An organoid three-dimensional image enhanced segmentation method and system

Abstract

The present application relates to the technical field of biomedical image processing and analysis, in particular to a kind of organoid three-dimensional image enhancement segmentation method and system;Contain data acquisition, light field analysis, physical enhancement, topological feature extraction and boundary evolution module;System utilizes deep learning to construct light transmission attenuation atlas, carries out voxel-level reverse light compensation, and carries out dynamic evolution segmentation in combination with centripetal vector and topological repulsive potential field;Its core is based on inverse process of Beer-Lambert law realizes digital transparency, eliminates light transmission physical degradation, so that the brightness distribution of deep and shallow layer cells is consistent;The present application realizes the depth invariance of feature extraction, effectively solves the problem that subsequent segmentation algorithm is sensitive to depth change.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical image processing and analysis, in particular to a three-dimensional image enhancement and segmentation method and system for organoids. BACKGROUND

[0002] With the in-depth application of organoid technology in the biomedical field, the demand for high-throughput three-dimensional imaging and fine analysis of organoids is increasing; such analysis is heavily dependent on high-quality three-dimensional microscopic image data.

[0003] Currently, confocal microscopes or two-photon microscopes are usually used to perform tomographic scanning on organoid samples, and conventional image processing algorithms are used for segmentation; however, due to the optical properties of biological tissues, the optical signal will be severely scattered and attenuated when penetrating deep tissues, resulting in a significant reduction in the signal-to-noise ratio of deep region images and loss of texture details; in addition, in high-density culture scenarios, adjacent organoids often come into close contact or even adhesion, making the optical boundary of the contact surface extremely blurred, presenting a zero gradient feature; existing segmentation methods are difficult to effectively recover structural information from degraded physical signals and cannot accurately define the blurred instance boundary, which easily leads to difficulties in identifying deep targets and failure in separating adhesion regions.

[0004] Therefore, how to overcome the signal attenuation caused by light transmission and achieve accurate separation of high-density organoids under weak boundary conditions has become a problem to be solved in the field. SUMMARY

[0005] To solve the above technical problems, the present application provides a three-dimensional image enhancement and segmentation method and system for organoids, specifically, the technical solution of the present application is as follows:

[0006] A three-dimensional image enhancement and segmentation method for organoids, comprising:

[0007] acquiring initial three-dimensional image data containing organoid targets;

[0008] performing light field analysis on the initial three-dimensional image data using a deep learning model to construct a light transmission attenuation atlas, the light transmission attenuation atlas including background light field data representing background illumination intensity and a local attenuation coefficient matrix representing medium light transmission characteristics;

[0009] based on the light transmission attenuation atlas, performing voxel-level inverse illumination compensation on the initial three-dimensional image data to generate physical enhancement image data;

[0010] inputting the physical enhanced image data into a preset topological prediction model to extract a multi-dimensional topological feature field, the multi-dimensional topological feature field including a target probability map representing a probability of existence of a target, a centripetal vector field representing a direction of a center of the target, and a topological repulsive potential field representing a separation barrier between targets;

[0011] constructing a dynamic evolution equation according to the centripetal vector field and a repulsive force field derived from the topological repulsive potential field, and performing instance boundary evolution on the organoid target to generate a three-dimensional segmentation result;

[0012] The topological repulsive potential field is used to define a segmentation boundary between different organoid targets in a region where a gradient modulus is lower than a preset threshold.

[0013] Preferably, the method for constructing the light transmission attenuation atlas comprises:

[0014] inputting the initial three-dimensional image data into a convolutional neural network;

[0015] outputting background light field distribution data consistent in size with the initial three-dimensional image data;

[0016] outputting a local attenuation coefficient matrix representing an attenuation rate of light in a biological tissue along a depth direction;

[0017] combining the background light field distribution data and the local attenuation coefficient matrix into the light transmission attenuation atlas.

[0018] Preferably, the method for generating the physical enhanced image data comprises:

[0019] subtracting the background light field distribution data from the initial three-dimensional image data to obtain net signal data;

[0020] integrating the local attenuation coefficient matrix along the depth direction to generate a cumulative attenuation factor;

[0021] performing exponential gain compensation on the net signal data based on the cumulative attenuation factor to restore voxel intensity in a deep region, thereby obtaining the physical enhanced image data.

[0022] Preferably, the topological repulsive potential field in the multi-dimensional topological feature field is generated by:

[0023] identifying local extreme value regions in the physical enhanced image data;

[0024] constructing a repulsive force field based on distances between the local extreme value regions;

[0025] defining an equipotential surface with zero potential in the repulsive force field as a segmentation boundary, and a voxel position corresponding to the equipotential surface being a contact interface of adjacent organoid targets.

[0026] Preferably, the method of evolving instance boundary of the organoid target comprises:

[0027] calculating a signal-to-noise ratio indicator of the organoid target at a current depth position;

[0028] adjusting a weight coefficient of the geometric rigidity constraint based on the signal-to-noise ratio indicator;

[0029] configured to: when the signal-to-noise ratio indicator is lower than a preset threshold, increase the weight coefficient to enhance the shape constraint force of the organoid target tending to a regular sphere;

[0030] combined with the aggregation force provided by the centripetal vector field, the separation force provided by the topological repulsive potential field, and the shape maintaining force provided by the geometric rigidity constraint, iteratively updating the contour position of the organoid target.

[0031] Preferably, adjusting the weight coefficient of the geometric rigidity constraint follows the following rules:

[0032] establishing a nonlinear mapping relationship between the depth coordinate and the weight coefficient;

[0033] monotonically increasing the weight coefficient with the increase of the depth coordinate;

[0034] wherein the geometric rigidity constraint is realized by minimizing the gradient modulus of the level set function.

[0035] Preferably, the training process of the topological prediction model comprises:

[0036] constructing a composite loss function comprising a physical consistency loss and a geometric topology loss;

[0037] using the physical consistency loss to constrain the generation of the light transmission attenuation atlas, ensuring that the physically enhanced image data conforms to the law of light transmission physics;

[0038] using the geometric topology loss to constrain the generation of the centripetal vector field and the topological repulsive potential field, ensuring the continuity of the three-dimensional topological structure under the condition of sparse annotation.

[0039] An organoid three-dimensional image enhancement and segmentation system comprises:

[0040] a data acquisition module for acquiring initial three-dimensional image data containing an organoid target;

[0041] a physical modeling module for performing light field analysis on the initial three-dimensional image data using a deep learning model to construct a light transmission attenuation atlas;

[0042] An inverse enhancement module is configured to perform voxel-level inverse light compensation on the initial three-dimensional image data based on the light transmission attenuation map to generate physical enhancement image data;

[0043] A feature prediction module is configured to input the physical enhancement image data into a preset topological prediction model to extract a multi-dimensional topological feature field, which includes a target probability map representing a target existence probability, a centripetal vector field representing a direction pointing to a target center, and a topological repulsive potential field representing a separation barrier between targets;

[0044] A topological segmentation module is configured to construct a dynamic evolution equation based on the centripetal vector field and a repulsive force field derived from the topological repulsive potential field to perform instance boundary evolution on the organoid target to generate a three-dimensional segmentation result.

[0045] Compared with the prior art, the present application has the following beneficial effects:

[0046] 1. The present application uses a deep learning model to decouple image degradation into a background light field and a local attenuation coefficient, and performs exponential gain compensation on voxels according to the inverse process of the Beer-Lambert law. This digital transparentization at the physical level makes deep cells have similar brightness distribution to shallow cells in the enhanced image, eliminates the influence of light transmission physical degradation on image quality, reduces the sensitivity of the subsequent segmentation algorithm to depth changes, and realizes the depth invariance of feature extraction.

[0047] 2. In view of the problem that the optical boundary of adjacent organoids is blurred or even presents zero gradient characteristics, causing the failure of traditional algorithms, the present application does not simply rely on image gradients, but constructs a repulsive force field by identifying local extreme regions, and defines an equipotential surface with zero potential as a segmentation boundary in the area where the gradient module length is lower than the threshold. This method is equivalent to establishing an invisible wall in the adhesion area, forcing the initial seed point to avoid crossing the boundary during the evolution process, thereby realizing accurate segmentation of closely contacted or even adhered targets.

[0048] 3. The present application establishes a nonlinear mapping relationship between depth coordinates, signal-to-noise ratio indicators, and geometric weight coefficients. In the shallow area where the image is clear, the system allows the segmentation result to present irregular real biological morphology. In the deep area where the signal-to-noise ratio is low, the system automatically increases the weight coefficient and uses strong prior mode to enhance the shape constraint force of the organoid tending to a regular sphere. This dynamic balance mechanism effectively counteracts the deep spot noise and prevents the outline edge burr and fragmentation caused by signal degradation.

[0049] 4. In the model training phase, the physical consistency loss forces the network to decompose the parameters to meet the physical law of light transmission, preventing the model from generating false hallucinated textures; at the same time, the geometric topology loss uses the vector field smoothness and potential field regression constraints to strengthen the spatial continuity between voxels; this dual constraint mechanism enables the system to maintain the integrity of the three-dimensional topological structure even under sparse annotation conditions, avoiding unreasonable discrete points or topological breaks in the segmentation results. BRIEF DESCRIPTION OF DRAWINGS

[0050] The application will be further explained in conjunction with the accompanying drawings and embodiments:

[0051] Figure 1 is a flowchart of the method of the application;

[0052] Figure 2 is a structural diagram of the system of the application. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical scheme and advantages of the application clearer and more apparent, the application will be further described in detail below in conjunction with specific embodiments.

[0054] Embodiment 1

[0055] Please refer to Figure 1 An organoid three-dimensional image enhanced segmentation method, comprising:

[0056] Collecting initial three-dimensional image data containing organoid targets;

[0057] Performing light field analysis on the initial three-dimensional image data using a deep learning model to construct a light transmission attenuation atlas, the light transmission attenuation atlas including background light field data representing background light intensity and a local attenuation coefficient matrix representing medium light transmission characteristics;

[0058] Based on the light transmission attenuation atlas, performing voxel-level inverse light compensation on the initial three-dimensional image data to generate physically enhanced image data;

[0059] Inputting the physically enhanced image data into a preset topology prediction model to extract a multi-dimensional topology feature field, the multi-dimensional topology feature field including a target probability map representing the probability of target existence, a centripetal vector field representing the direction pointing to the center of the target, and a topological repulsive potential field representing the separation barrier between targets;

[0060] According to the centripetal vector field and the topological repulsive potential field, constructing a dynamic evolution equation to perform instance boundary evolution on the organoid targets to generate a three-dimensional segmentation result;

[0061] The topological repulsive potential field is used to define the segmentation boundary between different organoid targets in regions where the gradient module length is lower than a preset threshold.

[0062] The embodiment discloses in detail the physical-topological double-cycle execution logic of the method, and aims to solve the problem of blurred boundary of deep area of high-density organoids due to attenuation of physical signals; the system performs a data acquisition step, uses a confocal microscope or a two-photon microscope to perform tomographic scanning on a fluorescence-labeled organoid sample, and obtains initial three-dimensional image data ; the data is derived from high-throughput sampling of an optical sensor, and has a physical meaning of original light intensity spatial distribution containing an attenuation signal affected by scattering of biological tissue and environmental background noise; the system starts a light field analysis program, uses a deep learning model to decouple brightness degradation of the image into two physical processes of additive noise and multiplicative attenuation, and constructs a light transmission attenuation atlas; on this basis, the system performs a digital transparency operation at a physical level, restores signals of each voxel according to an inverse process of the Beer-Lambert law, and generates physically enhanced image data ; further, the data is input into a topological prediction model, and target probability maps, centripetal vector fields and topological repulsive potential fields of cores are output; the system constructs a dynamic evolution equation according to the centripetal vector field and the topological repulsive potential field, and in response to an edge blurred area with a gradient modulus lower than a preset threshold value, the threshold value is preferably set as a 90th percentile value of a gradient modulus histogram of the current input image, or an adaptive low threshold value calculated by an Otsu method, and the topological repulsive potential field is used to forcibly define a segmentation boundary, so that initial seed points aggregate under the action of the centripetal field and avoid crossing the boundary under the action of the repulsive potential field, and finally generated equipotential surfaces are cell surfaces after segmentation;

[0063] The embodiment effectively restores texture details of a deep area of an organoid by constructing a light transmission attenuation atlas for reverse physical enhancement; in a high-density organoid culture scene, an artificial boundary defined by a geometric potential is established in a zero-gradient area with blurred optical boundaries by introducing a topological repulsive potential field, so that accurate separation of deep weak signal targets and high-density contact cells is realized, and the problem that a traditional watershed algorithm fails in a sticking area is solved.

[0064] Embodiment 2

[0065] The method for constructing the light transmission attenuation atlas comprises the following steps:

[0066] inputting the initial three-dimensional image data into a convolutional neural network;

[0067] outputting background light field distribution data consistent in size with the initial three-dimensional image data;

[0068] outputting a local attenuation coefficient matrix representing an attenuation rate of light in a biological tissue along a depth direction;

[0069] The background light field distribution data and the local attenuation coefficient matrix are combined into a light transmission attenuation atlas.

[0070] The embodiment is a further embodiment of the above light field analysis step; the process adopts a lightweight 3D full convolutional neural network, such as 3D U-Net and its variant structure, as the analysis core; specifically, the convolutional neural network adopts an encoder-decoder architecture, and is designed as a multi-task parallel output head at the end of the decoder: the first output head outputs a single-channel background light field B through a 1x1x1 convolutional layer, and the second output head outputs a single-channel local attenuation coefficient ; both output heads adopt a Softplus activation function to meet the non-negative physical constraint;

[0071] The initial three-dimensional image data is input into the network; the first branch of the network outputs the background light field B distribution data, which is derived from the network's feature extraction of the low-frequency components of the image, and has a physical meaning of representing the additive background noise distribution caused by non-specific fluorescence and ambient light; at the same time, the second branch of the network outputs the local attenuation coefficient , which is derived from the network's deep learning of tissue heterogeneity, and has a physical meaning of representing the light intensity attenuation rate per unit length when light passes through a specific voxel position; this parameter varies with spatial position and reflects the structural differences of the organoid, such as cystic cavities or dense cell clusters; in order to meet the non-negative constraint of the physical quantity of light transmission, i.e., the light intensity and attenuation coefficient cannot be negative, a Softplus activation function is introduced at the end of the output layer of the convolutional neural network, which is mathematically expressed as:

[0072] ;

[0073] This operation ensures that the output background light field B and local attenuation coefficient always satisfy and , thereby avoiding complex numbers or numerical divergence problems in the subsequent integration step; the system combines and outputs these two groups of physical parameters.

[0074] The method for generating physically enhanced image data comprises:

[0075] Subtracting the background light field distribution data from the initial three-dimensional image data to obtain net signal data;

[0076] Integrating the local attenuation coefficient matrix along the depth direction to generate a cumulative attenuation factor;

[0077] Based on the cumulative attenuation factor, the net signal data is exponentially gain compensated to restore the voxel intensity of the deep region, and the physically enhanced image data is obtained.

[0078] This embodiment is a further embodiment of the above-mentioned inverse light compensation step, aiming to mathematically strictly reverse the physical degradation process of light transmission; the system calculates the net signal data , so as to remove the interference of ambient light and non-specific staining; the local attenuation coefficient matrix is numerically integrated along the depth direction z-axis to generate the cumulative attenuation factor ; The source is the integral of the local attenuation coefficient along the optical path, and the physical meaning is the total attenuation amount experienced by the light from the sample surface to the depth z, with the unit of dimensionless factor; on this basis, the system compensates the net signal data with exponential gain, and the calculation formula is , so as to obtain the physical enhancement image data; this step is not blindly amplifying the brightness, but specifically compensating for the exponentially lost photon energy due to the increase in depth;

[0079] This embodiment, through the inverse compensation mechanism based on the physical model, enables the deep cells in the organoid to have a similar brightness distribution to the shallow cells in the enhanced image in the organoid deep imaging scene; this processing greatly reduces the sensitivity of the subsequent segmentation network to depth changes, realizes the depth invariance of feature extraction, and ensures the consistency and robustness of the segmentation algorithm under different imaging depths.

[0080] Embodiment 3

[0081] The topological repulsive potential field in the multi-dimensional topological feature field is generated in the following way:

[0082] Identify the local extreme value region in the physical enhancement image data;

[0083] Construct a repulsive force field based on the distance between the local extreme value regions;

[0084] In the repulsive force field, define the equipotential surface with zero potential as the segmentation boundary, and the voxel position corresponding to the equipotential surface is the contact interface of the adjacent organoid target.

[0085] This embodiment details the generation mechanism of the topological repulsive potential field, aiming to build an invisible wall between adjacent targets; the system identifies the local extreme value region representing each cell core in the physical enhancement image or target probability map; for any point in space, calculate the distance to the nearest two extreme points, and construct a repulsive force field based on the distance difference; specifically, let the space point The Euclidean distance to the nearest extreme point is , and the Euclidean distance to the second nearest extreme point is , and the repulsive potential is defined as the signed distance function of the distance difference:

[0086] ;

[0087] Under this definition, when a point is located at the midpoint of the centers of two cells, i.e. , the potential energy is zero, forming a zero potential energy surface; the repulsive force field is defined as the negative gradient of the potential energy field:

[0088] ;

[0089] The force field is opposite on both sides of the contact surface, thereby forming a separation potential barrier; in the repulsive force field, the potential energy ridge line or zero potential energy surface is defined as the dividing boundary, even if the two cell gray scales are continuous on the physical image, they are separated by a high energy barrier in the topological potential energy space.

[0090] The training process of the topological prediction model includes:

[0091] A composite loss function is constructed including physical consistency loss and geometric topology loss;

[0092] The physical consistency loss is used to constrain the generation of light transmission attenuation atlas, to ensure that the physically enhanced image data conforms to the law of light transmission physics;

[0093] The geometric topology loss is used to constrain the generation of centripetal vector field and topological repulsive potential field, to ensure the continuity of three-dimensional topological structure under sparse annotation conditions.

[0094] The embodiment further limits the training strategy of the model; a composite loss function is constructed ; wherein, and are the hyperparameter weight factors of the physical loss term and the geometric loss term respectively; the physical consistency loss is constructed according to the physical constraint of Beer-Lambert law, and the specific formula is:

[0095] ;

[0096] Wherein, is the input image, is the integral of the attenuation coefficient along the light path output by the network, is the background field output by the network, and the loss function forces the parameters decomposed by the network to satisfy the physical reconstruction equation, preventing the model from producing hallucinated textures; it is worth noting that the physical consistency loss constructs a self-supervised learning mechanism, so that the network can complete the training without obtaining the true labels of the background light field and the attenuation coefficient ; at the same time, the geometric topology loss is introduced to constrain the centripetal vector field and the topological repulsive potential field, and the combination of cosine similarity loss, smoothness constraint and potential field regression loss is adopted:

[0097] ;

[0098] wherein, denotes the total number of voxels in the sampling space denotes the current voxel index, denotes the current voxel index, denotes norm; is a vector field predicted by the model, is a ground truth vector field, is a predicted topological repulsive potential field, is a ground truth potential field calculated based on the true distance transform; wherein the ground truth vector field and the ground truth potential field are pre-generated by calculating the Euclidean distance transform and its gradient field, respectively, based on the artificially labeled organoid binary mask; is a vector field smoothing regularization coefficient, is a potential field constraint coefficient;

[0099] In the specific experimental setup of the present embodiment, in order to ensure the effectiveness of the physical constraint and take into account the learning of geometric features, the weight factor is preferably set to 10.0, is set to 1.0, and the regularization coefficient is set to , is set to 1.0; in response to the condition of sparse labeling, the loss function strengthens the spatial continuity between voxels, penalizes the main discrete points in the vector field, and forces the potential field to maintain the monotonicity of geometric distance in the unlabeled region, ensuring the continuity of the topological structure in three-dimensional space;

[0100] The present embodiment can still accurately delineate the boundary according to the geometric logic of cell center repulsion in the extremely blurred or even edge invisible region of the organoid image by introducing the geometric positional relationship instead of simply relying on the topological repulsive field of image gradient; in combination with the joint constraint of physical consistency and geometric topology loss, the zero-gradient contact surface separation problem is effectively solved, and the integrity of the three-dimensional topological structure is ensured in the case of scarce labeled data.

[0101] Embodiment 4

[0102] The method for instance boundary evolution of the organoid target comprises:

[0103] calculating a signal-to-noise ratio index of the organoid target at the current depth position;

[0104] adjusting a weight coefficient of the geometric rigidity constraint based on the signal-to-noise ratio index;

[0105] is configured to: when the signal-to-noise ratio index is lower than a preset threshold, the weight coefficient is increased to enhance the shape constraint force of the organoid target tending to a regular sphere.

[0106] The contour position of the organoid target is iteratively updated in combination with the aggregation force provided by the centripetal vector field, the separation force provided by the topological repulsive potential field, and the shape-preserving force provided by the geometric rigidity constraint.

[0107] The embodiment discloses a deep adaptive geometric rigidity constraint evolution strategy; the system calculates the local signal-to-noise ratio index of the organoid target at the current depth position ; The source is local statistical analysis of the image, and the physical meaning is the ratio of the signal quality to the noise level of the current depth slice; specifically, the calculation formula of the signal-to-noise ratio index

[0108] ;

[0109] wherein, is the average gray value of the foreground region voxel after Otsu threshold segmentation in the current depth slice , is the average gray value of the background region voxel, is the standard deviation of the gray value of the background region voxel, is a small constant to prevent the denominator from being zero, and the value is ;

[0110] The system establishes a negative correlation between the geometric regularization term weight in the evolution equation and the signal-to-noise ratio; in response to the signal-to-noise ratio index being lower than a preset threshold, that is, it is determined that it is a deep high-noise area, and the system automatically increases the weight coefficient; the contour position is iteratively updated using the level set function, and the specific dynamic evolution equation is constructed as the level set PDE:

[0111] ;

[0112] wherein, is the time step of evolution iteration, characterizes the evolution speed of the contour over time; is the level set function, is the average curvature, The calculation formula is:

[0113] ;

[0114] is a deep adaptive weight coefficient, wherein, is defined as a normalized unit vector field, and the calculation formula is:

[0115] ;

[0116] wherein, ​​​is the index of the current voxel, is the coordinate of the nearest target center to the voxel, denotes the Euclidean norm, is a small constant to prevent division by zero error; similarly, the topological repulsive force field also needs to be normalized, that is, to ensure the dimensional consistency of the dot product operation with the level set normal ; is the aggregation velocity provided by the centripetal vector field, is the separation force provided by the topological repulsive potential field, and are the convection term weight coefficients of the centripetal force field and the repulsive force field, respectively; to ensure the convergence and separation effect of evolution, in specific embodiments, is set to drive the main contour to converge, and is set to provide a moderate separation repulsive force on the contact surface;

[0117] It should be noted that the weight coefficients , and play a role in balancing the physical dimensions in numerical implementation; among them, the curvature flow term is of diffusion nature and related to the second-order spatial derivative, while the convection term is related to the first-order spatial derivative; in order to ensure the self-consistency of the dynamic equation in physics, the weight coefficient implies the physical meaning of the diffusion coefficient, thereby balancing the dimensional difference between the time variation rate on the left side of the equation and the geometric flow term on the right side, ensuring the stability of numerical evolution; this process combines the balance of three forces: the aggregation force provided by the centripetal vector field, the separation force provided by the topological repulsive potential field, and the shape-preserving force provided by the geometric rigidity constraint.

[0118] The weight coefficient of the adaptive adjustment of the geometric rigidity constraint follows the following rules:

[0119] A nonlinear mapping relationship between the depth coordinate and the weight coefficient is established;

[0120] The weight coefficient monotonically increases with the increase of the depth coordinate;

[0121] where the geometric rigidity constraint is realized by minimizing the gradient modulus of the level set function.

[0122] This embodiment is a further quantification of the above weight adjustment mechanism; a negative correlation mapping relationship between the weight coefficient and the signal-to-noise ratio index is established to realize depth adaptive adjustment, and the calculation formula is as follows:

[0123]

[0124] wherein, is the base geometry constraint weight, is the gain coefficient, is the signal-to-noise ratio sensitivity coefficient, in the preferred setting of the present embodiment, the value range of is 0.1 to 1.0, to control the smoothness of the weight change with the signal-to-noise ratio; the physical meaning of the formula is: when the signal-to-noise ratio is high, the shallow region, the exponential term tends to 0, and the evolution process is mainly driven by the image data term; when the signal-to-noise ratio decreases, the deep region, the weight increases exponentially, forcing to enhance the geometric rigidity constraint to maintain the spherical shape; in this process, the geometric rigidity constraint is realized by minimizing the gradient modulus of the level set function, i.e. the average curvature flow term ; when the weight increases in the deep layer, the curvature flow term in the PDE equation dominates the evolution process, and the algorithm mainly acts to forcibly smooth the burrs of the contour, so that the segmentation result tends to a smooth sphere or ellipsoid;

[0125] The present embodiment adopts a dynamic balance strategy, allowing the segmentation result to present irregular real biological morphology in the shallow layer where the image is clear, and automatically switching to a strong prior mode in the deep layer where the signal-to-noise ratio is extremely low; this counterintuitive design relies on the geometric assumption that the target should be spherical to resist noise and prevent the contour from being broken due to deep spot noise, significantly improving the segmentation robustness in the full depth range.

[0126] Embodiment 5

[0127] Referring to Figure 2 , a three-dimensional image enhancement segmentation system for organoids, comprising:

[0128] a data acquisition module for acquiring initial three-dimensional image data containing an organoid target;

[0129] a physical modeling module for performing light field analysis on the initial three-dimensional image data using a deep learning model to construct a light transmission attenuation atlas;

[0130] a reverse enhancement module for performing voxel-level reverse light compensation on the initial three-dimensional image data based on the light transmission attenuation atlas to generate physical enhancement image data;

[0131] a feature prediction module for inputting the physical enhancement image data into a preset topological prediction model to extract a multi-dimensional topological feature field, the multi-dimensional topological feature field including a target probability map representing the probability of the existence of the target, a centripetal vector field representing the direction of the target center, and a topological repulsive potential field representing the separation barrier between targets;

[0132] The topological segmentation module is configured to construct a dynamic evolution equation according to the centripetal vector field and the topological repulsive potential field, perform instance boundary evolution on the organoid target, and generate a three-dimensional segmentation result.

[0133] The embodiment provides a system architecture based on computer hardware; the system comprises the following core components: a data acquisition module connected to an optical microscope interface, configured to cache original three-dimensional data of high-throughput scanning in real time; a physical modeling module with a pre-trained neural network, such as a 3DU-Net variant, configured to construct and store background light fields and attenuation coefficient tensors in GPU memory; an inverse enhancement module calling a high-precision floating-point operation unit to perform voxel-level subtraction and exponential operation to generate physically enhanced image data; a feature prediction module running a topological prediction model, using a multi-channel output architecture to output probability maps, vector fields and potential fields in parallel; and a topological segmentation module as a core computing unit, running a level set evolution algorithm based on partial differential equations, with a depth-weight mapping table in memory for dynamically calling adaptive weights during evolution, and finally outputting a three-dimensional segmentation result with instance ID.

[0134] The embodiment closely integrates physical signal recovery and geometric topological segmentation through pipeline cooperation of hardware modules; the output of the physical modeling module is directly used as the input of the inverse enhancement module, ensuring the authenticity of data flow recovery at the physical level, and cooperating with dynamic evolution calculation of the topological segmentation module to realize high-precision and automated analysis in the imaging scene of complex biological tissues.

[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application rather than limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for enhancing and segmenting three-dimensional images of organoids, characterized in that, include: Acquire initial 3D image data containing organoid targets; The initial three-dimensional image data is analyzed using a deep learning model to construct an optical transmission attenuation map. The optical transmission attenuation map includes background light field data characterizing the background illumination intensity and a local attenuation coefficient matrix characterizing the light transmission properties of the medium. Based on the optical transmission attenuation spectrum, voxel-level inverse illumination compensation is performed on the initial three-dimensional image data to generate physically enhanced image data. The physical enhancement image data is input into a preset topology prediction model to extract a multidimensional topology feature field. The multidimensional topology feature field includes a target probability map representing the probability of the target's existence, a centripetal vector field representing the direction pointing towards the target's center, and a topological repulsion potential field representing the separation barrier between targets. A dynamic evolution equation is constructed based on the centripetal vector field and the repulsive force field originating from the topological repulsive potential field. The instance boundary evolution is performed on the organoid target to generate a three-dimensional segmentation result. The topological repulsion potential field is used to define the segmentation boundary between different organoid targets in regions where the gradient magnitude is lower than a preset threshold.

2. The organoid three-dimensional image enhancement and segmentation method according to claim 1, characterized in that, The method for constructing the optical transmission attenuation map includes: The initial 3D image data is input into a convolutional neural network; Output background light field distribution data with the same size as the initial three-dimensional image data; Outputs a local attenuation coefficient matrix that characterizes the rate attenuation of light along the depth direction within biological tissue; The background light field distribution data and the local attenuation coefficient matrix are combined to form the optical transmission attenuation map.

3. The organoid three-dimensional image enhancement and segmentation method according to claim 2, characterized in that, The method for generating physically enhanced image data includes: Subtract the background light field distribution data from the initial three-dimensional image data to obtain the net signal data; The local attenuation coefficient matrix is ​​integrated along the depth direction to generate the cumulative attenuation factor; Based on the cumulative attenuation factor, exponential gain compensation is performed on the net signal data to restore the voxel intensity in the deep region, thereby obtaining the physically enhanced image data.

4. The organoid three-dimensional image enhancement and segmentation method according to claim 1, characterized in that, The topological repulsive potential field in the multidimensional topological feature field is generated in the following manner: Identify local extrema regions in the physically enhanced image data; A repulsive force field is constructed based on the distance between the local extreme regions; In the repulsive force field, an equipotential surface with zero potential energy is defined as the dividing boundary, and the voxel position corresponding to the equipotential surface is the contact interface between adjacent organoid targets.

5. The organoid three-dimensional image enhancement and segmentation method according to claim 1, characterized in that, The method for performing instance boundary evolution on the organoid target includes: Calculate the signal-to-noise ratio of the organoid target at the current depth position; Based on the signal-to-noise ratio index, adjust the weighting coefficients of the geometric rigidity constraints; The configuration is as follows: when the signal-to-noise ratio index is lower than a preset threshold, the weighting coefficient is increased to enhance the shape constraint of the organoid target towards a regular sphere; The contour position of the organoid target is iteratively updated by combining the cohesive force provided by the centripetal vector field, the separation force provided by the topological repulsive potential energy field, and the shape-preserving force provided by the geometric rigidity constraint.

6. The organoid three-dimensional image enhancement and segmentation method according to claim 5, characterized in that, The weighting coefficients of geometrically rigid constraints are adjusted according to the following rules: Establish a nonlinear mapping relationship between depth coordinates and the weighting coefficients; The weighting coefficient increases monotonically as the depth coordinates increase; The geometric rigidity constraint is achieved by minimizing the gradient magnitude of the level set function.

7. The organoid three-dimensional image enhancement and segmentation method according to claim 1, characterized in that, The training process of the topology prediction model includes: Construct a composite loss function that includes physical consistency loss and geometric topology loss; The physical consistency loss is used to constrain the generation of the optical transmission attenuation map, ensuring that the physically enhanced image data conforms to the physical laws of optical transmission. The generation of the centripetal vector field and the topological repulsion potential field are constrained by the geometric topological loss, ensuring the continuity of the three-dimensional topological structure under sparse labeling conditions.

8. A three-dimensional image enhancement and segmentation system for organoids, used to implement the three-dimensional image enhancement and segmentation method for organoids as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire initial three-dimensional image data containing organoid targets; The physical modeling module is used to perform light field analysis on the initial three-dimensional image data using a deep learning model and construct an optical transmission attenuation map. The inverse enhancement module is used to perform voxel-level inverse illumination compensation on the initial three-dimensional image data based on the optical transmission attenuation spectrum, and generate physically enhanced image data. The feature prediction module is used to input the physical enhancement image data into a preset topology prediction model and extract a multidimensional topology feature field. The multidimensional topology feature field includes a target probability map representing the probability of the existence of the target, a centripetal vector field representing the direction pointing to the center of the target, and a topological repulsion potential field representing the separation barrier between targets. The topology segmentation module is used to construct a dynamic evolution equation based on the centripetal vector field and the repulsive force field originating from the topological repulsive potential energy field, perform instance boundary evolution on the organoid target, and generate a three-dimensional segmentation result.

Images