A three-dimensional spatial distribution reconstruction method of an in-vitro three-dimensional biological model

Abstract

The present disclosure relates to the field of three-dimensional spatial distribution reconstruction of in-vitro three-dimensional biological models, and provides a three-dimensional spatial distribution reconstruction method of in-vitro three-dimensional biological models, comprising: collecting multiple Z-axis layer scanning images of different focal planes of an in-vitro three-dimensional biological model culture sample; inputting the Z-axis layer scanning images into a target detection network for in-vitro three-dimensional biological model detection to obtain candidate targets; respectively calculating spatial blur scores and frequency domain sharpness scores thereof, and fusing the scores into a comprehensive sharpness score according to an adaptive weight, and determining candidate targets meeting adaptive score threshold requirements as effective candidate targets; and reconstructing the three-dimensional spatial distribution of the in-vitro three-dimensional biological model according to a processing result through cross-layer deduplication and vertical stacking classification processing. The present disclosure automatically reconstructs the three-dimensional spatial position distribution of in-vitro three-dimensional biological models without the need for fluorescent labeling, realizes adaptive optimal focal plane screening, and can correctly identify and retain different in-vitro three-dimensional biological models stacked in the vertical direction.

Description

Technical Field

[0001] This disclosure relates to the field of three-dimensional spatial distribution reconstruction technology of in vitro three-dimensional biological models, and particularly to a method for three-dimensional spatial distribution reconstruction of in vitro three-dimensional biological models. Background Technology

[0002] Organoids are miniature tissues that self-organize under in vitro three-dimensional culture conditions and can be induced from stem cells. Organoids can simulate the key structural and functional characteristics of real organs and have significant application value in drug screening, disease modeling, toxicology evaluation, and regenerative medicine. In addition to organoids, in vitro three-dimensional biological models also include all biological entities that can grow in suspension under in vitro three-dimensional culture conditions and possess recognizable morphological characteristics, such as, but not limited to, three-dimensional spheroids and 3D cell aggregates formed by single cells or multiple cell aggregates. Compared to traditional two-dimensional cell culture models, in vitro three-dimensional biological models better reflect the three-dimensional microenvironment, intercellular interactions, and differentiation processes of in vivo tissues. In recent years, the culture technology of in vitro three-dimensional biological models has gradually developed towards standardization, high throughput, and automation. The demand for accurate measurement of key indicators such as the number of in vitro three-dimensional biological models, growth status tracking, and quantitative drug sensitivity responses during the culture process is becoming increasingly urgent.

[0003] In the culture system of in vitro three-dimensional biological models, because the models grow in three-dimensional suspension within the culture medium, accurate measurement of their spatial location and diameter is crucial for downstream analysis. In vitro three-dimensional biological models exhibit not only distribution in the XY plane but also depth differences along the Z-axis in the three-dimensional culture system. Images of in vitro three-dimensional biological models at different focal planes can capture information about the models at different depths, such as... Figure 1 As shown, organoid images at different focal planes contain organoid information at different depths. Based on this, how to automatically and accurately reconstruct the true distribution of in vitro three-dimensional biological models in three-dimensional space from multi-focal plane images of bright-field microscopy has gradually become an urgent need and an important technical challenge in this field.

[0004] Currently, several methods have emerged in the existing technology for reconstructing the true distribution of in vitro 3D biological models in 3D space based on multi-focal plane images. Some methods employ fluorescence confocal microscopy imaging, which uses fluorescently labeled images to perform 3D confocal scanning of the in vitro 3D biological model, obtaining high-resolution 3D structural information. However, this method requires fluorescence staining, is complex and costly, and the phototoxicity of the fluorescent dyes may interfere with the normal growth and drug response of the in vitro 3D biological model, making it unsuitable for long-term dynamic monitoring of live cells. Other methods combine single-layer bright-field imaging with traditional image segmentation techniques, but these only capture bright-field images at a fixed focal plane, failing to acquire the Z-axis depth information of the in vitro 3D biological model. This makes it unable to handle the distribution differences of the in vitro 3D biological model across different focal planes. Furthermore, these methods also utilize traditional image segmentation methods such as thresholding and watershed algorithms to identify the in vitro 3D biological model, but traditional image segmentation methods are sensitive to imaging conditions and have poor robustness. Some methods employ multi-focal plane synthesis, combining images from multiple focal planes into a single fully focused image before image segmentation. While this improves the sharpness range of a single image, it loses Z-axis depth information during image synthesis, failing to reflect the true distribution of the in vitro 3D biological model in 3D space and unable to address the overlap problem of the in vitro 3D biological model at different depths. Other methods utilize deep learning-based single-frame object detection techniques, using models such as YOLO and Faster R-CNN to detect in vitro 3D biological models on a single bright-field image. This method is also limited to a single focal plane, unable to address the overlap and occlusion problems of in vitro 3D biological models at different depths. Furthermore, if multiple Z-axis tomographic images are used for in vitro 3D biological model detection, the same in vitro 3D biological model may be detected repeatedly in images with different focal planes, leading to counting bias and positional inconsistencies. However, this method lacks an effective cross-layer deduplication mechanism and cannot distinguish the vertical stacking of in vitro 3D biological models. Summary of the Invention

[0005] This disclosure aims to address at least one of the problems existing in the prior art by providing a method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model.

[0006] This disclosure provides a method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model, including: Z-axis tomography was performed on cultured samples of an in vitro three-dimensional biological model using a bright-field microscope, and multiple Z-axis tomography images with different focal planes were acquired. Each Z-axis tomographic image is input into a preset target detection network for in vitro three-dimensional biological model detection to obtain the corresponding candidate target. For each candidate target, a spatial ambiguity score and a frequency domain sharpness score are calculated respectively. The spatial ambiguity score and the frequency domain sharpness score are then merged into a comprehensive sharpness score according to an adaptive weight. Candidate targets whose comprehensive sharpness scores meet the adaptive score threshold requirements are determined as valid candidate targets. Each of the effective candidate targets is subjected to cross-layer deduplication and vertical stacking classification to obtain a unique set of in vitro three-dimensional biological models. The three-dimensional spatial distribution of the in vitro three-dimensional biological model is reconstructed based on the unique set of in vitro three-dimensional biological models.

[0007] Optionally, spatial ambiguity score and frequency domain sharpness score are calculated for each of the candidate targets, including: For each candidate target, the corresponding ROI sub-image is cropped from the Z-axis layer scan image in which it is located; Apply a Laplacian kernel of a preset size to the grayscale image of each ROI sub-image to obtain the corresponding response image; The variance of each response map is calculated as the spatial ambiguity score of the corresponding candidate target; Perform a two-dimensional Fourier transform on the grayscale image of each ROI sub-image and shift the zero frequency to the center. Calculate the power spectrum and count the proportion of high-frequency energy above the defined cutoff frequency as the frequency domain sharpness score of the corresponding candidate target.

[0008] Optionally, the spatial ambiguity score and the frequency domain sharpness score are fused into a comprehensive sharpness score using adaptive weights, including: The global statistical features of each Z-axis layer scan image are statistically analyzed. The frequency domain weights are dynamically adjusted according to preset rules based on the global statistical features. The spatial ambiguity score and the frequency domain sharpness score are normalized. The spatial ambiguity score and the frequency domain sharpness score are weighted and fused according to the following formula: ; in, The overall sharpness score is given as follows: The normalized spatial ambiguity score is the score of the spatial domain. The normalized frequency domain sharpness score is... The frequency domain weights, Spatial weights and satisfying .

[0009] Optionally, the global statistical features include the average contrast of the entire image, the average noise of the entire image, and the average high-frequency energy ratio of the entire image. Dynamically adjust the frequency domain weights according to the global statistical features and preset rules, including: Define the base value and range of the frequency domain weights; The frequency domain weights are dynamically adjusted according to the following preset rules: if the contrast of the candidate target is lower than the average contrast of the entire Z-axis tomographic image in which it is located, the frequency domain weights are appropriately reduced from the base value; if the noise of the candidate target is higher than the average noise of the entire Z-axis tomographic image in which it is located, or if the high-frequency energy ratio of the candidate target is higher than the average high-frequency energy ratio of the entire Z-axis tomographic image in which it is located, the frequency domain weights are appropriately increased from the base value; at the same time, it is ensured that the final value of the frequency domain weights is within the value range.

[0010] Optionally, each of the effective candidate targets undergoes cross-layer deduplication and vertical stacking classification to obtain a unique set of in vitro three-dimensional biological models, including: A Z-axis index and Z-axis depth are attached to each of the valid candidate targets, and the valid candidate targets are sorted in descending order according to the comprehensive sharpness score; Based on the two-dimensional spatial intersection-union ratio between any two valid candidate targets, suspected duplicate pairs are filtered out. Calculate the Z-axis physical distance between the valid candidate targets in each suspected duplicate pair and determine the corresponding Z-axis threshold; Based on the Z-axis physical distance and the corresponding Z-axis threshold, suspected height pairs are filtered out; Expand the effective candidate targets included in the highly suspected pair into corresponding 3D bounding boxes, calculate the 3D intersection-over-union ratio of the effective candidate targets in the highly suspected pair, and remove the effective candidate targets with lower overall sharpness scores in the highly suspected pair whose 3D intersection-over-union ratio is greater than a preset 3D intersection-over-union ratio threshold; All remaining valid candidate targets are combined into the unique in vitro three-dimensional biological model set.

[0011] Optionally, based on the two-dimensional spatial intersection-union ratio between any two of the valid candidate targets, suspected duplicate pairs are filtered out, including: If the two-dimensional spatial intersection-union ratio between two valid candidate targets is greater than a preset two-dimensional intersection-union ratio threshold, then the two valid candidate targets are identified as the suspected duplicate pair.

[0012] Optionally, based on the Z-axis physical distance and the corresponding Z-axis threshold, suspected height pairs are filtered out, including: If the Z-axis physical distance between the valid candidate targets in the suspected duplicate pair is less than its corresponding Z-axis threshold, then the suspected duplicate pair is determined as the highly suspected pair; The Z-axis threshold is expressed as: ; in, The Z-axis threshold, The median diameter of the valid candidate targets with the higher overall sharpness score among the suspected duplicate pairs. These are preset positive numbers.

[0013] Optionally, reconstructing the three-dimensional spatial distribution of the in vitro three-dimensional biological models based on the unique set of in vitro three-dimensional biological models includes: Calculate the center point coordinates, diameter, Z-axis depth, and volume of each valid candidate target in the unique in vitro three-dimensional biological model set; The pixel coordinates of each valid candidate target are converted into corresponding physical coordinates using a microscope pixel-physical size calibration coefficient. Based on the physical coordinates, each of the valid candidate targets is located and rendered on the Z-axis layer scan image to generate a three-dimensional reconstruction result.

[0014] Optionally, before converting the pixel coordinates of each of the valid candidate targets into their corresponding physical coordinates using a microscope pixel-physical size calibration coefficient, the method for reconstructing the three-dimensional spatial distribution of the in vitro three-dimensional biological model further includes: The center point coordinates, diameter, and Z-axis index of the valid candidate targets are verified, and abnormal values ​​are recorded and warnings are issued.

[0015] Optionally, when inputting each of the Z-axis tomographic images into a preset target detection network for in vitro three-dimensional biological model detection, an inference engine is used for optimization.

[0016] Compared with the prior art, this disclosure has the following technical effects: 1. By performing layer-by-layer detection on multi-focal plane image sequences acquired by bright-field microscopy and associating them with Z-axis coordinate information, the three-dimensional spatial distribution of in vitro three-dimensional biological models can be automatically reconstructed without the need for fluorescent labeling. This significantly reduces experimental costs and operational complexity, solves the problem of "missing Z-axis information" in existing methods, and realizes the reconstruction of the three-dimensional spatial distribution of in vitro three-dimensional biological models based on bright-field Z-axis tomography. 2. The detection results of all Z-axis layer images are aggregated into a global candidate set. The comprehensive sharpness score is introduced as the sorting criterion. A three-level cascaded cross-layer deduplication mechanism is established. The three-level cascaded pipeline of "2D IoU coarse screening → Z-axis distance fine screening → 3D IoU final judgment" achieves accurate deduplication and solves the counting deviation problem caused by "cross-layer repeated detection". 3. By adopting a candidate target quality assessment method based on weighted fusion of ambiguity score (spatial domain) and frequency domain sharpness score (frequency domain), and dynamically and adaptively calculating weights according to the current global statistical features of the image, the detection results of each in vitro three-dimensional biological model on the optimal focal plane are automatically retained, thus solving the problem of "lack of automated standards for focal plane selection" and realizing adaptive optimal focal plane screening. 4. By using only Z-axis tomographic images from a bright-field microscope, combined with a deep learning target detection model and inference engine acceleration, high-precision, high-throughput in vitro three-dimensional biological model detection and three-dimensional spatial distribution reconstruction under fluorescent label-free conditions were achieved, solving the problem of "fluorescent label dependence" and realizing high-precision detection under bright-field label-free conditions. 5. The Z-axis distance constraint and 3D IoU discrimination mechanism are introduced. In the cross-layer deduplication, the Z-axis distance constraint (second-level fine screening) and 3D IoU final judgment (third level) are introduced, which can correctly identify and retain different in vitro three-dimensional biological models stacked vertically, and solve the edge scene problem of "mis-merging of vertically stacked in vitro three-dimensional biological models". Attached Figure Description

[0017] One or more embodiments are illustrated by way of example with the corresponding pictures in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0018] Figure 1 The organoid images with different focal planes provided in the background art; Figure 2 A flowchart illustrating a method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model according to an embodiment of this disclosure; Figure 3 A flowchart illustrating a method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model, as provided in another embodiment of this disclosure; Figure 4 This is a schematic diagram illustrating the process and effect of a method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model, provided as another embodiment of this disclosure. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the various embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this disclosure to facilitate a better understanding of the disclosure. However, the technical solutions claimed in this disclosure can be implemented even without these technical details and with various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of this disclosure. The various embodiments can be combined with and referenced by each other without contradiction.

[0020] One embodiment of this disclosure relates to a method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model. This method reconstructs the three-dimensional spatial distribution based on adaptive sharpness scoring, cross-layer deduplication, and vertical stacking classification. The specific process is as follows: Figure 2 As shown, it includes steps S110 to S150.

[0021] Step S110: Perform Z-axis tomography imaging on the cultured sample of the in vitro three-dimensional biological model using a bright-field microscope, and acquire multiple Z-axis tomography images with different focal planes.

[0022] Specifically, step S110 is mainly used for image acquisition. For example, step S110 can use a bright-field microscope to perform Z-axis tomographic imaging of the cultured sample of an in vitro three-dimensional biological model, with a fixed tomographic step size along the optical axis. A bright-field image sequence was created by scanning multiple Z-axis tomographic images at different focal planes from the bottom of the culture medium to the top of the liquid surface of the culture sample in an in vitro three-dimensional biological model. The tomographic step size was specified. It can be set to , The depth of field along the Z-axis is used to avoid information gaps between adjacent focal planes. The number of Z-axis layer scan images. Set as , This represents the height from the bottom of the culture medium to the top of the liquid surface. For the i-th Z-axis tomographic image... Its Z-axis depth for ,in, This represents the Z-axis position at the bottom of the culture medium.

[0023] In particular, during the acquisition of multiple Z-axis tomographic images with different focal planes, the illumination and exposure parameters are kept constant to avoid interference caused by environmental changes.

[0024] It should be noted that the term "in vitro three-dimensional biological model" in the three-dimensional spatial distribution reconstruction method of in vitro three-dimensional biological models provided in this disclosure should be interpreted broadly. It includes not only "organoids," i.e., strictly speaking, miniature tissues that self-organize under in vitro three-dimensional culture conditions, such as those induced from stem cells, but also all biological entities capable of growing in suspension under in vitro three-dimensional culture conditions and possessing identifiable morphological characteristics. These may include, but are not limited to, three-dimensional spheroids and 3D cell aggregates formed by single or multiple cells. Like "organoids," these biological entities exhibit depth differences along the Z-axis, facing common problems such as cross-layer duplicate detection and vertical stacking blurring. Therefore, the three-dimensional spatial distribution reconstruction method of in vitro three-dimensional biological models based on adaptive sharpness scoring and cross-layer deduplication and vertical stacking classification provided in this disclosure is applicable not only to the three-dimensional spatial distribution reconstruction of organoids but also to the three-dimensional spatial distribution reconstruction of the aforementioned biological entities, without requiring substantial adjustments to its key steps.

[0025] Step S120: Input each Z-axis layer scan image into a preset target detection network to detect the in vitro three-dimensional biological model and obtain the corresponding candidate targets.

[0026] Specifically, step S120 is mainly used to perform layer-by-layer deep learning target detection, inputting all the acquired Z-axis layer scan images one by one into a preset target detection network, and obtaining the in vitro three-dimensional biological model target output by the target detection network as a candidate target.

[0027] The preset object detection network can be any deep learning network model capable of object detection in images. For example, the preset object detection network can be any one of YOLOv8, YOLOv11, YOLOX, RTMDet, DETR, etc. Preferably, the preset object detection network adopts the YOLOv11 architecture, which uses the CSPDarknet backbone and PAN feature pyramid for multi-scale feature fusion to achieve object detection. In particular, when each Z-axis layer scan image is input into the preset object detection network for in vitro 3D biological model detection, an inference engine can be used for optimization to accelerate high-throughput, low-latency batch processing and meet the real-time requirements of large-scale in vitro 3D biological model culture and screening scenarios. Those skilled in the art can choose the specific type of inference engine according to actual needs, for example, any one of TensorRT, ONNX Runtime, OpenVINO, ncnn, etc.

[0028] Specifically, after obtaining the candidate targets in each Z-axis layer scan image output by the preset target detection network, single-layer non-maximum suppression (NMS) processing can be performed on each Z-axis layer scan image to suppress candidate targets with a cross-union ratio greater than a preset suppression threshold such as 0.5, and to discard candidate targets with a confidence level less than a preset confidence threshold such as 0.25, so as to filter the detection results output by the preset target detection network and improve the detection accuracy.

[0029] Step S130: Calculate the spatial ambiguity score and frequency domain sharpness score for each candidate target, and merge the spatial ambiguity score and frequency domain sharpness score into a comprehensive sharpness score according to adaptive weights. Determine the candidate targets whose comprehensive sharpness scores meet the adaptive score threshold requirements as valid candidate targets.

[0030] Specifically, step S130 is mainly used to select effective candidate targets based on the comprehensive sharpness score, based on the candidate targets.

[0031] For example, in step S130, the spatial ambiguity score and frequency domain sharpness score are calculated for each candidate target, including: for each candidate target, the corresponding ROI sub-image is extracted from the Z-axis layer scan image where it is located; a Laplacian kernel of a preset size is applied to the grayscale image of each ROI sub-image to obtain the corresponding response image; the variance of each response image is calculated as the spatial ambiguity score of its corresponding candidate target; a two-dimensional Fourier transform is performed on the grayscale image of each ROI sub-image and the zero frequency is shifted to the center, the power spectrum is calculated, and the proportion of high-frequency energy above the defined cutoff frequency is counted as the frequency domain sharpness score of the corresponding candidate target.

[0032] Specifically, the candidate targets output by the preset object detection network are usually represented as bounding boxes. Therefore, when cropping ROI sub-images, the bounding boxes of each candidate target can be directly extracted from the Z-axis layer scan image where they are located as the corresponding ROI sub-image. If the ROI sub-image is in RGB format, it needs to be converted to grayscale to eliminate interference from color channel differences. If the size of the ROI sub-image is smaller than a preset pixel threshold, such as 3 pixels, it can be discarded directly to improve detection accuracy.

[0033] When calculating the spatial ambiguity score, a Laplacian kernel of a preset size (e.g., 3×3) is first applied to the grayscale image of each ROI sub-image to obtain a response image. The variance of the response image is then calculated as the spatial ambiguity score, which serves as an indicator of edge sharpness. A larger variance in the response image indicates a sharper corresponding ROI sub-image. Spatial Ambiguity Score Represented as ,in, This represents the response values ​​of all elements in the response graph. This represents the total number of elements in the response graph. This represents the mean response of all elements in the response graph.

[0034] When calculating the frequency domain sharpness score, a two-dimensional Fourier transform (2D-FFT) is performed on the grayscale image of each ROI sub-image, and the zero frequency is shifted to the center to calculate the power spectrum. ,in, Coordinates in the frequency domain The power level at that location, The matrix is ​​a complex number after undergoing a two-dimensional Fourier transform and shifting to its center from zero frequency. The proportion of high-frequency energy above a defined cutoff frequency is used as the frequency domain sharpness score for the corresponding candidate target, which is then used as an indicator of detail richness. The defined cutoff frequency... It can be , The width of the ROI sub-image. The height of the ROI sub-image. This is the proportionality coefficient, which can be set to... Frequency domain resolution score It can be represented as , High-frequency energy above the cutoff frequency. This represents the total energy above the cutoff frequency.

[0035] For example, in step S130, the spatial ambiguity score and the frequency domain sharpness score are fused into a comprehensive sharpness score according to adaptive weights, including: separately calculating the global statistical features of each Z-axis layer scan image; dynamically adjusting the frequency domain weights according to preset rules based on the global statistical features; normalizing the spatial ambiguity score and the frequency domain sharpness score; and weighting and fusing the spatial ambiguity score and the frequency domain sharpness score according to the following formula: ; in, To score overall clarity, The normalized spatial ambiguity score. The normalized frequency domain resolution score. For frequency domain weights, Spatial weights and satisfying .

[0036] For example, global statistical features include the overall image average contrast, overall image average noise, and overall image average high-frequency energy percentage. When calculating the global statistical features of each Z-axis tomographic image, the overall image average contrast, overall image average noise, and overall image average high-frequency energy percentage are calculated separately for each Z-axis tomographic image. Specifically, the overall average contrast is the mean of the pixel standard deviation of the Z-axis tomographic image; the overall average noise is the mean of the difference variance of the Z-axis tomographic image after median filtering and denoising; and the overall average high-frequency energy percentage is the proportion of high-frequency energy in the Z-axis tomographic image to the total energy of the entire image.

[0037] The frequency domain weights are dynamically adjusted according to global statistical characteristics and preset rules, including: setting a base value for the frequency domain weights, such as 0.6, and a range of values, such as [0.4, 0.8]; dynamically adjusting the frequency domain weights according to the following preset rules: if the contrast of a candidate target is lower than the average contrast of the entire Z-axis tomographic image, the frequency domain weights are appropriately reduced from the base value, such as by 0.1; if the noise of a candidate target is higher than the average noise of the entire Z-axis tomographic image, or if the proportion of high-frequency energy of a candidate target is higher than the average proportion of high-frequency energy of the entire Z-axis tomographic image, the frequency domain weights are appropriately increased from the base value, such as by 0.1; and ensuring that the final value of the frequency domain weights is within the range of values, such as [0.4, 0.8].

[0038] Before weighted fusion of spatial ambiguity scores and frequency domain sharpness scores, the spatial ambiguity scores are first... Frequency domain resolution score Normalization processes, such as Min-Max normalization, were performed separately to eliminate dimensional differences. Then, a comprehensive sharpness score was calculated. The calculation formula for the normalized spatial ambiguity score and normalized frequency domain sharpness score Weighted fusion is performed based on frequency domain weights and spatial domain weights.

[0039] After obtaining the overall sharpness score for each candidate target, those targets with an overall sharpness score lower than the adaptive score threshold are marked as blurry and invalid and discarded. The remaining candidate targets are considered valid candidate targets carrying an overall sharpness score. The adaptive score threshold is... It can be represented as , The average of the overall sharpness scores for all current candidate targets. The standard deviation of the overall sharpness score for all current candidate targets. This is a scaling factor, which can be set to... .

[0040] Step S140: Perform cross-layer deduplication and vertical stacking classification on each valid candidate target to obtain a unique set of in vitro three-dimensional biological models.

[0041] Specifically, step S140 is mainly used to complete cross-layer deduplication and vertical stacking classification processing through three-level cascade deduplication.

[0042] For example, step S140 includes: attaching a Z-axis index and Z-axis depth to each valid candidate target, and sorting the valid candidate targets in descending order according to the overall sharpness score; filtering out suspected duplicate pairs based on the two-dimensional spatial intersection-union ratio between any two valid candidate targets; calculating the Z-axis physical distance between valid candidate targets in each suspected duplicate pair and determining the corresponding Z-axis threshold; filtering out highly suspected pairs based on the Z-axis physical distance and the corresponding Z-axis threshold; expanding the valid candidate targets included in the highly suspected pairs into corresponding three-dimensional bounding boxes, calculating the three-dimensional intersection-union ratio of the valid candidate targets in the highly suspected pairs, and removing the valid candidate targets with lower overall sharpness scores in the highly suspected pairs whose three-dimensional intersection-union ratio is greater than a preset three-dimensional intersection-union ratio threshold; and forming a unique in vitro three-dimensional biological model set from all remaining valid candidate targets.

[0043] Specifically, the Z-axis index of a valid candidate target is the index of the Z-axis layer scan image in which it is located. For example, if a valid candidate target is located in the i-th Z-axis layer scan image, then the Z-axis index of that valid candidate target is... That is, the index i of the i-th Z-axis layer scan image, i.e. Similarly, the Z-axis depth of a valid candidate target is the Z-axis depth of the Z-axis tomographic image in which it is located. For example, if a valid candidate target is located in the i-th Z-axis tomographic image, then the Z-axis depth of that valid candidate target is... That is, the Z-axis depth of the i-th Z-axis tomographic image. ,Right now .

[0044] Specifically, if the number of valid candidate targets satisfy This indicates that there is only one valid candidate target. In this case, there will be no overlap of in vitro 3D biological models at different depths. Therefore, there is no need to perform cross-layer deduplication and vertical stacking classification. This valid candidate target can be directly used as an element in the unique in vitro 3D biological model set.

[0045] For example, filtering out suspected duplicate pairs based on the two-dimensional spatial intersection-union ratio between any two valid candidate targets includes: if the two-dimensional spatial intersection-union ratio between two valid candidate targets is greater than a preset two-dimensional intersection-union ratio threshold, then the two valid candidate targets are identified as suspected duplicate pairs.

[0046] Specifically, the screening of suspected duplicate pairs is achieved by performing a first-level deduplication, namely a two-dimensional spatial intersection-union ratio (2DLU) coarse screening. First, the 2D spatial intersection-union ratio between any two valid candidate targets is calculated. ,in, This represents the area of ​​the intersection of two valid candidate targets in two-dimensional space, i.e., the area of ​​the intersection region. This represents the area of ​​the union of two valid candidate targets in two-dimensional space, which is the sum of the areas of the two valid candidate targets in two-dimensional space minus the area of ​​their intersection in two-dimensional space. Then, starting with the valid candidate target with the highest overall sharpness score, iterative processing is performed to determine whether the two-dimensional intersection-union ratio (IUU) between each valid candidate target and other valid candidate targets is less than or equal to a preset two-dimensional IUU threshold. If yes, the two corresponding valid candidate targets are determined to be different in vitro 3D biological models that do not overlap in the XY plane; otherwise, the two corresponding valid candidate targets are determined to be different in vitro 3D biological models that overlap in the XY plane. These two valid candidate targets are marked as "suspected duplicate pairs" for subsequent Z-axis fine screening. Screening of suspected duplicate pairs requires processing all valid candidate targets, but each pair of valid candidate targets only requires 4 multiplication-addition operations during the screening process, and the computational cost of a single pair of valid candidate targets is extremely small.

[0047] For example, based on the Z-axis physical distance and the corresponding Z-axis threshold, highly suspected pairs are filtered out, including: if the Z-axis physical distance between valid candidate targets in a suspected duplicate pair is less than its corresponding Z-axis threshold, then the suspected duplicate pair is determined as a highly suspected pair, wherein the Z-axis threshold is expressed as: ; in, Z-axis threshold The median diameter of the valid candidate targets with high overall sharpness scores, which are suspected of being duplicates. Preset positive constants and satisfy .

[0048] Specifically, the screening of highly suspected pairs is achieved by performing a second-level deduplication process, namely, Z-axis distance constraint-based fine screening. First, the Z-axis physical distance between the two valid candidate targets is calculated for each suspected duplicate pair. and the corresponding Z-axis threshold Among them, the physical distance along the Z-axis The calculation formula is ,in, The Z-axis depth of one of the valid candidate targets in the suspected duplicate pair. Let Z be the Z-axis depth of another valid candidate target in this suspected duplicate pair. If a suspected duplicate pair satisfies If the two valid candidate targets in the suspected duplicate pair are determined to be vertically stacked different in vitro three-dimensional biological models, then the suspected duplicate pair will not undergo a third-level deduplication process. If a suspected duplicate pair satisfies If the suspected duplicate pair is identified, it is marked as a "highly suspected pair" for further deduplication. The screening process for highly suspected pairs involves only one subtraction operation, one absolute value operation, and one comparison operation for each suspected duplicate segment, resulting in minimal computational overhead.

[0049] The third-level deduplication process is primarily used for cross-layer final determination through 3D intersection-union ratio (IUU) to filter out all unique in vitro 3D biological model targets. In practice, the effective candidate targets within each highly probable pair are first expanded into their corresponding 3D bounding boxes. , can be represented as ,in, These represent the minimum X-axis coordinate and minimum Y-axis coordinate corresponding to the valid candidate targets, respectively. These represent the maximum X-axis coordinate and maximum Y-axis coordinate corresponding to the valid candidate item, with a fixed step size. The value can be , The step size for layer scanning, The diameter of the valid candidate target is given. Next, the 3D intersection-union ratio (IUU) of the valid candidate targets in each highly probable pair is calculated based on the 3D bounding box. , represented as , Let be the intersection volume of the two 3D bounding boxes. Let be the union volume of two 3D bounding boxes. If the 3D intersection-union ratio of a highly probable candidate target in a pair is... Greater than the preset 3D intersection-union ratio threshold Preset 3D intersection-union ratio threshold Set as If the highly suspected pair is determined to be a spatially overlapping in vitro 3D biological model, the candidate target with the lower overall sharpness score is removed. Otherwise, both candidate targets in the highly suspected pair are retained. All remaining candidate targets are then used as the final unique in vitro 3D biological model targets to form a unique in vitro 3D biological model set. While the computational cost of the 3D intersection-union ratio is relatively high compared to the first two deduplication stages, the proportion of highly suspected pairs among all candidate targets is usually small, perhaps only 1% to 5%, thus the impact on overall efficiency is controllable.

[0050] Step S150: Reconstruct the three-dimensional spatial distribution of the in vitro three-dimensional biological model based on the unique set of in vitro three-dimensional biological models.

[0051] For example, step S150 includes: calculating the center point coordinates, diameter, Z-axis depth, and volume of each valid candidate target in the unique in vitro three-dimensional biological model set, wherein the volume can be the volume of a sphere directly fitted based on the diameter of the valid candidate target; converting the pixel coordinates of each valid candidate target into corresponding physical coordinates using a microscope pixel-physical size calibration coefficient; and locating and rendering each valid candidate target on the Z-axis tomographic image based on the physical coordinates to generate a three-dimensional reconstruction result.

[0052] Specifically, the localization data of the in vitro three-dimensional biological model is first calculated, and the center point coordinates, diameter, and Z-axis depth of each valid candidate target in the unique in vitro three-dimensional biological model set are calculated respectively. Among them, the center point coordinates of the valid candidate target ( In the diagram, the X-axis direction coordinates for Y-axis coordinate for Diameter of valid candidate targets for The diameter is determined by taking the maximum value of width and height because in vitro 3D biological models are usually approximately spherical, and their projection is nearly circular. Taking the maximum value of width and height provides a more conservative estimation method that is closer to the true physical size of the diameter. The Z-axis depth of the effective candidate target is... .

[0053] For example, before converting the pixel coordinates of each valid candidate target to the corresponding physical coordinates using the microscope pixel-physical size calibration coefficient, the three-dimensional spatial distribution reconstruction method of the in vitro three-dimensional biological model also includes: verifying the center point coordinates, diameter, and Z-axis index of the valid candidate targets, recording outliers, and issuing warnings.

[0054] Specifically, when verifying the center point coordinates, diameter, and Z-axis index of valid candidate targets, the main checks are whether the diameter of the valid candidate targets is within a reasonable range, such as [5μm, 1000μm], whether the center point coordinates exceed the limit, and whether the Z-axis index is valid. Abnormal values ​​detected are recorded and warnings are issued, but they are retained for manual review to prevent verification deviations.

[0055] When converting the pixel coordinates of each valid candidate target to their corresponding physical coordinates using the microscope pixel-physical size calibration coefficient, each pixel coordinate can be multiplied by the microscope pixel-physical size calibration coefficient to obtain the corresponding physical coordinates. The X-axis coordinates of the center point can then be used as the reference coordinates. Corresponding physical coordinates for The Y-axis coordinate of the center point coordinates Corresponding physical coordinates for ,diameter Corresponding physical coordinates for ,in, This is the microscope pixel-to-physical size calibration factor, in units of .

[0056] During localization rendering, a circular border and center point marker can be drawn on the acquired original bright-field images (multiple Z-axis tomographic images), with the center point of each unique in vitro 3D biological model target as the center and half the diameter as the equivalent radius. These are labeled with a serial number, generating a localized rendering image with annotations. The circular border can be green, and the center point marker can be red for distinction. If different annotations overlap, radial offset processing can be performed, and the portion exceeding the image edge can be cropped when it approaches the image edge.

[0057] To enable those skilled in the art to better understand the above embodiments, a specific example is provided below for illustration.

[0058] This example provides a method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model, the process of which is as follows: Figure 3 As shown, the process includes steps S01 to S14. Steps S01 to S02 belong to the image acquisition and target detection part, steps S03 to S07 belong to the candidate target quality scoring part, steps S08 to S11 belong to the cross-layer global non-maximum suppression, i.e., the three-level cascade deduplication part, and steps S12 to S14 belong to the three-dimensional spatial distribution output and visualization part. Figure 4 Given Figure 3 The flowchart and effect of the three-dimensional spatial distribution reconstruction method for the in vitro three-dimensional biological model are shown below. Figure 3 and Figure 4 This example provides a detailed explanation of the method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model. Specifically, this example selects organoids as the in vitro three-dimensional biological model and reconstructs their three-dimensional spatial distribution.

[0059] Step S01: Bright-field Z-axis tomography acquires multi-focal-plane image sequences for image acquisition and input. Specifically, Z-axis tomography imaging of organoid culture samples is performed using a bright-field microscope, with a fixed tomography step size along the optical axis. Scan from the bottom of the culture medium to the top of the liquid surface to collect data. A sequence of bright-field images composed of Z-axis tomographic images from different focal planes is used as the original image, with a tomographic step size of [missing information]. Must meet To avoid information gaps between adjacent focal planes, among which, The depth of field is along the Z-axis. The i-th Z-axis slice image. Corresponding Z-axis depth During the acquisition process, keep the lighting and exposure parameters constant.

[0060] Step S02: YOLOv11 detects organoid candidate bounding boxes layer by layer to perform layer-by-layer deep learning object detection. Specifically, it will... One Z-axis tomographic image is input into the YOLOv11 object detection network. The output is obtained, where the output of the YOLOv11 object detection network for the i-th Z-axis tomographic image is: The YOLOv11 object detection network uses a CSPDarknet backbone and a PAN feature pyramid for multi-scale feature fusion, and runs after serialization optimization by the TensorRT engine. For each Z-axis layer scan image, the detection result output by the YOLOv11 object detection network is obtained as a candidate bounding box six-tuple (xmin, ymin, xmax, ymax, confidence, class). id ), where xmin, ymin, xmax, and ymax are the minimum, minimum, maximum, and maximum X-axis coordinates and Y-axis coordinates of the candidate bounding box, respectively, and confidence is the confidence level, type. Corresponding organoids. Single-layer NMS processing is performed on each Z-axis tomographic image to suppress cross-union ratio. The bounding box, for The bounding boxes are pre-filtered and discarded to obtain the filtered result. No cross-layer deduplication is performed at this stage.

[0061] Step S03: Extract the ROI region and convert it to a grayscale image for ROI cropping and grayscale conversion. Specifically, for each candidate bounding box, crop the ROI sub-image from its source image, i.e., the Z-axis layer scan image where it is located. , The column coordinates within the ROI sub-image. These are the row coordinates within the ROI sub-image. If the ROI sub-image is in RGB format, it is converted to a grayscale image G to eliminate color channel difference interference. If the size of the ROI sub-image is less than 3 pixels, it is discarded directly.

[0062] Step S04: Calculate the spatial ambiguity score using the Laplace variance method. Specifically, a 3×3 Laplacian kernel convolution is applied to the grayscale image G to obtain the response image L, and the variance of L is calculated. As a score of spatial ambiguity ,Right now This variance is used as an indicator of edge sharpness; the larger the variance, the sharper the image.

[0063] Step S05: Calculate the frequency domain sharpness score using two-dimensional FFT high-frequency energy percentage. . Specifically, perform 2D-FFT on the grayscale G and shift the zero frequency to the center, and calculate the power spectrum . Define the cutoff frequency , and count the high-frequency energy proportion as the frequency-domain sharpness score , that is , which is used as the detail richness index.

[0064] Step S06: Dynamically calculate adaptive fusion weights based on global statistical features. Specifically, three indicators of global statistical features are calculated for M full images, i.e., M Z-axis layer-scanned images: the full-image average contrast (mean of pixel standard deviations), the full-image average noise (mean of differential variances after median filtering denoising), the full-image average high-frequency energy proportion . Dynamically adjust the frequency-domain weight according to preset rules : the base value is 0.6, subtract 0.1 when the contrast is low, add 0.1 when the noise is high or the details are rich, constrain it to the interval [0.4, 0.8], and the spatial-domain weight .

[0065] Step S07: Perform weighted fusion of scores and adaptively filter low-quality targets with a threshold. Specifically, first perform Min-Max normalization on the spatial-domain blurriness score BlurScore and the frequency-domain sharpness score FFTScore to eliminate dimensional differences, according to perform weighted fusion to obtain a comprehensive sharpness score , is the normalized spatial-domain blurriness score, is the normalized frequency-domain sharpness score. Calculate the adaptive score threshold , mark candidate targets with the comprehensive sharpness score < T as blurred and invalid, discard them, and retain valid candidate targets carrying the Score.

[0066] Step S08: Aggregate the global candidate set and sort it in descending order of sharpness score. Specifically, attach the Z-axis index and Z-axis depth to each valid candidate target output in step S07, where the Z-axis index of the valid candidate target located on the i-th Z-axis layer-scanned image , the Z-axis depth of the valid candidate target located on the i-th Z-axis layer-scanned image , construct a 9-tuple (xmin, ymin, xmax, ymax, 0, 0, Score, z index , z depth ). Aggregate all valid candidate targets into a global set C and sort them in descending order of Score. If , skip the subsequent deduplication step and directly proceed to step S12.

[0067] Step S09: Perform the first-level deduplication process—2D IoU coarse screening to eliminate non-overlapping candidate pairs in the XY plane. Specifically, calculate the two-dimensional spatial intersection-union ratio (IoU) of any two valid candidate targets. Iterate starting with the highest score: If... If they are determined to be different organoids with non-overlapping XY structures, they are skipped directly; if The pairs marked as "suspected duplicates" are sent to step S10 for Z-axis fine screening. This step requires only 4 multiply-add operations for each pair of valid candidate targets. Although all candidate pairs need to be processed, the overhead of a single candidate pair is minimal.

[0068] Step S10: Perform the second-level deduplication process—Z-axis distance constraint-based fine screening to identify vertically stacked organoids. Specifically, for the "suspected duplicate pairs" marked in step S09, calculate the Z-axis physical distance. and Z-axis threshold , This represents the median diameter of valid candidate targets with high overall sharpness scores that are suspected of being duplicates. If... If suspected duplicates are vertically stacked different organoids, they are retained and not proceeded to step S11; If the suspected duplicate pair is marked as a "highly suspected pair", it will be sent to step S11 for 3D final judgment. This step requires only one subtraction to obtain the absolute value and one comparison operation for each candidate pair.

[0069] Step S11: Perform the third-level deduplication process—3D IoU final judgment—to accurately determine spatial overlap relationships. Specifically, highly likely candidate targets are expanded into 3D bounding boxes. Among them, fixed step size , The step size for layer scanning, The diameter of the valid candidate targets. Calculate the 3D intersection-union ratio. .like If two or more organoids are identified as spatially overlapping, valid candidate targets with lower overall sharpness scores in highly probable pairs are suppressed; otherwise, they are retained. Although the computational cost of 3D IoU is relatively large (~20 multiply-add / pair), highly probable pairs usually account for a small percentage of all candidate targets, possibly only 1%~5%, so the impact on overall efficiency is controllable. The final output consists of L unique organoid targets forming a unique organoid set.

[0070] Step S12: Calculate the organoid's center coordinates, diameter, and volume data to perform organoid localization data calculation. Specifically, for the unique set of organoids after three-level cascade deduplication, calculate the center point coordinates of the organoid corresponding to each valid candidate target. , ,diameter Z-axis depth is The rationale for using the maximum width and height as the diameter estimate is that organoids are approximately spherical, and their projections are nearly circular. Using the maximum width and height is more conservative and closer to the actual physical dimensions. The volume of the approximate sphere is then calculated. The volume of organoids corresponding to valid candidate targets.

[0071] Step S13: Outlier detection and conversion of pixel coordinates to physical dimensions. Specifically, verifying whether the diameter is within a reasonable range [5μm, 1000μm], whether the center point coordinates exceed the limits, and whether the Z-axis index is valid. Outlier warnings are recorded but retained by default for manual review. The pixel-to-physical-size calibration coefficient is determined using a microscope. Convert pixel coordinates to their corresponding physical coordinates , , .

[0072] Step S14: Visualize and render the labeled image. Specifically, in the original bright-field image, i.e., the acquired... On the Z-axis tomographic image, a green circular border and a red center point marker are drawn with the center point of each unique organoid target in the unique organoid set as the center and a radius of diameter / 2 as the equivalent radius. The serial numbers are labeled, and a localized rendering image with annotations is generated to obtain the final 3D reconstruction result. If there is overlap between different annotations, radial offset processing is performed, and the part that exceeds the image edge is cropped when it is close to the image edge.

[0073] The method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model provided in this disclosure has the following technical advantages compared to the prior art: 1. By performing layer-by-layer detection on multi-focal plane image sequences acquired by bright-field microscopy and associating them with Z-axis coordinate information, the three-dimensional spatial distribution of in vitro three-dimensional biological models can be automatically reconstructed without the need for fluorescent labeling. This significantly reduces experimental costs and operational complexity, solves the problem of "missing Z-axis information" in existing methods, and realizes the reconstruction of the three-dimensional spatial distribution of in vitro three-dimensional biological models based on bright-field Z-axis tomography. 2. The detection results of all Z-axis layer images are aggregated into a global candidate set. The comprehensive sharpness score is introduced as the sorting criterion. A three-level cascaded cross-layer deduplication mechanism is established. The three-level cascaded pipeline of "2D IoU coarse screening → Z-axis distance fine screening → 3D IoU final judgment" achieves accurate deduplication and solves the counting deviation problem caused by "cross-layer repeated detection". 3. By adopting a candidate target quality assessment method based on weighted fusion of ambiguity score (spatial domain) and frequency domain sharpness score (frequency domain), and dynamically and adaptively calculating weights according to the current global statistical features of the image, the detection results of each in vitro 3D biological model on the optimal focal plane are automatically retained, thus solving the problem of "lack of automated standards for focal plane selection" and realizing adaptive optimal focal plane screening. 4. By using only Z-axis tomographic images from a bright-field microscope, combined with a deep learning target detection model (YOLO architecture) and TensorRT inference engine acceleration, high-precision, high-throughput in vitro three-dimensional biological model detection and three-dimensional spatial distribution reconstruction under fluorescent label-free conditions were achieved, solving the problem of "fluorescent label dependence" and realizing high-precision detection under bright-field label-free conditions. 5. By introducing Z-axis distance constraints and 3D IoU discrimination mechanism, Z-axis distance constraints (second-level fine screening) and 3D IoU final judgment (third-level deduplication) are introduced in cross-layer deduplication. This can correctly identify and retain different in vitro 3D biological models stacked vertically, and solve the edge scene problem of "mis-merging of vertically stacked in vitro 3D biological models".

[0074] In particular, the three-dimensional spatial distribution reconstruction method for in vitro three-dimensional biological models provided in this disclosure is model-independent and inference engine-independent. Its core innovative framework (adaptive sharpness assessment + three-level cascaded cross-layer deduplication) is compatible with multiple object detection architectures such as YOLOv8 / v11 / YOLOX / RTMDet / DETR and multiple inference engines such as TensorRT / ONNX Runtime / OpenVINO / ncnn, and has a wide range of adaptability.

[0075] Those skilled in the art will understand that the above embodiments are specific implementations of this disclosure, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of this disclosure.

Claims

1. A method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model, characterized in that, include: Z-axis tomography was performed on cultured samples of an in vitro three-dimensional biological model using a bright-field microscope, and multiple Z-axis tomography images with different focal planes were acquired. Each Z-axis tomographic image is input into a preset target detection network for in vitro three-dimensional biological model detection to obtain the corresponding candidate target. For each candidate target, a spatial ambiguity score and a frequency domain sharpness score are calculated respectively. The spatial ambiguity score and the frequency domain sharpness score are then merged into a comprehensive sharpness score according to an adaptive weight. Candidate targets whose comprehensive sharpness scores meet the adaptive score threshold requirements are determined as valid candidate targets. Each of the effective candidate targets is subjected to cross-layer deduplication and vertical stacking classification to obtain a unique set of in vitro three-dimensional biological models. The three-dimensional spatial distribution of the in vitro three-dimensional biological model is reconstructed based on the unique set of in vitro three-dimensional biological models.

2. The method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model according to claim 1, characterized in that, For each of the candidate targets, a spatial ambiguity score and a frequency domain sharpness score are calculated, including: For each candidate target, the corresponding ROI sub-image is cropped from the Z-axis layer scan image in which it is located; Apply a Laplacian kernel of a preset size to the grayscale image of each ROI sub-image to obtain the corresponding response image; The variance of each response map is calculated as the spatial ambiguity score of the corresponding candidate target; Perform a two-dimensional Fourier transform on the grayscale image of each ROI sub-image and shift the zero frequency to the center. Calculate the power spectrum and count the proportion of high-frequency energy above the defined cutoff frequency as the frequency domain sharpness score of the corresponding candidate target.

3. The method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model according to claim 1, characterized in that, The spatial ambiguity score and the frequency domain sharpness score are fused into a comprehensive sharpness score using adaptive weights, including: The global statistical features of each Z-axis layer scan image are statistically analyzed. The frequency domain weights are dynamically adjusted according to preset rules based on the global statistical features. The spatial ambiguity score and the frequency domain sharpness score are normalized. The spatial ambiguity score and the frequency domain sharpness score are weighted and fused according to the following formula: ； in, The overall sharpness score is given as follows: The normalized spatial ambiguity score is the score of the spatial domain. The normalized frequency domain sharpness score is... The frequency domain weights, Spatial weights and satisfying .

4. The method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model according to claim 3, characterized in that, The global statistical features include the average contrast of the entire image, the average noise of the entire image, and the average high-frequency energy ratio of the entire image. Dynamically adjust the frequency domain weights according to the global statistical features and preset rules, including: Define the base value and range of the frequency domain weights; The frequency domain weights are dynamically adjusted according to the following preset rules: if the contrast of the candidate target is lower than the average contrast of the entire Z-axis tomographic image in which it is located, the frequency domain weights are appropriately reduced from the base value; if the noise of the candidate target is higher than the average noise of the entire Z-axis tomographic image in which it is located, or if the high-frequency energy ratio of the candidate target is higher than the average high-frequency energy ratio of the entire Z-axis tomographic image in which it is located, the frequency domain weights are appropriately increased from the base value; at the same time, it is ensured that the final value of the frequency domain weights is within the value range.

5. The method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model according to claim 1, characterized in that, Each of the aforementioned valid candidate targets undergoes cross-layer deduplication and vertical stacking classification to obtain a unique set of in vitro three-dimensional biological models, including: A Z-axis index and Z-axis depth are attached to each of the valid candidate targets, and the valid candidate targets are sorted in descending order according to the comprehensive sharpness score; Based on the two-dimensional spatial intersection-union ratio between any two valid candidate targets, suspected duplicate pairs are filtered out. Calculate the Z-axis physical distance between the valid candidate targets in each suspected duplicate pair and determine the corresponding Z-axis threshold; Based on the Z-axis physical distance and the corresponding Z-axis threshold, suspected height pairs are filtered out; Expand the effective candidate targets included in the highly suspected pair into corresponding 3D bounding boxes, calculate the 3D intersection-over-union ratio of the effective candidate targets in the highly suspected pair, and remove the effective candidate targets with lower overall sharpness scores in the highly suspected pair whose 3D intersection-over-union ratio is greater than a preset 3D intersection-over-union ratio threshold; All remaining valid candidate targets are combined into the unique in vitro three-dimensional biological model set.

6. The method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model according to claim 5, characterized in that, Based on the two-dimensional spatial intersection-union ratio between any two of the valid candidate targets, suspected duplicate pairs are filtered out, including: If the two-dimensional spatial intersection-union ratio between two valid candidate targets is greater than a preset two-dimensional intersection-union ratio threshold, then the two valid candidate targets are identified as the suspected duplicate pair.

7. The method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model according to claim 5, characterized in that, Based on the Z-axis physical distance and the corresponding Z-axis threshold, suspected height pairs are filtered out, including: If the Z-axis physical distance between the valid candidate targets in the suspected duplicate pair is less than its corresponding Z-axis threshold, then the suspected duplicate pair is determined as the highly suspected pair; The Z-axis threshold is expressed as: ； in, The Z-axis threshold, The median diameter of the valid candidate targets with the higher overall sharpness score among the suspected duplicate pairs. These are preset positive numbers.

8. The method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model according to claim 1, characterized in that, Reconstructing the three-dimensional spatial distribution of the in vitro three-dimensional biological models based on the unique set of in vitro three-dimensional biological models includes: Calculate the center point coordinates, diameter, Z-axis depth, and volume of each valid candidate target in the unique in vitro three-dimensional biological model set; The pixel coordinates of each valid candidate target are converted into corresponding physical coordinates using a microscope pixel-physical size calibration coefficient. Based on the physical coordinates, each of the valid candidate targets is located and rendered on the Z-axis layer scan image to generate a three-dimensional reconstruction result.

9. The method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model according to claim 8, characterized in that, Before converting the pixel coordinates of each effective candidate target to their corresponding physical coordinates using a microscope pixel-physical size calibration coefficient, the method for reconstructing the three-dimensional spatial distribution of the in vitro three-dimensional biological model further includes: The center point coordinates, diameter, and Z-axis index of the valid candidate targets are verified, and abnormal values ​​are recorded and warnings are issued.

10. The method for reconstructing the three-dimensional spatial distribution of an in vitro three-dimensional biological model according to any one of claims 1 to 9, characterized in that, When each of the Z-axis tomographic images is input into a preset target detection network for in vitro three-dimensional biological model detection, an inference engine is used for optimization.

Images