Skin aging diagnosis system based on physical perception diffusion and grade determination method

Abstract

The application discloses a skin aging diagnosis system and grade determination method based on physical perception diffusion, relates to the field of biomedical imaging and artificial intelligence technology assisted diagnosis, and comprises a data acquisition and preprocessing module, a training data set construction module, a three-dimensional diagnosis module and a report generation and output module.The data acquisition and preprocessing module is used for receiving continuous two-dimensional skin tissue slice images of a target object, and performing standardization processing on the continuous two-dimensional skin tissue slice images.The training data set construction module is used for mapping and stacking two-dimensional labeling information by using a space alignment and dimensionality increasing algorithm based on the labeling of collagen fiber age state, constructing a three-dimensional collagen structure true value label body with spatial continuity, forming a gold standard data set to train a physically constrained end-to-end deep learning model.The application can significantly improve the accuracy and interpretability of skin aging diagnosis.

Description

Technical Field

[0001] This application relates to the field of biomedical imaging and artificial intelligence-assisted diagnosis, specifically to a skin aging diagnostic system and grade determination method based on physical perception diffusion. Background Technology

[0002] During skin aging, the collagen fiber network gradually degenerates and thins out, a slow and progressive process of destruction; while in cases of trauma such as injury or surgery, collagen fibers undergo acute and severe rupture. Although their causes differ, the core pathological changes in both can be attributed to an imbalance between the destruction and remodeling of the collagen microstructure.

[0003] Clinical diagnosis of collagen damage in skin tissue typically involves physical examination, skin biopsy, and imaging studies. Physicians confirm the diagnosis by examining changes in the hardness, color, and elasticity of the affected skin, combined with microscopic observation of collagen histological changes during skin biopsy. Current clinical methods still primarily rely on manual interpretation of two-dimensional slide images, which suffers from strong subjectivity and a lack of spatial awareness. These methods cannot directly provide specific information about collagen, nor can they reflect its three-dimensional spatial arrangement, resulting in superficial diagnostic information. Obtaining a true three-dimensional microstructure usually requires expensive confocal microscopy or two-photon microscopy for tomography, accompanied by complex and costly immunofluorescence staining procedures. Although some studies have attempted to use deep learning to analyze macroscopic skin images, they have not reached the pathological gold standard of tissue microstructure, lacking depth and interpretability, making it difficult to guide subsequent biomaterial preparation and 3D-printed skin scaffold development. Therefore, how to reconstruct the three-dimensional collagen network structure using low-cost, conventionally stained slides as input, and achieve personalized and quantifiable skin aging, has become an urgent problem to be solved. Summary of the Invention

[0004] This application provides a skin aging diagnostic system and grade determination method based on physical sensing diffusion, which can significantly improve the accuracy and interpretability of skin aging diagnosis.

[0005] In a first aspect, embodiments of this application provide a skin aging diagnostic system based on physical sensing diffusion, the skin aging diagnostic system based on physical sensing diffusion includes:

[0006] The data acquisition and preprocessing module is used to receive continuous two-dimensional skin tissue slice images of the target object, and to perform standardization processing on the continuous two-dimensional skin tissue slice images based on image processing algorithms. The continuous two-dimensional skin tissue slice images include at least two slices of different depths.

[0007] The training dataset construction module is used to map and stack two-dimensional annotation information based on the labeling of collagen fiber age status, using spatial alignment and dimensionality increase algorithms to construct a three-dimensional collagen structure truth label body with spatial continuity, forming a gold standard dataset to train a physically constrained end-to-end deep learning model.

[0008] The 3D diagnostic module is used to take standardized continuous two-dimensional skin tissue slice images as input based on a pre-trained end-to-end deep learning model with physical constraints. It performs cross-modal reconstruction and semantic segmentation of collagen fiber structures in the sequence images in 3D space, and generates and outputs a 3D collagen structure map and corresponding physical parameters.

[0009] The report generation and output module is used to analyze the output of the three-dimensional diagnostic module and output a quantitative report on skin aging level to guide clinical diagnosis and biomaterial preparation.

[0010] In conjunction with the first aspect, in one implementation method,

[0011] The continuous two-dimensional skin tissue slice images are multi-channel imaging data acquired by confocal microscopy. The multi-channel imaging data consists of two modal images synchronously acquired from the same tissue slice under the same field of view.

[0012] The modal images include confocal transmission light microscopy and multichannel fluorescence confocal microscopy.

[0013] The confocal transmission light microscopy is a bright-field image of a tissue section that has undergone standard histological staining, displayed under the transmission light detector channel.

[0014] The multichannel fluorescence confocal microscopy imaging consists of fluorescence images labeled with DAPI and FITC or Cy3 dyes, respectively, showing the spatial distribution of cell nuclei and type I collagen.

[0015] In conjunction with the first aspect, in one implementation method,

[0016] The image processing algorithms include Reinhard color migration algorithm, scale-invariant feature transform matching algorithm, and Otsu threshold segmentation algorithm.

[0017] The standardization process includes geometric correction, color standardization, and artifact removal.

[0018] In conjunction with the first aspect, in one implementation method,

[0019] The training dataset construction module also includes an image registration strategy for structure preservation.

[0020] The image registration strategy involves using a scale-invariant feature transformation algorithm to detect key feature points and estimate the initial affine transformation matrix to correct local nonlinear displacements.

[0021] In conjunction with the first aspect, in one implementation method,

[0022] The physical constraint end-to-end deep learning model includes a cross-modal structure diffusion module, a differentiable physical quantization module, and a microstructure diagnostic Transformer.

[0023] The cross-modal structure diffusion module is used to receive continuous two-dimensional skin tissue slice images as conditional guidance, and uses a feature injection mechanism to infer and reconstruct three-dimensional virtual fluorescent voxel data through a reverse diffusion process.

[0024] The differentiable physical quantization module is connected after the cross-modal structure diffusion module. The differentiable physical quantization module is equipped with a set of differentiable tensor operators for real-time calculation of the generated three-dimensional virtual fluorescent voxel data, and extraction of physical parameters including collagen volume fraction, protein fiber anisotropy index and curvature.

[0025] The microstructure diagnostic Transformer is used to simultaneously receive three-dimensional virtual fluorescent voxel data generated by the cross-modal structure diffusion module and physical parameters output by the differentiable physics quantization module. It uses a multimodal fusion architecture to analyze the topology of the collagen network and outputs the level characterizing the skin aging state.

[0026] In conjunction with the first aspect, in one implementation method,

[0027] The cross-modal structure diffusion module is a conditional feature injection type 3D U-Net model, and the backbone network of the conditional feature injection type 3D U-Net model contains symmetrical 3D encoder and 3D decoder paths.

[0028] The backbone network integrates a feature injection layer at each resolution level of the 3D encoder and 3D decoder.

[0029] The feature injection layer is used to receive feature maps of continuous two-dimensional skin tissue slice images, broadcast and copy them in the depth dimension to match the size of the three-dimensional feature map, and fuse two-dimensional texture information into three-dimensional spatial features through convolution operation, so as to recover the three-dimensional collagen fluorescence distribution that is aligned with the two-dimensional slice texture in terms of anatomical structure from Gaussian random noise.

[0030] In conjunction with the first aspect, in one implementation method,

[0031] The differentiable physical quantization module is constructed as a differentiable computational layer and integrates tensor operators for physical parameter extraction.

[0032] The tensor operators include anisotropic extraction operators and density-morphological operators, and both anisotropic extraction operators and density-morphological operators support automatic differentiation.

[0033] In conjunction with the first aspect, in one implementation method,

[0034] The anisotropic extraction operator is used to calculate the gradient field of voxels using a three-dimensional Sobel convolution kernel, construct a local structure tensor and solve for eigenvalues ​​to quantify the directional consistency of collagen fibers.

[0035] The density morphology operator is used to calculate the collagen volume fraction using a differentiable sigmoid approximation threshold function, and to calculate the crimp index of collagen fibers using a skeletal approximation algorithm.

[0036] In conjunction with the first aspect, in one implementation method,

[0037] The microstructure diagnostic Transformer is a hybrid feature encoding network with multimodal input. The input end is used to simultaneously receive three-dimensional virtual fluorescent voxel data generated by the cross-modal structure diffusion module and physical parameters output by the differentiable physics quantization module.

[0038] The hybrid feature encoding network consists of multiple stacked hybrid stages, and each hybrid stage is composed of stacked local texture encoding blocks and global dependency building blocks;

[0039] The local texture coding block is based on a three-dimensional convolutional neural network and is used to capture and extract the microscopic fracture and cross-linking details of collagen fibers.

[0040] The global dependency modeling module is used to model the topology and spatial arrangement of collagen networks based on the Transformer self-attention mechanism.

[0041] Secondly, embodiments of this application provide a method for determining skin aging levels, based on the aforementioned skin aging diagnostic system, wherein the method for determining skin aging levels includes:

[0042] Receive continuous two-dimensional skin tissue slice images of the target object and input them into the skin aging diagnosis system;

[0043] Guided by received continuous two-dimensional skin tissue slice images, three-dimensional virtual fluorescent voxel data characterizing the specific distribution of collagen are inferred and generated;

[0044] Real-time differentiable calculations were performed on the generated three-dimensional virtual fluorescent voxel data to extract physical parameters including collagen volume fraction, protein fiber anisotropy index, and curvature.

[0045] By integrating three-dimensional virtual fluorescent voxel data and extracted physical parameters, a level characterizing the skin's aging state is output;

[0046] Based on the physical parameters and the output levels characterizing skin aging, a quantitative report is generated to guide clinical diagnosis and biomaterial preparation.

[0047] The beneficial effects of the technical solutions provided in this application include:

[0048] By efficiently converting low-cost two-dimensional image input into objective and quantifiable three-dimensional structural output, and reconstructing the three-dimensional microstructure of collagen through cross-modal generation technology, combined with differentiable physical quantification methods, this method enables high-precision and automated quantitative assessment of skin collagen status and aging degree, and assists in clinical decision-making. This significantly improves the accuracy and interpretability of skin aging diagnosis, and provides precise and quantifiable three-dimensional structural evidence for the clinical diagnosis, targeted repair treatment, and prognostic assessment of skin collagen status. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the skin aging diagnostic system based on physical sensing diffusion according to this application;

[0050] Figure 2 This is a schematic diagram of the overall workflow of the skin aging diagnostic system of this application;

[0051] Figure 3 This is a schematic diagram of the structure of a physically constrained end-to-end deep learning model.

[0052] Figure 4 A schematic diagram of the network architecture of the cross-modal structure diffusion module;

[0053] Figure 5 This is a schematic diagram of the architecture of the differentiable physics quantization module;

[0054] Figure 6 This is a schematic diagram of the architecture of the microstructure diagnostic module;

[0055] Figure 7 Visual comparison verification diagram of three-dimensional virtual fluorescent collagen network image generated for skin aging diagnostic system and real confocal microscopic fluorescence image;

[0056] Figure 8 Examples of visualization output for aging levels and interpretable heatmap visualization based on the Grad-CAM algorithm;

[0057] Figure 9 A schematic diagram illustrating the correlation between DPQ output physical parameters and predicted age;

[0058] Figure 10 Example of quantitative indicator output;

[0059] Figure 11 This is a schematic diagram illustrating the statistical distribution and dimensionality reduction clustering analysis of collagen microstructure characteristics. Detailed Implementation

[0060] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0062] Firstly, embodiments of this application provide a skin aging diagnostic system based on physically sensed diffusion, addressing the dual shortcomings of current systems that rely on physical 3D imaging or complex tissue staining, resulting in cumbersome and costly sample acquisition processes, as well as highly subjective diagnostic results lacking 3D quantitative information. To achieve the above objectives, this application provides a skin aging diagnostic system based on a physically sensed cross-modal diffusion model. Through end-to-end computational dimensionality enhancement and physical constraint generation, it achieves a direct mapping from two-dimensional bright-field information to three-dimensional fluorescence structures and physical parameters, significantly improving the accuracy and interpretability of skin aging diagnosis. This provides precise and quantifiable 3D structural evidence for the clinical diagnosis, targeted repair treatment, and prognostic assessment of skin collagen status.

[0063] In one embodiment, reference is made to Figure 1 , Figure 1 This is a schematic diagram of the skin aging diagnostic system based on physical sensing diffusion, as described in this application. Figure 1 As shown, the skin aging diagnostic system based on physical perception diffusion includes: a data acquisition and preprocessing module, a training dataset construction module, a three-dimensional diagnostic module, and a report generation and output module. The skin aging diagnostic system of this application can perform the following: reconstruct and quantitatively analyze the three-dimensional collagen network structure of a target tissue region from a series of digitized continuous two-dimensional skin tissue slice images, and generate a report on the skin aging level accordingly.

[0064] The data acquisition and preprocessing module receives continuous two-dimensional skin tissue slice images (biological samples) of the target object and performs standardization processing on the continuous two-dimensional skin tissue slice images based on image processing algorithms. The continuous two-dimensional skin tissue slice images include at least two slices at different depths. The image processing algorithms include the Reinhard color transfer algorithm, the scale-invariant feature transform matching algorithm, and the Otsu threshold segmentation algorithm; the standardization processing includes geometric correction, color standardization, and artifact removal.

[0065] It should be noted that the source of biological samples covers multiple age groups (<20 years old, 20-30 years old, 30-40 years old, 40-50 years old, >50 years old), and is skin tissue that has been legally removed and discarded in clinical surgical procedures such as double eyelid surgery, facelift, and abdominoplasty, and the skin tissue covers at least one anatomical site among the eyelids, cheeks, and abdomen.

[0066] The training dataset construction module is used to map and stack two-dimensional annotation information based on the labeled age state of collagen fibers, using spatial alignment and dimensionality increase algorithms to construct a three-dimensional collagen structure ground truth (GT) label body with spatial continuity, forming the gold standard dataset for training the physically constrained end-to-end deep learning model. In other words, the training dataset construction module is used to construct the annotation of the training set and train the physically constrained end-to-end deep learning model.

[0067] The 3D diagnostic module is used to perform cross-modal reconstruction and semantic segmentation of collagen fiber structures in the sequence of images in 3D space based on a pre-trained end-to-end deep learning model with physical constraints. It takes standardized continuous 2D skin tissue slice images as input and generates and outputs a 3D collagen structure map and corresponding physical parameters.

[0068] The report generation and output module analyzes the output of the three-dimensional diagnostic module, outputting a skin aging grade and a quantitative report to guide clinical diagnosis and biomaterial preparation. Specifically, by analyzing the output of the three-dimensional diagnostic module and based on multiple three-dimensional structural parameters such as collagen spatial distribution density, fiber network anisotropy, average curvature, and breakpoint density, it outputs a clear skin aging grade (Grade IV) and a quantitative report to guide clinical diagnosis and biomaterial preparation.

[0069] In this application, the continuous two-dimensional skin tissue slice images are multi-channel imaging data acquired by confocal microscopy. The multi-channel imaging data consists of two modal images synchronously acquired from the same tissue slice under the same field of view. The modal images include confocal transmission light microscopy and multi-channel fluorescence confocal microscopy.

[0070] Confocal transmission light microscopy is a bright-field image of a tissue section stained with standard histology under the transmission light detector channel; multichannel fluorescence confocal microscopy is a fluorescence image showing the spatial distribution of cell nuclei and type I collagen, labeled with DAPI (a fluorescent dye that can strongly bind to DNA) and FITC (fluorescein isothiocyanate) or Cy3 dye, respectively.

[0071] In this application, the training dataset construction module also includes an image registration strategy for structure preservation. This strategy utilizes a scale-invariant feature transform algorithm to detect key feature points and estimates the initial affine transformation matrix to correct local nonlinear displacements. Specifically, the structure-preserving image registration strategy employed in the training dataset construction module includes using a scale-invariant feature transform algorithm to detect key feature points and estimating the initial affine transformation matrix to correct local nonlinear displacements.

[0072] In this application, the physically constrained end-to-end deep learning model includes a Cross-Modal Structural Diffusion Module (CMSD), a Differentiable Physics Quantifier (DPQ), and a Microstructure Diagnostic Transformer (MDT). Specifically, the internal implementation of the 3D diagnostic module is a physically constrained end-to-end deep learning model processing flow, which includes at least three cascaded sub-modules: the Cross-Modal Structural Diffusion Module, the Differentiable Physics Quantifier, and the Microstructure Diagnostic Transformer.

[0073] The cross-modal structural diffusion module receives continuous two-dimensional skin tissue slice images (i.e., receives two-dimensional bright-field microscopic images) as conditional guidance. Utilizing a feature injection mechanism, it infers and reconstructs three-dimensional virtual fluorescent voxel data through a reverse diffusion process. In other words, the cross-modal structural diffusion module receives continuous two-dimensional skin tissue slice images as conditional guidance and, using a built-in feature injection mechanism, directly infers and reconstructs three-dimensional virtual fluorescent voxel data through a reverse diffusion process. CMSD learns the nonlinear cross-modal mapping relationship between two-dimensional optical texture features and the three-dimensional spatial distribution of collagen molecules.

[0074] The differentiable physical quantization (DPQ) module is connected after the cross-modal structure diffusion module. This module is equipped with a set of differentiable tensor operators for real-time computation of the generated three-dimensional virtual fluorescent voxel data, extracting physical parameters including collagen volume fraction, protein fiber anisotropy index, and curvature. DPQ supports gradient backpropagation to constrain the CMSD generation process with the rationality of the physical parameters.

[0075] The microstructure diagnostic Transformer is used to simultaneously receive three-dimensional virtual fluorescent voxel data generated by the cross-modal structure diffusion module and physical parameters output by the differentiable physics quantization module. It uses a multimodal fusion architecture to analyze the topology of the collagen network and outputs the level characterizing the skin aging state.

[0076] In this application, the cross-modal structural diffusion module is a conditional feature-injected 3D U-Net model. The backbone network of the conditional feature-injected 3D U-Net model includes symmetrical 3D encoder and 3D decoder paths. The backbone network integrates a feature injection layer at each resolution level of the 3D encoder and 3D decoder. The feature injection layer receives feature maps from continuous 2D skin tissue slice images (i.e., receives feature maps from 2D bright-field images), performs broadcast replication in the depth dimension to match the size of the 3D feature map, and fuses 2D texture information into the 3D spatial features through convolution operations to recover the 3D collagen fluorescence distribution aligned anatomically with the 2D slice texture from Gaussian random noise.

[0077] In this application, the differentiable physical quantization module is constructed as a differentiable computational layer, and integrates tensor operators for physical parameter extraction. The tensor operators include anisotropic extraction operators and density-morphological operators, both of which support automatic differentiation, enabling the diagnostic gradient of the subsequent MDT to backpropagate through DPQ, thereby constraining the physical rationality of the generated CMSD during the training phase.

[0078] An anisotropy extraction operator is used to calculate the gradient field of voxels using a 3D Sobel convolution kernel, construct a local structure tensor and solve for eigenvalues ​​to quantify the directional consistency of collagen fibers; a density morphology operator is used to calculate the collagen volume fraction through a differentiable Sigmoid approximation threshold function and to calculate the crimp index of collagen fibers through a skeletonization approximation algorithm.

[0079] In this application, the microstructure diagnostic Transformer is a hybrid feature encoding network with multimodal input. The input end is used to simultaneously receive three-dimensional virtual fluorescent voxel data generated by the cross-modal structure diffusion module and physical parameters output by the differentiable physics quantization module.

[0080] The hybrid feature encoding network consists of multiple stacked hybrid stages, each composed of a Local Texture Encoding Block (LTEB) and a Global Dependency Modeling Block (GDMB). The LTEB, based on a 3D convolutional neural network, is used to capture and extract the microscopic details of collagen fiber breakage and cross-linking. The GDMB is used to model the topology and spatial arrangement of the collagen network based on a Transformer-based self-attention mechanism. Through the collaborative work of LTEB and GDMB, physical parameters are concatenated into the high-dimensional semantic features of the image before the fully connected layer, enabling the final aging level determination to be based on both microscopic image features and macroscopic physical quantitative indicators.

[0081] Furthermore, for cross-modal structure diffusion modules, this application aims to address the issue of diffusion from a single 2D image I 2D Inferring 3D volume V 3D To address the ill-conditioned inverse problem, this paper proposes a deep feature broadcasting and injection mechanism. Unlike simple input-end concatenation, this application intervenes during the feature extraction stage of the network. Assume F... l 3D For the 3D feature map of the l-th layer of U-Net, conditional guidance is achieved through the following formula:

[0082] F l injected =Conv 1×1×1 (Concat[F l 3D Broadcast(E) l 2D (I 2D ))])

[0083] The Broadcast(•) operation copies and extends the 2D features along the depth axis, matching their dimensions to the 3D feature space. This mechanism ensures that the generated 3D voxels are subject to strict geometric constraints of the original 2D texture on each slice, guaranteeing the consistency of the anatomical structure. l injected This represents the fused 3D feature map of the l-th layer after feature injection, Conv 1×1×1 Used for feature fusion and channel dimensionality reduction, it represents a 3D convolution operation with a kernel size of 1×1×1, and Concat represents a concatenation operation along the channel dimension. l 2D This represents the two-dimensional feature map extracted by the two-dimensional encoder at layer l.

[0084] Furthermore, for the differentiable physics quantization module, taking the calculation of collagen fiber anisotropy as an example, this application utilizes convolutional layers to implement the 3D Sobel operator to calculate the spatial gradient of the generated voxel V. V=(Gx, Gy, Gz):

[0085] Gx = V K sobelx Gy = V K sobely Gz = V K sobelz ;

[0086] Based on the gradient, the pseudo-anisotropy index (PAI) is calculated as a surrogate loss during training:

[0087] PAI = (1 / |Ω|)∑_(v∈Ω) [√((Gx-μ x )²+ (Gy-μ y )²+ (Gz-μ z )²) / (ε+Gmag) ]

[0088] Where PAI represents the calculated pseudo-anisotropy exponent scalar, Ω represents the spatial domain of a 3D voxel, v represents any voxel point within the domain Ω, Gx, Gy, and Gz represent the gradient components of the voxel point in the x, y, and z directions, respectively, and μ... x μ y μ z G represents the mean of the gradient component in the local neighborhood, ε represents a small constant (1e-8) to prevent the denominator from being zero, and Gmag represents the gradient magnitude, i.e., √(Gx²+Gy²+Gz²).

[0089] This formula is implemented through fully differentiable tensor operations. This means that if the fiber structure generated by the model is not clear enough or its orientation does not conform to aging characteristics, the classifier's gradient will directly penalize the generator through the PAI formula, forcing it to generate fiber structures with clear physical meaning.

[0090] Furthermore, for the microstructure diagnosis Transformer, the ultimate optimization objective of this application is the multi-task joint loss:

[0091] L total =L diffusion +λ1L projection +λ2L diagnosis

[0092] Among them, L diffusion L represents the standard propagation noise prediction error, reflecting the generation quality; projectionEnsure that the texture contour of the generated 3D collagen mesh matches the input 2D bright-field image after flattening, preventing illusion constraints; L diagnosis The classification cross-entropy is based on the generated graph and physical parameters, reflecting the accuracy of loss diagnosis; λ1 is used to balance structure fidelity and represents the weight coefficient of projection consistency loss; λ2 is used to adjust the gradient influence of the auxiliary task on the main generation task and represents the weight coefficient of diagnostic classification loss.

[0093] The following is a detailed description of the skin aging diagnostic system based on physical sensing diffusion proposed in this application.

[0094] See Figure 2 The diagram shown is a schematic of the overall workflow of the skin aging diagnosis system of this application, which mainly includes a data acquisition and preprocessing module, a training dataset construction module, a three-dimensional diagnosis module, and a report generation and output module.

[0095] The data acquisition and preprocessing module is responsible for data acquisition and standardization preparation. In this application, the biological sample data (image data) comes from legally obtained and discarded human skin tissue during clinical surgical procedures (such as blepharoplasty, abdominoplasty, etc.). The samples cover individuals of different age groups (e.g., <20 years old, 20-30 years old, 30-40 years old, 40-50 years old, >50 years old) and different anatomical locations (eyelids, cheeks, abdomen, etc.).

[0096] The obtained tissue samples were fixed in formalin and embedded in paraffin, and then prepared into serial sections with a thickness of 5–10 micrometers. Subsequently, the sections were subjected to routine histological staining to clearly display the collagen fiber structure under a bright-field microscope.

[0097] Using the transmission light detector (TD) channel of a confocal microscope, stained sections were digitally scanned at high resolution to acquire a series of continuous two-dimensional bright-field images as input for the skin aging diagnostic system. Simultaneously, to construct the gold standard three-dimensional ground truth required for training the CMSD, paired sections were immunofluorescently stained (using anti-type I collagen antibody and labeled with FITC / Cy3; DAPI-labeled cell nuclei) and imaged in the corresponding fluorescence channels to acquire three-dimensional fluorescent voxel data.

[0098] The acquired continuous two-dimensional skin tissue slice images are subjected to standardized preprocessing, which includes: image registration to correct translation, rotation and deformation between slices; color standardization to eliminate color deviation caused by differences in staining or imaging conditions; and artifact removal algorithm to automatically identify and remove air bubbles and dust interference.

[0099] See Figure 3The diagram shows the structure of a physically constrained end-to-end deep learning model, which is constructed as a physically constrained end-to-end perceptual deep learning framework. Unlike traditional cascaded processing, this application couples generation and diagnostic tasks through differentiable physical operators, achieving a direct mapping from two-dimensional bright-field images to three-dimensional pathological diagnosis. The physically constrained end-to-end deep learning model specifically includes the following three co-optimized sub-stages:

[0100] Phase 1, Cross-Modal Structure Diffusion Module (CMSD):

[0101] See Figure 4 The diagram shows the network architecture of the cross-modal structure diffusion module, illustrating a two-dimensional conditionally guided three-dimensional generation mechanism based on feature injection, used to establish a nonlinear mapping relationship between two-dimensional bright-field texture features and three-dimensional fluorescence spatial structure. CMSD adopts a conditional 3D U-Net-based architecture, which includes a contraction path (3D Encoder) and a symmetrical expansion path (3D Decoder). The skin aging diagnostic system first receives the two-dimensional bright-field image feature map I. 2D As conditional input, the original 2D feature map (Input) is first copied and expanded in the depth dimension through a broadcast operation, and directly injected into the top layer of the 3D Encoder, providing high-resolution texture constraints for the shallow network. The Input image simultaneously enters the 2D Feature Extractor, which consists of a multi-layer convolutional neural network, extracting 2D semantic features at different resolution levels from shallow to deep. The extracted 2D features at each level are also expanded into 3D tensors through a broadcast operation, and then enter the corresponding layer of the 3D Encoder through the Feature Injection path. For example... Figure 4 As shown by the dashed arrow, the injection process uses concat to fuse two-dimensional conditional features and three-dimensional spatial features in the channel dimension.

[0102] The 3D encoder is responsible for extracting 3D spatial features and performing dimensionality reduction. Each encoding stage includes a feature injection concat layer and a 3D residual block (3D ResBlock). The 3D ResBlock is responsible for the fusion and extraction of deep features. Layers are connected by downsampling layers. This application preferably uses max pooling / stride convolution to halve the resolution and gradually expand the receptive field. At the bottom of the network, there is a bottleneck layer, composed of 3D ResBlocks, used to process highly abstract deep semantic features.

[0103] The 3D decoder is responsible for restoring low-resolution features to high-resolution 3D voxel data. The data flow follows a bottom-up restoration path. The output of the bottleneck layer first passes through an upsampling layer; this application preferably uses a 3D transposed convolution (TransConv) to achieve resolution doubling. Skip connections are the core of the U-Net architecture. Figure 4 As indicated by the horizontal solid arrow, each level of the 3D Decoder has a Concat layer, which simultaneously receives: (1) the upsampled feature map from below; and (2) the original feature map from the same level of the encoder on the left. This design effectively preserves the high-frequency spatial details of the 3D structure. The concatenated feature map is then fed into 3DResBlock for further refinement.

[0104] After progressive upsampling and feature fusion, the top layer of the 3D Decoder outputs the final feature tensor. This tensor is then processed by the output convolutional layer to generate the target's 3D virtual fluorescent voxel data (V). gen ).

[0105] The second phase, Differentiable Physics Quantization Module (DPQ):

[0106] See Figure 5 The diagram shows the architecture of the differentiable physics quantization module. DPQ is constructed as a differentiable computation layer embedded in a deep neural network, configured to receive three-dimensional virtual fluorescent voxel data (V3) generated by CMSD. gen It calculates biophysical parameters in real time through two parallel computing paths and supports gradient backpropagation.

[0107] Anisotropy extraction unit: Input voxel V gen First, the data is processed by a set of fixed, non-trainable 3D Sobel convolution kernels to calculate the gradient field in space for each voxel. V), capturing texture edge information. A local structure tensor matrix is ​​calculated based on the gradient field to describe the main texture direction within the local neighborhood. Eigenvalue analysis (Eigen Solver) is performed on the structure tensor using a differentiable eigenvalue decomposition algorithm to obtain eigenvalues ​​in three orthogonal directions. An anisotropy index (FA) is calculated based on the eigenvalue differences, and this index is directly output to the biophysical feature vector (V). phys )middle.

[0108] Density and Morphology Calculation Unit: To maintain the differentiability of the calculation process, this application employs a sigmoid soft thresholding function instead of the traditional binarization thresholding operation, continuously mapping voxel intensity to the (0,1) interval to generate a soft mask. Subsequently, a skeletonization approximation operation is performed on the soft mask to extract the centerline features of collagen fibers to capture the fiber's topological morphology. Based on the above processing results, the proportion of collagen in the tissue volume, i.e., density, is calculated globally using statistical methods as a key aging indicator.

[0109] The scalar metrics such as FA and Density calculated from the two paths mentioned above are aggregated and output as Biophysical Feature (Vphys) vectors. The core characteristic of the differentiable physical quantization module lies in its gradient backpropagation flow. Since all operators within the differentiable physical quantization module, including convolutions, algebraic operations, and activation functions, are built upon fundamental differentiable operations, the diagnostic classification loss generated by the subsequent MDT can be transformed into a loss against Vphys. phys The gradient is backpropagated losslessly to the input V. gen This, in turn, updates the parameters of the front-end CMSD. This mechanism enables explicit constraints on the generation process by physical parameters, forcing the network to generate three-dimensional structures that conform to biophysical laws.

[0110] The third stage, the microstructure diagnostic module (MDT):

[0111] See Figure 6 The diagram shows the architecture of the microstructure diagnostic module, illustrating the hierarchical image feature extraction process and physical-visual feature fusion mechanism. The MDT is constructed as a hybrid visual Transformer architecture with multimodal input.

[0112] The MDT is designed with dual-stream input interfaces, receiving high-dimensional image features (F) output from the CMSD. img ) and the physical feature vector (V) output by DPQ phys The input 3D voxels are first subjected to preliminary feature mapping and downsampling through Stem layers composed of stride convolutions, and then enter four cascaded feature extraction stages (Stage 1~4), in which the resolution of the feature map decreases step by step and the number of channels increases step by step.

[0113] Stages 1 and 2 are shallow texture extraction stages, both consisting entirely of stacked Local Texture Encoding Blocks (LTEBs). The LTEBs preferably employ a 3D residual convolutional structure, with each LTEB containing two paths: the main path sequentially passes through a 3D convolutional layer (Conv 3D), batch normalization (BN), and a ReLU activation function to extract the microscopic edges and texture details of collagen fibers; the residual path directly transmits input features to alleviate the gradient vanishing problem. To balance the capture of local texture and global semantics, Stage 3 serves as a hybrid transition stage, employing a hybrid architecture that connects the LTEB and GDMB. This design allows the network to maintain sensitivity to fiber breakpoints while beginning to establish preliminary long-range spatial dependencies. Stage 4 is the deep semantic modeling stage, entirely composed of the Global Dependency Modeling Block (GDMB). The microstructure diagnosis module introduces a self-attention mechanism, which includes: (i) a three-dimensional convolutional projection layer, used to introduce positional information bias; (ii) an efficient multi-head self-attention unit (E-MHSA), used to calculate the correlation strength between any two patches in three-dimensional space, thereby capturing the overall collapse, relaxation and other macroscopic topological structures of the dermis; and (iii) a feedforward neural network (FFN), used to compress the high-dimensional feature map output by feature transformation / Stage 4 into a one-dimensional image feature vector through global average pooling.

[0114] The skin aging diagnostic system introduces a biophysical feature vector (V) calculated by DPQ, which includes parameters such as density and anisotropy, at the fusion point. phys The fusion operation employs a channel concatenation strategy to combine image features with physical features, forming a multimodal diagnostic vector. The fused vector is then fed into a fully connected layer classifier (Diagnostic Output), which ultimately outputs the predicted skin aging level (Level IV).

[0115] It is necessary to further explain the interpretability and physical validity verification of the skin aging diagnostic system in this application. To verify whether the skin aging diagnostic system (Model_ColRejuv) makes decisions based on correct pathological features, gradient-weighted class activation mapping (Grad-CAM) technology was used to visualize and analyze the diagnostic process. See [link to relevant documentation]. Figure 7The image shown is a visual comparison verification diagram of the three-dimensional virtual fluorescent collagen network image generated by the skin aging diagnosis system of this application and the real confocal microscopic fluorescence image (Ground Truth) (green channel: FITC-labeled collagen; red channel: Cy3-labeled background). Figure 7 In the middle, the top three images are: generated virtual fluorescence → comparison of the generated image with the true value checkerboard pattern (verifying alignment) → true fluorescence ground truth (FITC). The bottom three images are: input two-dimensional bright field image → generated background / red light channel → true red light ground truth (Cy3).

[0116] See Figure 8 The image shows a visualization output of aging levels and an example of an interpretable heatmap visualization based on the Grad-CAM algorithm. Figure 8 As shown on the right, the skin aging diagnostic system outputs attention heatmaps for typical samples: For Level I (<20 years old) samples, a high confidence prediction of 0.92 is given. The warm-toned areas of the heatmap are highly concentrated in areas of intact, densely arranged collagen fiber bundles in the image, indicating that the LTEB coding block in the MDT successfully extracted the continuity and density features of collagen fibers as a basis for judging youthfulness. For Level III (30-40 years old) samples, a prediction confidence of 0.96 is given. The heatmap shows that the model's focus area has shifted, concentrating on areas with less blurred fiber texture and relatively loose structure. This visualization result confirms that the diagnostic logic of the skin aging diagnostic system in this application is highly consistent with the gold standard of clinical pathology, that is, the skin aging diagnostic system can automatically identify and locate microstructural changes that lead to skin aging, such as collagen breakage and decreased density.

[0117] To verify the validity of DPQ, see [link to relevant documentation]. Figure 9 As shown, this is a schematic diagram of the correlation between the physical parameters output by DPQ and the predicted age. After the skin aging diagnosis system performs gridding processing on the input image and makes predictions, it statistically analyzes the physical parameters of all sub-regions and their corresponding aging level classification results. Figure 9 The statistical chart on the left shows that the collagen volume fraction (Density) calculated by the skin aging diagnostic system of this application can clearly distinguish between the middle-aged group (30-40 years old) and the elderly group (>40 years old). The density value of the elderly group is significantly reduced, which is consistent with the pre-set pathological prior knowledge and proves the accuracy of the density morphology operator in DPQ. Figure 9 The statistical chart on the right shows a significant numerical transition in the fiber anisotropy index between different aging levels. Utilizing this characteristic, the skin aging diagnostic system can assist MDT (Multidisciplinary Team) in making auxiliary corrections based on the distribution characteristics of physical parameters when texture features are not obvious, thereby improving the robustness of the diagnosis.

[0118] Furthermore, regarding the output and statistical validation of diagnostic reports, the report generation and output module is the final output of this skin aging diagnostic system, responsible for converting complex three-dimensional data into a format easily understood by clinicians and researchers. This module receives classification results from the MDT (such as aging grade IV) and simultaneously outputs various quantitative parameters V. phys To assist in clinical decision-making.

[0119] For specific output formats, see [link to specific output formats]. Figure 10 As shown, this application demonstrates an example of quantization index output. It can perform gridded scanning and reconstruction of large-field-of-view two-dimensional slices, outputting a quantized map containing spatial positioning information. Specifically, Figure 10 The left image shows a collagen density distribution map, which visually displays the changes in the volume fraction of collagen in the tissue through pseudo-color mapping. The dark areas represent areas with sparse collagen. Figure 10 The middle image shows the fiber anisotropy diagram, which quantifies the degree of orderliness of collagen fiber arrangement. The bright areas represent that the fiber arrangement has strong directionality. Figure 10 The image on the right shows a classification map of local aging levels, with values ​​2 to 4 representing progressive levels from youth to old age. The three maps correspond one-to-one in spatial coordinates, achieving precise digital localization of the entire slice.

[0120] See the data validation section. Figure 11 As shown, in order to verify the effectiveness of the feature space and the classification criteria, a large-scale clinical sample was statistically analyzed. Figure 11 The joint distribution map on the left reveals a significant structure-morphology coupling evolution pattern during skin aging: the young group samples (blue) are concentrated in the high-density-low curvature region, while the aging group samples (red) migrate to the low-density-high curvature region, confirming the negative correlation between collagen loss and fiber relaxation. Figure 11 As shown in the principal component analysis (PCA) scatter plot on the right, the young group and the aging group form independent clusters with clear boundaries in the feature space. The first principal component (PC1, which explains 67.5% of the variance) is defined as the morphological aging axis, providing a solid mathematical foundation for high-precision classification.

[0121] This application's embodiment of a skin aging diagnostic system based on physical sensing diffusion constructs a highly clinically representative gold standard training dataset and proposes an end-to-end computational dimensionality enhancement and quantitative analysis framework. It receives clinically readily available two-dimensional bright-field images of tissues and then reconstructs three-dimensional virtual fluorescent voxels representing the specific distribution of collagen using a feature injection mechanism via a cross-modal structural diffusion module. Subsequently, a differentiable physical quantification module calculates biophysical parameters such as collagen density, anisotropy, and curvature in real time, using physical constraints to guide the generation process. Finally, a microstructure diagnostic module fuses image features and physical parameters to output the aging level. This application efficiently transforms low-cost two-dimensional image input into objectively quantified three-dimensional structural output. By reconstructing the three-dimensional microstructure of collagen through cross-modal generation technology and combining it with differentiable physical quantification methods, it provides high-precision, automated quantitative assessment of skin collagen status and aging degree, assisting clinical decision-making. This significantly improves the accuracy and interpretability of skin aging diagnosis, providing precise and quantifiable three-dimensional structural evidence for clinical diagnosis, targeted repair treatment, and prognostic assessment of skin collagen status.

[0122] Secondly, this application also provides a method for determining the level of skin aging, based on the skin aging diagnostic system described above.

[0123] The method for determining the skin aging grade in this application includes:

[0124] S1: Receive continuous two-dimensional skin tissue slice images of the target object and input them into the skin aging diagnosis system;

[0125] S2: Guided by the received continuous two-dimensional skin tissue slice images, infer and generate three-dimensional virtual fluorescent voxel data characterizing the specific distribution of collagen;

[0126] S3: Perform real-time differentiable calculations on the generated three-dimensional virtual fluorescent voxel data to extract physical parameters including collagen volume fraction, protein fiber anisotropy index, and curvature;

[0127] S4: Integrates three-dimensional virtual fluorescent voxel data and extracted physical parameters to output a level characterizing the skin's aging state;

[0128] S5: Based on the physical parameters and the output levels characterizing skin aging, generate a quantitative report to guide clinical diagnosis and biomaterial preparation.

[0129] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0130] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0131] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0132] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0133] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0134] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A skin aging diagnostic system based on physical sensing diffusion, characterized in that, The skin aging diagnostic system based on physical sensing diffusion includes: The data acquisition and preprocessing module is used to receive continuous two-dimensional skin tissue slice images of the target object, and to perform standardization processing on the continuous two-dimensional skin tissue slice images based on image processing algorithms. The continuous two-dimensional skin tissue slice images include at least two slices of different depths. The training dataset construction module is used to map and stack two-dimensional annotation information based on the labeling of collagen fiber age status, using spatial alignment and dimensionality increase algorithms to construct a three-dimensional collagen structure truth label body with spatial continuity, forming a gold standard dataset to train a physically constrained end-to-end deep learning model. The 3D diagnostic module is used to take standardized continuous two-dimensional skin tissue slice images as input based on a pre-trained end-to-end deep learning model with physical constraints. It performs cross-modal reconstruction and semantic segmentation of collagen fiber structures in the sequence images in 3D space, and generates and outputs a 3D collagen structure map and corresponding physical parameters. The report generation and output module is used to analyze the output of the three-dimensional diagnostic module and output a quantitative report on skin aging level to guide clinical diagnosis and biomaterial preparation; The physical constraint end-to-end deep learning model includes a cross-modal structure diffusion module, a differentiable physical quantization module, and a microstructure diagnostic Transformer. The cross-modal structure diffusion module is used to receive continuous two-dimensional skin tissue slice images as conditional guidance, and uses a feature injection mechanism to infer and reconstruct three-dimensional virtual fluorescent voxel data through a reverse diffusion process. The differentiable physical quantization module is connected after the cross-modal structure diffusion module. The differentiable physical quantization module is equipped with a set of differentiable tensor operators for real-time calculation of the generated three-dimensional virtual fluorescent voxel data, and extraction of physical parameters including collagen volume fraction, protein fiber anisotropy index and curvature. The microstructure diagnostic Transformer is used to simultaneously receive three-dimensional virtual fluorescent voxel data generated by the cross-modal structure diffusion module and physical parameters output by the differentiable physics quantization module. It uses a multimodal fusion architecture to analyze the topology of the collagen network and outputs the level characterizing the skin aging state.

2. The skin aging diagnostic system based on physical sensing diffusion as described in claim 1, characterized in that: The continuous two-dimensional skin tissue slice images are multi-channel imaging data acquired by confocal microscopy. The multi-channel imaging data consists of two modal images synchronously acquired from the same tissue slice under the same field of view. The modal images include confocal transmission light microscopy and multichannel fluorescence confocal microscopy. The confocal transmission light microscopy is a bright-field image of a tissue section that has undergone standard histological staining, displayed under the transmission light detector channel. The multichannel fluorescence confocal microscopy imaging consists of fluorescence images labeled with DAPI and FITC or Cy3 dyes, respectively, showing the spatial distribution of cell nuclei and type I collagen.

3. The skin aging diagnostic system based on physical sensing diffusion as described in claim 1, characterized in that: The image processing algorithms include Reinhard color migration algorithm, scale-invariant feature transform matching algorithm, and Otsu threshold segmentation algorithm. The standardization process includes geometric correction, color standardization, and artifact removal.

4. The skin aging diagnostic system based on physical sensing diffusion as described in claim 1, characterized in that: The training dataset construction module also includes an image registration strategy for structure preservation. The image registration strategy involves using a scale-invariant feature transformation algorithm to detect key feature points and estimate the initial affine transformation matrix to correct local nonlinear displacements.

5. The skin aging diagnostic system based on physical sensing diffusion as described in claim 1, characterized in that: The cross-modal structure diffusion module is a conditional feature injection type 3D U-Net model, and the backbone network of the conditional feature injection type 3D U-Net model contains symmetrical 3D encoder and 3D decoder paths. The backbone network integrates a feature injection layer at each resolution level of the 3D encoder and 3D decoder. The feature injection layer is used to receive feature maps of continuous two-dimensional skin tissue slice images, broadcast and copy them in the depth dimension to match the size of the three-dimensional feature map, and fuse two-dimensional texture information into three-dimensional spatial features through convolution operation, so as to recover the three-dimensional collagen fluorescence distribution that is aligned with the two-dimensional slice texture in terms of anatomical structure from Gaussian random noise.

6. The skin aging diagnostic system based on physical sensing diffusion as described in claim 1, characterized in that: The differentiable physical quantization module is constructed as a differentiable computational layer and integrates tensor operators for physical parameter extraction. The tensor operators include anisotropic extraction operators and density-morphological operators, and both anisotropic extraction operators and density-morphological operators support automatic differentiation.

7. The skin aging diagnostic system based on physical sensing diffusion as described in claim 6, characterized in that: The anisotropic extraction operator is used to calculate the gradient field of voxels using a three-dimensional Sobel convolution kernel, construct a local structure tensor and solve for eigenvalues ​​to quantify the directional consistency of collagen fibers. The density morphology operator is used to calculate the collagen volume fraction using a differentiable sigmoid approximation threshold function, and to calculate the crimp index of collagen fibers using a skeletal approximation algorithm.

8. The skin aging diagnostic system based on physical sensing diffusion as described in claim 1, characterized in that: The microstructure diagnostic Transformer is a hybrid feature encoding network with multimodal input. The input end is used to simultaneously receive three-dimensional virtual fluorescent voxel data generated by the cross-modal structure diffusion module and physical parameters output by the differentiable physics quantization module. The hybrid feature encoding network consists of multiple stacked hybrid stages, and each hybrid stage is composed of stacked local texture encoding blocks and global dependency building blocks; The local texture coding block is based on a three-dimensional convolutional neural network and is used to capture and extract the microscopic fracture and cross-linking details of collagen fibers. The global dependency modeling module is used to model the topology and spatial arrangement of collagen networks based on the Transformer self-attention mechanism.

9. A method for determining the level of skin aging, based on the skin aging diagnostic system according to any one of claims 1 to 8, characterized in that, The method for determining the level of skin aging includes: Receive continuous two-dimensional skin tissue slice images of the target object and input them into the skin aging diagnosis system; Guided by received continuous two-dimensional skin tissue slice images, three-dimensional virtual fluorescent voxel data characterizing the specific distribution of collagen are inferred and generated; Real-time differentiable calculations were performed on the generated three-dimensional virtual fluorescent voxel data to extract physical parameters including collagen volume fraction, protein fiber anisotropy index, and curvature. By integrating three-dimensional virtual fluorescent voxel data and extracted physical parameters, a level characterizing the skin's aging state is output; Based on the physical parameters and the output levels characterizing skin aging, a quantitative report is generated to guide clinical diagnosis and biomaterial preparation.

Images