental to ensuring the consistency of system diagnosis. The system has constructed a database of tens of thousands of cervical cell images confirmed by the gold standard of pathology, covering a complete pathological sequence from normal, LSIL, HSIL to malignant cancer. During the data preparation phase, data balancing is performed, and early lesion data with limited samples are expanded through rotation, scaling, and generative adversarial network techniques to ensure that the proportion of lesion samples at all levels in the training set conforms to scientific statistics. After the image is input into the system, the deformation deconstruction module and the physical property simulation module extract multi-dimensional physical semantic features, which are used as the input benchmark for training. The system optimizes the loss function by using the cross-entropy loss function to calculate the deviation between the predicted distribution and the gold standard distribution, and uses L2 regularization to suppress the excessive expansion of model weights. The system also uses the Adam optimizer to execute the backpropagation algorithm to dynamically adjust the weight parameters of the neural network. During training, the system sets the initial learning rate to 0.0001 and executes a periodic learning rate decay strategy. By monitoring the accuracy curve of the validation set, the system ensures the global stability of the model during convergence and avoids getting trapped in local optima.

[0029] The intelligent screening result feedback center receives the classification probability distribution output by the multidimensional heterogeneity evaluation engine, performs the visualization of the results and auxiliary diagnostic logic. The system first maps continuous probability values ​​to clinical grades such as normal, low-grade lesion, high-grade lesion and suspected malignancy through a preset risk threshold matrix. The core function of the feedback center is to generate a visual diagnostic atlas. This function uses the reverse gradient propagation technology to mark the key areas that trigger the judgment logic on the original microscopic image. Specifically, the system highlights the areas with the most significant microscopic deformation or the locations of nuclear membrane notches with abnormal physical stress inversion, generating a heat map overlay. This traceable feedback mechanism eliminates the black-box effect common in artificial intelligence algorithms, allowing pathologists to intuitively examine the abnormal details that the algorithm focuses on during review. This mechanistic auxiliary evidence not only improves the efficiency of image reading but also provides traceable morphological data support for precision medicine, greatly enhancing the certainty of clinical diagnosis.

[0030] The system operation logic of this invention ensures deep coupling and signal interaction between various modules. The normalized digital tensor output by the standardized preprocessing unit flows to the boundary construction module, and the generated continuous closed contour coordinates are used as spatial constraint information. These are then input in parallel to the deformation deconstruction module and the physical modeling module. The deformation deconstruction module uses the Laplacian operator to strip out the high-frequency geometric feature components of the nuclear membrane, while the physical modeling module calculates the mechanical stability index of the cell nucleus by constructing a stiffness matrix. These two sets of features with deep pathological semantics are finally subjected to feature-level tensor fusion in the multidimensional heterogeneity evaluation engine. Through the nonlinear feature mapping of the neural network, the system transforms complex physical quantities into final screening instructions. This end-to-end process processing scheme ensures that each calculation step has clear pathological and physical significance.

[0031] In a further embodiment, the system performs more refined control for image processing under complex backgrounds. When processing extremely high-resolution images, the adaptive light field compensation algorithm in the preprocessing unit adopts a block-parallel computing architecture and uses GPU-accelerated kernel functions to realize real-time brightness reconstruction of millions of pixels. The non-local mean filtering filters irrelevant pixel blocks through a preset similarity threshold, ensuring the denoising strength while keeping the computational complexity within a reasonable range. During the evolution process, the boundary construction module introduces dynamic time step control logic. When the contour tends to the edge of high gradient, the evolution step size is automatically reduced to improve the positioning accuracy, while the step size is increased in the uniform background area to improve the computational efficiency.

[0032] When performing curvature flow decomposition, the micro-high frequency deformation feature deconstruction module adopts a multi-scale analysis strategy to extract wave dynamics distributions at ten different frequency bandwidths. These components are combined into a morphological feature fingerprint, which can identify the heterogeneous evolution of specific subtypes. The biophysical property dynamics modeling module further introduces a nonlinear coupling term for the nucleus-mass ratio; In the computer simulation, the physical damping effect of the cytoplasm is taken into account, making the inversion result of the elastic response closer to the biological reality of living cells. Through these in-depth technical optimizations, this invention not only solves the feature loss problem in theory, but also improves the robustness in engineering implementation.

[0033] The embodiments employ the complete system described in this invention, which includes physical continuity boundary construction and dynamic simulation.

[0034] Comparative Example 1 uses a traditional machine learning scheme based on edge detection operators and simple geometric descriptors (roundness, area), while Comparative Example 2 uses a general convolutional neural network (CNN) to directly classify the original image without introducing physical modeling and deformation deconstruction modules.

[0035] Table comparing experimental data of the examples and comparative examples:

[0036] As can be seen from the data analysis in the table above, in terms of sensitivity for early cancer (CIN I / II stage), the present invention achieved 97.45%, far exceeding the 71.20% of Comparative Example 1 and 84.50% of Comparative Example 2. This significant improvement is attributed to the efficient capture of tiny serrated protrusions and notches of the nuclear membrane by the microscopic high-frequency deformation feature deconstruction module. Comparative Example 1, due to the use of a global geometric average descriptor, caused these microscopic aberration features to be treated as noise smoothing, resulting in missed diagnoses. Although Comparative Example 2 has a certain feature extraction capability, due to the lack of explicit constraints on physical continuity boundaries, it is very easy to extract incorrect edge features when the image quality is poor, resulting in limited diagnostic accuracy.

[0037] Under complex occlusion conditions, the misdiagnosis rate of this invention is only 1.12%, while that of Comparative Example 1 is as high as 15.45%. This strongly proves the effectiveness of the biophysical property dynamic modeling module. By simulating mechanical elastic response, this invention can accurately distinguish between physiological geometric deformation and pathological isomerism caused by cell overlap, and solve the long-standing problem of overlapping cell identification in the field of auxiliary diagnosis. It performs well in terms of edge detection integrity and early micro-trace capture rate, demonstrating the superiority of the physical potential field model and energy minimization evolution logic; In terms of generalization accuracy, this invention, through the combination of mechanistic model and data-driven approach, demonstrates stronger cross-center adaptability compared to the purely data-driven Comparative Example 2, proving that biomechanical parameters have stronger species universality when used as diagnostic criteria.

[0038] Furthermore, in the execution details of the multidimensional heterogeneity evaluation engine, the weight distribution learned by the attention mechanism layer shows that for early lesion samples, the weight of the micro-deformation component is significantly higher than that of the conventional geometric features. During operation, the system adaptively adjusts the sensitive frequency band of the neural network to achieve targeted identification of latent lesions. The visualized heat map generated by the intelligent screening result feedback center has a high degree of spatial overlap with the hand-drawn annotation area of ​​the pathologist in clinical comparison tests, and its average intersection-over-union ratio (IoU) reaches the preset high reliability range.

[0039] The modules of this system adopt a standardized data exchange protocol. The normalized tensor output by the standardized preprocessing unit is checked and verified before entering the boundary construction module to ensure that no bit flips occur during data transmission. The residual block structure design inside the evaluation engine enables the system to dynamically reduce the computational load while ensuring accuracy when performing large-scale concurrent screening tasks through model pruning technology. This flexibility allows the system to adapt to various hardware environments, from large servers in central laboratories to portable terminals in community screening centers.

[0040] In the identification logic targeting specific pathological features, the microscopic high-frequency deformation feature deconstruction module can identify gray-level gradient anomalies caused by nuclear membrane thickening and chromatin margination. By coupling local binary pattern features based on curvature decomposition, the system realizes second-order quantization of the fine texture of the nuclear membrane. The physical modeling module simulates the morphological collapse threshold of abnormal cell nuclei under high-speed centrifugal force by introducing a nonlinear elastic constitutive equation. This dimensional physical simulation provides additional dynamic criteria for determining the invasion risk of carcinoma in situ, extending the application scope of this invention from simple screening to prognostic risk assessment.

[0041] The intelligent screening result feedback center also has the ability to learn autonomously and iterate its knowledge base. By receiving feedback data from pathologists who review it online, the system can automatically add misdiagnosed or missed samples to the training set, triggering the incremental learning logic of the multidimensional heterogeneity evaluation engine. Utilizing online learning technology, the system can continuously correct parameter deviations in physical property modeling during continuous operation, achieving self-evolution of diagnostic performance. This closed-loop optimization path ensures that the system of this invention always maintains a leading performance level in the industry.

[0042] In summary, by integrating core technical features such as adaptive boundary continuity construction, microscopic high-frequency deformation deconstruction, and biophysical dynamics modeling, this invention not only solves the problems of image fragility, loss of microscopic features, and lack of mechanistic support in existing technologies, but also demonstrates its progress in improving early detection rate and reducing misdiagnosis rate through specific experimental data.

[0043] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A cervical cancer early screening system based on cell morphology modeling, characterized in that, include: The image acquisition and normalization preprocessing unit is used to capture raw cervical cell images and perform quality enhancement processing, outputting a digital tensor with a normalized grayscale distribution. The adaptive boundary continuity building module is used to build a physical potential field model based on the principle of energy minimization, and to generate a closed outline of the cell nucleus with physical continuity by guiding the convergence of the evolution curve. The micro-high frequency deformation feature deconstruction module is used to perform multi-scale curvature flow decomposition on the closed contour of the cell nucleus to dissociate the micro-high frequency deformation components of the nuclear membrane boundary. The biophysical property dynamics modeling module is used to convert the closed contour of the cell nucleus into a virtual mesh model and obtain the biomechanical characteristics of the cell nucleus by executing elastic dynamics simulation operators; A multidimensional heterogeneity evaluation engine is used to construct a multi-branch feature fusion neural network to perform integrated evaluation on the microscopic high-frequency deformation components, the biomechanical features, and the geometric features of the cell nucleus closure contour, and output the lesion grading results. The intelligent screening result feedback center is used to receive the lesion grading results and generate a visual diagnostic atlas with pathological morphology annotations.

2. The cervical cancer early screening system based on cell morphology modeling according to claim 1, characterized in that, The logic for quality enhancement processing performed by the image acquisition and normalization preprocessing unit includes: The image acquisition and normalization preprocessing unit uses a sliding window to traverse the pixel spatial domain, calculates the pixel mean and variance distribution characteristics in the current window in real time, and constructs a nonlinear dynamic gain function based on the pixel mean and variance distribution characteristics. The dynamic gain function is used to perform differentiated grayscale mapping on contrast-deficient regions to achieve adaptive compensation of the image light field. The nonlocal mean filter operator is invoked to find a reference block with a similar topological structure to the target pixel block within the global search window. The corresponding weight coefficients are assigned by calculating the Euclidean distance between the blocks and weighted averaging is performed to suppress background noise while preserving the original gradient features of the cell nucleus edges.

3. The cervical cancer early screening system based on cell morphology modeling according to claim 2, characterized in that, The logic for constructing the physical potential field model by the adaptive boundary continuity construction module includes: The adaptive boundary continuity construction module constructs a gradient vector flow field in the pixel space, and radiates the gradient information of the cell nucleus edge from the local to the surrounding area by solving the partial differential diffusion equation, forming an attractive potential energy. Define a joint energy functional consisting of an internal constraint energy and an external potential energy function, wherein the internal constraint energy is used to simulate the physical tension and bending stiffness of the nuclear membrane, and the evolution curve is constrained by first-order and second-order derivative terms to maintain second-order geometric continuity. The external potential energy function uses the gradient vector flow field to guide the evolution curve to converge toward the boundary of the real anatomical structure; When the evolution curve encounters a region with a missing nuclear membrane or a physical breakpoint, the curvature extrapolation logic driven by the internal constraint energy is used to automatically complete the physical breakpoint, ensuring that the generated closed outline of the cell nucleus is closed in the topological structure.

4. The cervical cancer early screening system based on cell morphology modeling according to claim 1, characterized in that, The logic for performing multi-scale curvature flow decomposition in the microscopic high-frequency deformation feature decomposition module includes: The microscopic high-frequency deformation feature deconstruction module projects the closed contour of the cell nucleus onto a complex plane, constructing it into a periodic spatial frequency signal; The spatial frequency signal is decomposed into macroscopic low-frequency geometric components and microscopic high-frequency deformation components using the Laplace-Beltrami operator. The area, perimeter, and major-minor axis ratio are calculated using the macroscopic low-frequency geometric components to characterize the volume expansion and out-of-roundness of the cell nucleus; Using a preset high-frequency deformation operator, the serrated protrusion features, wrinkle features, and deep notch features of the nuclear membrane surface are extracted and quantified from the microscopic high-frequency deformation components to obtain the nonlinear fluctuation parameters of the nuclear membrane boundary within a small spatial scale.

5. The cervical cancer early screening system based on cell morphology modeling according to claim 1, characterized in that, The logic for obtaining biomechanical characteristics by the biophysical property dynamics modeling module includes: The biophysical property dynamics modeling module triangulates the closed contour of the cell nucleus, transforming it into a virtual mesh model with physical mass, damping and elastic modulus. Stress inversion is performed on the curvature distribution of the virtual mesh model surface, and the equivalent stiffness coefficient and surface tension distribution of the nuclear membrane under the current physiological state are calculated using the generalized Hooke's law tensor. In a digital simulation environment, a preset fictitious fluid shear force load is applied to the virtual mesh model to simulate the physical compression of cells during the slide preparation process; By monitoring the dynamic response path of the morphological evolution of the virtual mesh model, the linear elastic recovery characteristics corresponding to healthy cell nuclei and the nonlinear local collapse or stress concentration characteristics corresponding to unhealthy cell nuclei are identified, thereby distinguishing between physiological compression deformation and pathological morphological abnormalities.

6. The cervical cancer early screening system based on cell morphology modeling according to claim 1, characterized in that, The architecture of the multi-branch feature fusion neural network in the multi-dimensional heterogeneity evaluation engine includes: The input layer is used to receive a 4096-dimensional anisotropic feature vector derived from deformation deconstruction, geometric extraction and dynamic modeling. The anisotropic feature vector is divided into geometric feature branch, micro-deformation energy spectrum branch and biomechanical parameter branch according to attributes. The feature extraction layer employs a deep residual connection mechanism, which introduces an identity mapping function to ensure the stable transmission of the feature gradients of the microscopic high-frequency deformation components between network layers. The attention mechanism layer is used to identify the contribution of different features to cancer grading through a weight matrix generated by self-learning, and to assign a preset weight coefficient to the microscopic high-frequency deformation component reflecting early cancer, so as to focus on weak anomalous signals. The feature fusion layer is used to concatenate the deep features of each branch into tensors and map them to the Softmax classification output layer through a fully connected layer to output the grading probability distribution corresponding to the lesion grading result.

7. The cervical cancer early screening system based on cell morphology modeling according to claim 6, characterized in that, The training logic of the artificial intelligence model of the multidimensional heterogeneity evaluation engine includes: A database of cervical cell images confirmed by pathological gold standards was constructed, and data balancing was performed. Generative adversarial network technology was used to expand the sample size of early lesion data that was below a preset threshold. The sample image is input into the system, and the microscopic high-frequency deformation feature deconstruction module and the biophysical property dynamics modeling module extract multidimensional physical semantic features as training input benchmarks. The cross-entropy loss function is used to calculate the deviation between the predicted distribution and the gold standard distribution, and L2 regularization is used to suppress model weight inflation. The Adam optimizer is used to dynamically adjust the weights of the neural network using the backpropagation algorithm. The initial learning rate is set to 0.0001 and periodic learning rate decay is performed. The model convergence is guided by monitoring the accuracy curve of the validation set.

8. The cervical cancer early screening system based on cell morphology modeling according to claim 1, characterized in that, The logic for generating a visual diagnostic atlas by the intelligent screening result feedback center includes: The intelligent screening result feedback center uses reverse gradient propagation technology to locate the key area that triggers the judgment logic on the original microscopic image; The regions with significant microscopic high-frequency deformation components or the locations of physical stress anomalies determined by the biophysical property dynamics modeling module are highlighted and marked to generate a heat map overlay layer. The heatmap overlay is spatially aligned and superimposed with the original microscopic image to output visualized diagnostic evidence, thereby eliminating the black-box effect of artificial intelligence algorithms and providing traceable morphological data.

9. The cervical cancer early screening system based on cell morphology modeling according to claim 1, characterized in that, The screening system internally executes a full-link signal interaction process, specifically including: The normalized digital tensor output by the image acquisition and normalization preprocessing unit flows to the adaptive boundary continuity construction module. The continuous closed contour coordinates generated by the adaptive boundary continuity construction module are used as spatial constraint information and are input in parallel to the microscopic high-frequency deformation feature deconstruction module and the biophysical property dynamics modeling module. The microscopic high-frequency deformation feature deconstruction module uses the Laplacian operator to strip out the high-frequency geometric feature components of the nuclear membrane, and the biophysical property dynamic modeling module calculates the mechanical stability index of the cell nucleus by constructing a stiffness matrix. The high-frequency geometric feature components and the mechanical stability index undergo feature-level tensor fusion in the multidimensional anisotropy evaluation engine, and are transformed into the final screening instructions through the nonlinear feature mapping of the neural network.

10. A cervical cancer early screening system based on cell morphology modeling according to claim 9, characterized in that, The screening system performs fine-grained control logic for complex background images, specifically including: The adaptive light field compensation algorithm in the standardized preprocessing unit adopts a block-parallel computing architecture and uses GPU-accelerated kernel functions to realize real-time reconstruction of pixel brightness. The adaptive boundary continuity construction module introduces dynamic time step control when executing evolution logic. When the evolution curve tends to the edge of high gradient, the evolution step size is automatically reduced, while the evolution step size is increased in the uniform background region to optimize computational efficiency. The biophysical property dynamics modeling module introduces a nonlinear coupling term of nucleocytoplasmic ratio during computer simulation, incorporating the physical damping effect of the cytoplasm into the inversion calculation of elastic response, thereby improving the biological fit of the simulated load test.

Images