Method for reconstructing ultrahigh-resolution pathological full slice

Abstract

The invention discloses a method for reconstructing an ultrahigh-resolution pathological full slice. The method comprises the following steps: acquiring an effective tissue area image of a pathological tissue fragment by utilizing a fusion mask; an edge classification model and an image enhancement strategy are adopted to obtain an effective splicing edge of the effective tissue area image; constructing a multi-dimensional matching scoring system to calculate a geometric matching degree of a splicing edge, and screening out an optimal Top-N splicing scheme in combination with splicing quality evaluation of a deep neural network; and based on a genetic algorithm, combining global topological optimization, sub-pixel level fine tuning and a multi-dimensional matching scoring system to evaluate splicing quality, realizing iterative stitching of the effective tissue area image, and finally outputting an optimal splicing scheme. Reconstructing a full-slice image of the pathological tissue based on the optimal splicing scheme; the method has efficient and accurate automatic splicing and iterative optimization capability, and rapid reconstruction of the ultrahigh-resolution pathological full slice is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing and pattern recognition, and in particular relates to a method for reconstructing ultra-high resolution pathological full slices. Background Art

[0002] With rising public health awareness and the advancement of precision medicine strategies, the demand for pathology technologies, the gold standard for disease diagnosis, has grown significantly. Through the integrated analysis of tissue morphology and molecular testing, WMS is driving iterative upgrades in clinical diagnosis and treatment. In the field of histopathology, whole mount sections (WMS) offer multiple advantages in improving pathological diagnostic efficiency by fully preserving the macromorphology and microstructure of resected tissue. First, their panoramic histological features enhance the spatial correspondence between preoperative imaging and pathology findings, providing a morphological basis for precision medicine. Second, whole-layer embedding effectively avoids tissue stratification artifacts associated with traditional sectioning, reducing the risk of diagnostic bias due to sample fragmentation. Third, the preservation of complete tissue context helps pathologists diagnose the spatial relationship between lesions and surrounding structures, significantly improving the accuracy of tumor boundary determination and staging assessment. Despite its numerous advantages, WMS faces multiple challenges in its practical application. The primary bottleneck is the standardized supply of large-format slides, which not only complicates laboratory storage and management but also may force some institutions to abandon this technology due to equipment compatibility issues. Secondly, for organ resection specimens such as the prostate and breast, or large tumor tissues, the largest available slides still cannot meet the full embedding requirements, forcing technicians to perform secondary segmentation of the specimens. In addition, the preparation of whole sections requires higher expertise and technical proficiency than traditional quartered sections, resulting in the failure of WMS to be widely popularized.

[0003] The rise of digital pathology offers the potential to address the bottlenecks of traditional WMS technology. By developing AI-based tissue fragment reconstruction algorithms, multiple whole slide images (WSIs) are intelligently stitched together to simulate the cross-sections of the original specimen and generate artificial WMSs. Research has shown that this digital whole slide can significantly shorten pathologists' reading time, significantly improve diagnostic consistency, and reduce interobserver variability. However, WSI images are fundamentally different from general images. Their high-resolution characteristics (0.25–0.50 μm / pixel) required for microscopic pathology pose multiple technical challenges to the development of stitching algorithms: First, the ultra-high resolution characteristics of WSI mean that the conventional size of a single image can exceed 100,000 × 100,000 pixels, resulting in severe computing resource and time cost pressures for basic image processing operations; Second, unlike traditional full-field digital slides, WSI samples are physically divided into biological tissue fragments with non-overlapping areas before scanning, which completely ineffectively implements the core stitching mechanism based on coordinate matching of overlapping areas; Further complicating matters, the non-uniform shrinkage deformation that is common in tissue samples during preparation makes it difficult for solutions based on direct geometric feature matching to achieve precise alignment.

[0004] To address the above challenges, several tissue fragment stitching algorithms have been proposed. HistoStitcher, developed by early researchers as a semi-automatic registration software, uses pathologists to manually mark edge reference points of adjacent fragments and a four-quadrant iterative stitching strategy that minimizes the mean square error through linear transformation. However, its manual point selection mechanism is not only time-consuming and susceptible to subjective factors, but the mathematical model of a single linear transformation is also difficult to adapt to complex tissue morphological characteristics. To overcome this bottleneck, AutoStitcher implements an automatic stitching process by constructing a multidimensional cost function that includes tissue similarity and contour continuity, and significantly improves stitching efficiency by combining low-resolution pre-alignment and optimization algorithms. However, the algorithm still requires the user to manually mark the position of tissue fragments in the original cross-section (such as the upper left or lower right quadrant). This manual labeling method is impractical when processing slices lacking pathological reference information. Furthermore, due to the closed architecture of the Matlab platform, its deployment flexibility and computational performance still need to be improved. Although PythonStitcher has optimized the system deployment process through Python reconstruction and uses an affine transformation algorithm based on Euclidean distance of corner points to align tissue fragments, its core performance is still highly dependent on the quality of manually provided tissue masks, and is subject to the dual constraints of an upper limit on the number of fragments and sensitivity to tissue morphological diversity.

[0005] In summary, existing tissue fragment splicing algorithms rely heavily on manual annotation and lack fully automated processing capabilities. Their linear registration models struggle to overcome the trade-off between accuracy and efficiency. Furthermore, existing algorithms have limited generalization capabilities and are unable to adapt to complex tissue morphologies and multi-quadrant splicing scenarios. Summary of the Invention

[0006] In view of this, it is necessary to provide a method for reconstructing ultra-high resolution pathological full slices to address the above technical issues, including the following steps:

[0007] Step 1: Generate a fusion mask using threshold segmentation and tissue topology to obtain a valid tissue region image of the pathological tissue fragment;

[0008] Step 2, using a deep learning-based edge classification model and a virtual integrated image enhancement strategy to obtain the effective stitching edge of the effective tissue area image;

[0009] Step 3: Based on Theil-Sen regression, anti-noise modeling is performed on the effective splicing edges, and the alignment transformation matrix is ​​calculated to complete the edge alignment of the effective tissue area image. A multi-dimensional matching scoring system is constructed to calculate the geometric matching degree of the splicing edges. The optimal Top-N splicing solution is selected by combining the splicing quality evaluation of the deep neural network;

[0010] Step 4: A multi-resolution stitching optimization algorithm based on a genetic algorithm performs global topology optimization at the low-resolution level of the image and sub-pixel fine-tuning at the high-resolution level. The stitching quality is evaluated using the multi-dimensional matching scoring system, and high-fitness individuals are selected for crossover mutation to achieve iterative stitching of the effective tissue region images, ultimately outputting the optimal stitching solution.

[0011] Step 5: reconstruct the full-slice image of the pathological tissue based on the optimal splicing solution.

[0012] Furthermore, the specific steps of step 1 are as follows:

[0013] Step 101: Obtain and use the pyramid-level downsampling factor of the pathological tissue fragment image to establish a coordinate transformation relationship from the original resolution level to the target resolution level image. Use the contour scaling function to achieve geometric coordinate transformation of different resolution levels to ensure the topological structure consistency of the tissue region across resolutions.

[0014] Step 102: After downsampling the pathological tissue fragment image to the target resolution level, perform HSV color space conversion and extract the saturation channel. A median filter is used to eliminate the granular noise generated during the slide preparation process. An adaptive Otsu threshold segmentation algorithm is used to generate an initialized binary mask. Morphological closing operation is used for edge smoothing. Connected component analysis is combined to filter out discrete background noise to obtain an optimized Otsu threshold segmentation mask.

[0015] Step 103: Extract the effective tissue region mask of the target resolution hierarchical image based on the preprocessing method of the multi-instance learning framework CLAM, perform vector graphics processing using OpenCV, map the edges of the tissue region and the hole into vector polygons, and convert the tissue region into positive fill polygons and the inner contour of the hole into negative space fill polygons using hierarchical fill rules to generate a topological mask of the pathological tissue fragment;

[0016] In step 104, the Otsu threshold segmentation mask and the pathological tissue fragment topology mask are pixel-wise fused to obtain a fused mask, and then the effective tissue area confirmed by the two masks is retained through a matrix dot multiplication operation. Based on the fused mask and the coordinate transformation relationship, geometric alignment and synchronous cropping are performed on the pathological tissue fragment image to obtain an effective tissue area image of the pathological tissue fragment.

[0017] Furthermore, the specific steps of step 2 are as follows:

[0018] Step 201: Determine the principal axis direction of the effective tissue region image by extracting the contour of the fused mask and analyzing the minimum bounding rectangle (MBR). That is, calculate the angle between the major axis or minor axis of the MBR and the horizontal axis of the effective tissue region image. Then, construct an affine transformation matrix with the center of mass of the effective tissue region as the rotation center to synchronously correct the rotation directions of the fused mask and the effective tissue region image.

[0019] Step 202: Use EfficientNet-B0 as the basic network architecture to build a feature extractor for splicing edge classification, remove the original classification layer of the basic network architecture, and retain the original convolution layer to obtain the backbone network EfficientNet Backbone , used to perform feature encoding on the corrected effective tissue area image to obtain a corresponding feature map;

[0020] Step 203: Design an adaptive classification head, compress the feature map through a global average pooling layer, introduce a batch normalization layer and a 0.5 dropout layer for regularization, and use a fully connected layer with softmax activation to output the category probability distribution to obtain the original predicted label of the effective splicing edge;

[0021] Step 204: Using a virtual integrated image enhancement strategy, 16 geometrically transformed versions of the effective tissue region image are generated. The geometric transformations include any combination of image rotation transformations and mirror flip transformations to form a pseudo image set of the effective tissue region image. The feature extractor is then used to obtain the corresponding transformed image predicted pseudo labels in the pseudo image set. At the same time, a label space inverse transform decoder is designed to map the transformed image predicted pseudo labels back to the original predicted labels, thereby achieving multi-view decision fusion of a single model and improving the confidence of the classification results.

[0022] Step 205, based on the edge classification model and the virtual integrated image enhancement strategy, a majority voting mechanism is used to process the classification prediction results of the 16 image enhancements, a prediction label frequency statistical histogram is constructed, the highest frequency prediction label is selected as the final output label, and the effective splicing edge of the effective tissue area image of the corresponding pathological tissue fragment is obtained.

[0023] Furthermore, the specific steps of step 3 are as follows:

[0024] Step 301: Use an edge detection algorithm to obtain the effective splicing edge contour coordinate information of each pathological tissue fragment in a single full-slice sample, and generate all possible tissue fragment pair combinations through Cartesian product operations to ensure that all potential tissue fragment matching possibilities are traversed;

[0025] Step 302 , using the anti-outlier feature of the Theil-Sen regression algorithm to perform anti-noise modeling of effective splicing edge features, constructing an alignment transformation matrix by calculating the centroid of the splicing edge line and the relative components of the rotation angle, achieving geometric alignment of the splicing edges of the tissue fragments, and verifying the overlap of the tissue fragments to ensure the splicing quality;

[0026] In step 303, a multi-dimensional matching scoring system is designed to calculate the stitching edge contour length difference matching score and the fusion mask overlap matching score, and the stitching edge contour similarity matching score is calculated using the Hausdorff distance algorithm. The fitness matching score is obtained through linear combination and dynamic truncation. At the same time, a low-resolution automatic stitching network based on deep learning is used to further evaluate the stitching quality, and finally a stitching solution with the top-N matches is obtained.

[0027] Furthermore, step 302 includes the following steps:

[0028] Step 30201: After obtaining all matching combinations of the pathological tissue fragments, a Theil-Sen regression algorithm is used to establish models for the horizontal edge and the vertical edge of each splicing edge line;

[0029] Step 30202: The centroid of each splicing edge line is calculated using the endpoint midpoint instead of the traditional centroid. The rotation angle of each splicing edge line is determined by the difference between the regression line endpoints. The relative translation component of any two splicing edges is calculated to compensate for the coordinate offset caused by the initial position difference and rotation. The alignment coordinate transformation matrix is ​​constructed to achieve geometric correction of the splicing edges of the tissue fragments.

[0030] In step 30203, a mismatch detection criterion based on fusion mask area statistics is constructed. When the fusion mask overlap rate of the tissue fragment image is detected to be greater than 20%, the 180° rotation compensation mechanism is triggered and the corresponding translation relative component is updated. The alignment coordinate transformation matrix is ​​recalculated to resolve the mirror matching error caused by contour symmetry.

[0031] Furthermore, step 303 includes the following steps:

[0032] Step 30301, using the splicing edge length consistency assessment, set the splicing edge line lengths of any two valid tissue area images to be L1 and L2 respectively, and then the splicing edge contour length difference matching score fitness between the two len for:

[0033]

[0034] If the lengths of the splicing edge lines are similar, the score is higher; if the lengths differ greatly, the score is lower;

[0035] Step 30302: Use spatial exclusive constraints to construct an exponential fusion mask overlap matching score. Set the fusion mask overlap rate of any two valid tissue region images to ρ, and the fusion mask overlap matching score fitness ρ for:

[0036] fitness ρ =100(1-ρ) 6

[0037] If the fusion mask overlap ratio ρ is larger, the score is lower; if the fusion mask overlap ratio ρ is smaller, the score is higher;

[0038] Step 30303: Use contour similarity for measurement, establish an inverse value mapping based on the bidirectional Hausdorff distance, and set the Hausdorff distance between the edge lines of any two effective tissue area images A and B as d H (A→B),d H (B→A), the contour similarity matching score fitness based on Hausdorff distance Hausdorff for:

[0039]

[0040] If the shapes of the two splicing edges are similar, the smaller the Hausdorff distance, the higher the score, and vice versa;

[0041] In step 30304, the edge contour length difference, fusion mask overlap, and contour similarity matching scores are comprehensively scored. The robustness of the evaluation is ensured through linear combination and dynamic truncation to obtain the fitness matching score. The formula is as follows:

[0042] fitness=clip(fitness len +fitness p +fitness Hausdorff ,0,400)

[0043] The clipping operation limits the influence of outliers, and the 400-point threshold is designed so that the ideal match is:

[0044] fitness hausdorff >250 and fitness len >250; qualified match is fitness>180; invalid match is fitness<100;

[0045] Step 30305: Using a data storage format compatible with the JigsawNet framework, save the pairing information of all pathological tissue fragment images, including node index, matching score, transformation matrix, and stitching edge classification results. Simultaneously, calculate the dilated mask intersection of any two pathological tissue fragment images, skeletonize the fusion mask to create a pseudo stitching line representing the optimal stitching path, and convert the skeletonized coordinates to the original image coordinates for visualization.

[0046] In step 30306, the low-resolution automatic stitching network JigsawNetWithROI pre-trained with the traditional fragment stitching dataset is used to further evaluate the image edge stitching quality of the tissue fragment pairs, improve the confidence of the fitness matching score, and finally obtain the top-N matching stitching solutions.

[0047] Furthermore, the specific steps of step 4 are as follows:

[0048] Step 401: construct a four-level resolution pyramid using macro-topology, regional structure, local features, and sub-pixel downsampling factors. Global topology optimization is performed at the low-resolution level of the image, while sub-pixel fine-tuning is performed at the high-resolution level, effectively balancing computational efficiency and stitching accuracy.

[0049] Step 402 , obtaining the stitching information of the multi-resolution tissue, including the coordinate transformation relationship of the multi-resolution, the corner points in the tissue fragments, and the transformation matrix, and extracting the Theil-Sen lines of the relevant edges as robust approximations of the stitching edges to initialize the tissue fragment pairing transformation;

[0050] In step 403, the quality of the paired tissue fragment splicing is initialized using the multi-dimensional matching scoring system described in step 3, and individuals with high fitness are selected as parents for crossover mutation. The optimal solution offspring are retained for directed mutation to achieve the iterative suturing process of pathological tissue fragments and ultimately output the optimal splicing solution.

[0051] Furthermore, step 403 includes the following steps:

[0052] Step 40301: Initialize the number of iterations, fitness matching score, and crossover mutation probability, use the top-N splicing schemes as the initial population, and use the splicing information parameters as the gene values, and then run the genetic algorithm for iteration.

[0053] Step 40302, repeating the iterative process of evaluating all solutions in each generation until a predefined maximum number of iterations is reached, or the best solution fails to improve for 50 consecutive generations;

[0054] In step 40303, a four-level resolution layer-by-layer optimization strategy is adopted to complete the main image registration at the lowest resolution level, and then the preliminary registration results are fine-tuned using the high-resolution level to finally output the optimal stitching solution.

[0055] Furthermore, step 5 further includes the following optimization steps:

[0056] Step 501: Divide the high-resolution pathological tissue fragment image into 64×64 pixel basic image block units, and independently perform multi-resolution optimization on each image block. The obtained rigid transformation parameters include relative rotation components, relative translation components, and transformation matrices. By converting the high-resolution pathological tissue fragment image into calculations for each image block, computational overhead is reduced during the reconstruction process.

[0057] Step 502 : The overlapping areas of the adjacent pathological tissue fragment images are fused based on bilinear interpolation to make the splicing edges between the pathological tissue fragments more visually attractive, and finally reconstruct a full-slice image with the best fitting effect.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] The method of the present invention is an efficient and accurate automatic stitching and iterative optimization method that achieves rapid reconstruction of ultra-high-resolution pathology whole slides. It can be used to reconstruct artificial WMS from pathology tissue fragments to assist clinical evaluation and promote radiology-pathology related workflows. The reconstruction method is only affected by the shape of the pathology tissue fragments and does not require any further image information during the optimization process, making it an essentially stain-independent solution suitable for accurately reconstructing the original WMS on a range of different datasets. In addition, similar stitching accuracy and speed are demonstrated based on automatically detected fiducials. The efficient algorithm design and algorithm containerization improve the reproducibility of reconstruction results on conventional clinical workstations, which will help clinicians shorten case evaluation time, improve radiology-pathology related workflows, and promote research in the field of medical multimodality.

[0060] At the same time, the present invention also uses fusion masks and edge classification models to detect the splicing edges of effective tissue areas, performs precise alignment through multi-dimensional edge matching comprehensive scoring, and designs a deep iteration of an intelligent optimization strategy based on a genetic algorithm to obtain the optimal splicing solution, thereby achieving rapid reconstruction of ultra-high-resolution full pathology slices. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic flow chart of an implementation method of the present invention is shown.

[0062] Figure 2 A schematic diagram of a reconstructed full slice according to an implementation method of the present invention is shown. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0064] In order to better understand the technical content of this embodiment, the following explanations of the terms involved in this embodiment are first introduced.

[0065] Whole Mount Section (WMS): refers to the pathological section of the entire tissue or organ.

[0066] EfficientNet Edge Classifier: refers to an efficient convolutional neural network architecture, a model for edge detection and classification.

[0067] Morphological Operations: It is an image processing technique that is mainly used to analyze and process the shape and structure of images, including erosion, dilation, opening, closing, etc.

[0068] Connected Component Analysis (CCA): is an image processing technique used to identify and label interconnected pixel regions in an image, helping to identify independent objects or regions in the image.

[0069] Majority Voting: It is an ensemble learning method used to combine the prediction results of multiple classifiers to improve the overall classification performance.

[0070] Theil-Sen regression algorithm: A robust linear regression method for estimating the parameters of a linear model that is more robust to outliers.

[0071] Hausdorff distance algorithm: used to measure the similarity between two point sets. It is defined as the maximum and minimum distance from one point set to another point set, that is, the maximum distance from each point to the nearest point.

[0072] Skeletonization: An image processing technique used to extract the skeleton of an object in a binary image. This technique preserves the object's topology while reducing its width to a thin line. Skeletonization is commonly used in shape analysis, pattern recognition, and image segmentation.

[0073] Genetic Algorithm (GA): An optimization algorithm based on natural selection and genetic mechanisms. It iteratively optimizes solutions to problems by simulating the biological evolution process. The main steps include selection, crossover, and mutation.

[0074] Figure 1 A method for reconstructing ultra-high resolution pathology full slices is shown in a flow chart of an embodiment of the present invention. The method comprises the following steps:

[0075] Step 1: Generate a fusion mask using threshold segmentation and tissue topology to obtain a valid tissue region image of the pathological tissue fragment;

[0076] Step 2, using a deep learning-based edge classification model and a virtual integrated image enhancement strategy to obtain effective stitching edges of the effective tissue area image;

[0077] Step 3: Based on Theil-Sen regression, anti-noise modeling is performed on the effective splicing edges, and the alignment transformation matrix is ​​calculated to complete the edge alignment of the effective tissue area image. A multi-dimensional matching scoring system is constructed to calculate the geometric matching degree of the splicing edges. The optimal Top-N splicing solution is selected by combining the splicing quality evaluation of the deep neural network;

[0078] Step 4: A multi-resolution stitching optimization algorithm based on a genetic algorithm performs global topology optimization at the low-resolution level of the image and sub-pixel fine-tuning at the high-resolution level. The stitching quality is evaluated using the multi-dimensional matching scoring system, and high-fitness individuals are selected for crossover mutation to achieve iterative stitching of the effective tissue region images, ultimately outputting the optimal stitching solution.

[0079] Step 5: reconstruct the full-slice image of the pathological tissue based on the optimal splicing solution.

[0080] The following example illustrates the process of reconstructing ultra-high-resolution pathology full slides.

[0081] Specifically, the step 1 includes the following steps:

[0082] Step 101: Obtain and use the pyramid level downsampling factor of the pathological tissue fragment image to establish a coordinate transformation relationship from the original resolution level (level 0) to the target resolution level (level n). Use the contour scaling function to achieve geometric coordinate transformation of different resolution levels to ensure the topological structure consistency of the tissue region across resolutions.

[0083] Step 102: After downsampling the pathological tissue fragment image to the target resolution level, perform HSV color space conversion and extract the saturation channel. A median filter is used to eliminate the granular noise generated during the slide preparation process. An adaptive Otsu threshold segmentation algorithm is used to generate an initialized binary mask. Morphological closing operation is used for edge smoothing. Connected component analysis is combined to filter out discrete background noise to obtain an optimized Otsu threshold segmentation mask.

[0084] Among them, Otsu threshold segmentation is a threshold method that automatically selects the image binarization threshold, which determines the optimal threshold by minimizing the intra-class variance of the image or equivalently maximizing the inter-class variance, thereby segmenting the image into background and foreground with valid labels;

[0085] Step 103: Extract the effective tissue region mask of the target resolution hierarchical image based on the preprocessing method of the multi-instance learning framework CLAM, perform vector graphics processing using OpenCV, map the edges of the tissue region and the hole into vector polygons, and convert the tissue region into positive filling polygons and the inner contour of the hole into negative space filling polygons (i.e., tissue topology structure) using hierarchical filling rules to generate a topological mask of the pathological tissue fragment;

[0086] Among them, the pathological tissue fragment topological mask is a mask for marking pathological tissue fragments. It not only marks the tissue area, but also takes into account the topological structure of the tissue to ensure the connectivity and integrity of the tissue area.

[0087] In step 104, the Otsu threshold segmentation mask and the pathological tissue fragment topology mask are pixel-wise fused to obtain a fused mask, and then the effective tissue area confirmed by the two masks is retained through a matrix dot multiplication operation. Based on the fused mask and the coordinate transformation relationship, geometric alignment and synchronous cropping are performed on the pathological tissue fragment image to obtain an effective tissue area image of the pathological tissue fragment.

[0088] Specifically, the step 2 specifically includes the following steps:

[0089] Step 201: Determine the main axis direction of the effective tissue region image by extracting the contour of the fusion mask and analyzing the minimum bounding rectangle, that is, calculate the angle θ∈[-45°, 45°] between the major axis or minor axis of the minimum bounding rectangle and the horizontal axis of the effective tissue region image, and then use the centroid (x c ,y c ) is used as the rotation center to construct the affine transformation matrix. The affine transformation matrix M is calculated as follows:

[0090]

[0091] Wherein, α=cosθ represents the cosine value of the angle θ, and β=cosθ represents the sine value of the angle θ, so as to synchronously correct the rotation directions of the fusion mask and the effective tissue area image, facilitating subsequent edge detection and classification;

[0092] Step 202: construct a feature extractor for splicing edge classification, use EfficientNet-B0 as the basic network architecture (B0 represents the benchmark version B0 of EfficientNet), remove the original classification layer of the basic network architecture, and retain the original convolution layer to obtain the backbone network EfficientNet-B0 BackboneIt is used to perform feature encoding on the corrected effective tissue area image to obtain a feature map, and adopt a transfer learning strategy to load pre-trained weights and implement custom initialization. The feature extraction process is as follows:

[0093] feat=EfficientNet Backbone (input)

[0094] Among them, input is the image downsampled to the target resolution level, feat∈R C×H×W is the feature map initially extracted from the image, C is the number of channels, H and W are the height and width of the image; EfficientNet Backbone Represents the backbone network of EfficientNet-B0;

[0095] Step 203: Design an adaptive classification head, compress the feature map feat through a global average pooling layer, introduce a batch normalization layer and a 0.5 dropout layer for regularization, and use a fully connected layer with Softmax activation to output the category probability distribution to obtain the original prediction labels of the four categories of the effective splicing edge. The formula is as follows:

[0096] l=Classifier adaptive (GPA(feat))

[0097] Among them, l∈{0:upper right,1:lowerright,2:lowerleft,3:upperleft}) represents the original predicted labels of the four categories of the effective splicing edge and the corresponding position information (for example, the label numbered 0 is located in the upper right, the label numbered 1 is located in the lower right, etc.); feat is the feature map, GPA represents the global average pooling layer; Classifier adaptive Represents the batch normalization layer BN, Dropout regularization layer and fully connected layer with Softmax activation;

[0098] Step 204: Using a virtual integrated image enhancement strategy, 16 geometrically transformed versions of the effective tissue region image are generated. The geometric transformation includes any combination of image rotation transformation and mirror flip transformation to form a pseudo image set I of the effective tissue region image. The rotation transformation includes four rotation modes of 0°, 90°, 180°, and 270°, and the mirror flip includes four flip modes of horizontal flip (hor), vertical flip (ver), no flip (nonr), and composite flip (both) (the geometric transformation type is the product of rotation transformation and mirror flip). The feature extractor is then used to obtain predicted pseudo labels for the transformed images corresponding to the pseudo image set I.

[0099] At the same time, we design the label space inverse transform decoder Γ -1 , the formula is as follows:

[0100] Γ -1 (k,r,f)=rot -r (flip - f(k))

[0101] Among them, k∈{0,1,2,3} is the predicted pseudo label of the transformed image in the pseudo image set I, r∈{0°,90°,180°,270°} is the rotation parameter, f∈{hor,ver,none,both} is the flip parameter, rot -r Indicates the reverse operation of rotation, flip -f represents the inverse operation of flipping, mapping the transformed image predicted pseudo label back to the original predicted label, realizing multi-view decision fusion of a single model and improving the confidence of the classification result;

[0102] It should be noted that this method first predicts the label of the image to obtain the corresponding original predicted label, then performs geometric transformation and label prediction on the image to obtain the transformed image predicted pseudo label, and then maps the image predicted pseudo label back to the original predicted label to evaluate whether the model can recognize the same valid splicing edge for images with different rotations and flips, thereby improving the confidence of the classification results;

[0103] Step 205: Based on the edge classification model and the virtual integrated image enhancement strategy, a majority voting mechanism is used to process the classification prediction results of the 16 image enhancements and construct a prediction label frequency statistical histogram Counter:

[0104]

[0105] Among them, l represents the original predicted label of the effective splicing edge category, k i represents the predicted pseudo label of the transformed image of the i-th transformation, Γ -1 represents the label space inverse transform decoder that maps the transformed image predicted pseudo labels to the original predicted labels, and δ represents the Dirac function;

[0106] Select the most frequently predicted label as the final output label When the highest frequency of ties occurs, the label with the smallest geometric center distance is preferentially selected, and the other is used as a candidate label for the next evaluation. Finally, the effective splicing edge of the effective tissue region image of the pathological tissue fragment is obtained based on the predicted label.

[0107] Specifically, the step 3 includes the following steps:

[0108] Step 301: Use an edge detection algorithm to obtain the effective splicing edge contour coordinate information of each pathological tissue fragment in a single full-slice sample, and generate all possible tissue fragment pair combinations through Cartesian product operations to ensure that all potential tissue fragment matching possibilities are traversed;

[0109] Step 302 , using the anti-outlier feature of the Theil-Sen regression algorithm to perform anti-noise modeling of effective splicing edge features, constructing an alignment transformation matrix by calculating the centroid of the splicing edge line and the relative components of the rotation angle, achieving geometric alignment of the splicing edges of the tissue fragments, and verifying the overlap of the tissue fragments to ensure the splicing quality;

[0110] In step 303, a multi-dimensional matching scoring system is designed to calculate the stitching edge contour length difference matching score and the fusion mask overlap matching score, and the stitching edge contour similarity matching score is calculated using the Hausdorff distance algorithm. The fitness matching score is obtained through linear combination and dynamic truncation. At the same time, a low-resolution automatic stitching network based on deep learning is used to further evaluate the stitching quality, and finally a stitching solution with the top-N matches is obtained.

[0111] Preferably, the step 302 specifically includes the following steps:

[0112] Step 30201: After obtaining all matching combinations of the pathological tissue fragments, for each splicing edge line, the Theil-Sen regression algorithm is used to establish a y=f(x) model for the horizontal edge and an x=f(t) model for the vertical edge; wherein, in the y=f(x) model, any two data points (x i ,y i ),(x j ,y j ), the coordinate difference in the X direction is Δx, the coordinate difference in the Y direction is Δy, and the slope is expressed as Hyperparameter ∈ = 10 -5 Used to avoid numerical instability and ensure the validity of slope calculation for non-horizontal edges; calculate the median of the slope of the line segment between the data points to determine the slope of the best fit line β m , the calculation formula is as follows:

[0113] β m =Median(β ij )

[0114] Among them, Median represents the median calculation, β m represents the slope of the best fitting line and the intercept of the data point α m =Median(y i -β m x i), and finally construct the Theil-Sen fitting line of each splicing edge to represent y=β m x+α m , the fitting line of the vertical edge x=f(y) model is similar to the above;

[0115] Step 30202: Use the endpoint midpoint instead of the traditional centroid to calculate the centroid Q of each spliced ​​edge line, and determine its respective rotation angle by the regression line endpoint difference. The two effective splicing edge rotation angles corresponding to any two of the effective tissue area images A and B are respectively The center of mass positions are Q A , Q B , calculate the relative component of rotation Compensate for rotational offset and calculate relative translational component Compensating for the initial position difference and the coordinate offset caused by rotation to construct an alignment coordinate transformation matrix W to achieve geometric correction of the effective tissue area image B. The alignment coordinate transformation matrix is ​​calculated as follows:

[0116]

[0117] Among them, t x represents the relative translation component of the effective tissue region image B in the X-axis direction, t y represents the relative translation component of the effective tissue region image B in the Y-axis direction;

[0118] Step 30203: Construct a mismatch detection criterion based on fusion mask area statistics. The formula is as follows:

[0119]

[0120] Where ρ represents the fusion mask overlap ratio of the tissue fragment image, A and B are the two valid tissue region images mentioned above, A∩W(B) represents the overlapping image area between image A and image B after coordinate transformation, and S represents the area of ​​the image fusion mask region;

[0121] When the overlap rate of the fusion mask of the tissue fragment image is detected to be greater than 20%, it proves that the splicing fitting effect is not good, and the 180° rotation compensation mechanism R is triggered. new =R+180°, R is the relative rotation component, and the corresponding relative translation component R is updated new , recalculate the alignment coordinate transformation matrix to resolve the mirror matching error caused by the contour symmetry.

[0122] Furthermore, the step 303 specifically includes the following steps:

[0123] Step 30301, using the splicing edge length consistency assessment, set the splicing edge line lengths of the two pathological tissue fragments to be L1 and L2 respectively, and then the splicing edge contour length difference matching score fitness between the two len for:

[0124]

[0125] If the lengths of the splicing edge lines are similar, the score is higher; if the lengths differ greatly, the score is lower;

[0126] Step 30302: Using spatial exclusive constraints, construct an exponential fusion mask overlap matching score, set the fusion mask overlap rate of the two pathological tissue fragments to ρ, and the fusion mask overlap matching score fitness ρ for:

[0127] fitness ρ =100(1-ρ) 6

[0128] If the fusion mask overlap ratio ρ is larger, the score is lower; if the fusion mask overlap ratio ρ is smaller, the score is higher;

[0129] Step 30303: Use contour similarity for measurement, establish an inverse value mapping based on the bidirectional Hausdorff distance, and set the Hausdorff distance between the edge lines of the two effective tissue area images A and B as d H (A→B),d H (B→A), the contour similarity matching score fitness based on Hausdorff distance Hausdorff for:

[0130]

[0131] If the shapes of the spliced ​​edges of the two effective tissue region images A and B are more similar, the Hausdorff distance is smaller and the score is higher, and vice versa;

[0132] In step 30304, the edge contour length difference, fusion mask overlap, and contour similarity matching scores are comprehensively scored. The robustness of the evaluation is ensured through linear combination and dynamic truncation. The fitness matching score formula is as follows:

[0133] fitness=clip(fitness len +fitness ρ +fitness Hausdorff ,0,400)

[0134] The clipping operation limits the impact of outliers, and the 400-point threshold is designed so that the ideal match is fitness. hausdorff >250 and fitness len >250; qualified match is fitness>180; invalid match is fitness<100;

[0135] Step 30305: Using a data storage format compatible with the JigsawNet framework, save the pairing information of all pathological tissue fragment images, including node index, matching score, transformation matrix, and stitching edge classification results. Simultaneously, calculate the dilated mask intersection of any two pathological tissue fragment images, skeletonize the fusion mask to create a pseudo stitching line representing the optimal stitching path, and convert the skeletonized coordinates to the original image coordinates for visualization.

[0136] In step 30306, the low-resolution automatic stitching network JigsawNetWithROI pre-trained with the traditional fragment stitching dataset is used to further evaluate the image edge stitching quality of the tissue fragment pairs, improve the confidence of the fitness matching score, and finally obtain the top-N matching stitching solutions.

[0137] Specifically, the step 4 includes the following steps:

[0138] Step 401: construct a four-level resolution pyramid using downsampling factors of 2560× for macro topology, 853× for regional structure, 284× for local features, and 128× for sub-pixels. Global topology optimization is performed at the low-resolution level of the image, while sub-pixel fine-tuning is performed at the high-resolution level, effectively balancing computational efficiency and stitching accuracy.

[0139] Step 402 , obtaining the stitching information of the multi-resolution tissue, including the coordinate transformation relationship of the multi-resolution, the corner points in the tissue fragments, and the transformation matrix, and extracting the Theil-Sen lines of the relevant edges as robust approximations of the stitching edges to initialize the tissue fragment pairing transformation;

[0140] Step 403: Initialize the stitching quality of paired tissue fragments using the multi-dimensional matching scoring system described in step 3, select individuals with high fitness as parents for crossover mutation, retain the optimal solution offspring for directed mutation, and implement an iterative stitching process for pathological tissue fragments, ultimately outputting the optimal stitching solution.

[0141] Among them, the iterative stitching process is equivalent to basically aligning the position of the image block. Since there are still many gaps between the corroded edges, it is necessary to further fine-tune the position coordinates of the tissue fragments for fitting. By selecting a solution with a high fitness score and continuously iteratively adjusting, the best stitching solution is obtained.

[0142] Preferably, the step 403 specifically includes the following steps:

[0143] Step 40301 initializes the number of iterations n, the fitness matching score fitness, and the probability of crossover mutation P. The top-N splicing solutions are used as the initial population N, and the splicing information parameters are used as the gene value B. Then, the genetic algorithm GA is run iteratively to obtain the splicing solution as follows:

[0144] solution=GA(N,V,n,fitness,P)

[0145] The splicing solution includes the total number of the pathological tissue fragments, the global position (x, y) and rotation angle of each pathological tissue fragment, the connection relationship between the pathological tissue fragments, and the fitness matching score and other information.

[0146] Step 40302, repeating the iterative process of evaluating all solutions in each generation until a predefined maximum number of iterations is reached, or the best solution fails to improve for 50 consecutive generations;

[0147] In step 40303, a four-level resolution layer-by-layer optimization strategy is adopted to complete the main image registration at the lowest resolution level, and then the preliminary registration results are fine-tuned using the high-resolution level to finally output the optimal stitching solution.

[0148] Specifically, step 5 further includes the following optimization steps:

[0149] Step 501: Divide the high-resolution pathological tissue fragment image into 64×64 pixel basic image block units, and independently perform multi-resolution optimization on each image block. The obtained rigid transformation parameters include relative rotation components, relative translation components, and transformation matrices. By converting the high-resolution pathological tissue fragment image into calculations for each image block, computational overhead is reduced during the reconstruction process.

[0150] Step 502 : The overlapping areas of the adjacent pathological tissue fragment images are fused based on bilinear interpolation to make the splicing edges between the pathological tissue fragments more visually attractive, and finally reconstruct a full-slice image with the best fitting effect.

[0151] The present invention provides an efficient and accurate automatic stitching and iterative optimization method to achieve rapid reconstruction of ultra-high-resolution pathology full slices. It can be used to reconstruct artificial WMS from pathology tissue fragments to assist clinical evaluation and promote radiology-pathology related workflows. This reconstruction method is only affected by the shape of the pathological tissue and does not require any further image information during the optimization process, making it essentially a stain-independent solution suitable for accurately reconstructing the original WMS on a range of different datasets. In addition, efficient algorithm design and algorithm containerization improve the repeatability of reconstruction results on conventional clinical workstations, which will help clinicians shorten case evaluation time, improve radiology-pathology related workflows, and promote research in the field of medical multimodality.

[0152] The various technical features or steps of the above embodiments can be arbitrarily combined. To make the description concise, not all possible combinations of the various technical features or steps in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0153] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Claims

1. A method for reconstructing ultra-high resolution pathological full sections, characterized in that: The following steps are involved: Step 1: Generate a fusion mask using threshold segmentation and tissue topology to obtain a valid tissue region image of the pathological tissue fragment; Step 2, using a deep learning-based edge classification model and a virtual integrated image enhancement strategy to obtain the effective stitching edge of the effective tissue area image; Step 3: Based on Theil-Sen regression, anti-noise modeling is performed on the effective splicing edges, and the alignment transformation matrix is ​​calculated to complete the edge alignment of the effective tissue area image. A multi-dimensional matching scoring system is constructed to calculate the geometric matching degree of the splicing edges. The optimal Top-N splicing solution is selected by combining the splicing quality evaluation of the deep neural network; Step 4: A multi-resolution stitching optimization algorithm based on a genetic algorithm performs global topology optimization at the low-resolution level of the image and sub-pixel fine-tuning at the high-resolution level. The stitching quality is evaluated using the multi-dimensional matching scoring system, and high-fitness individuals are selected for crossover mutation to achieve iterative stitching of the effective tissue region images, ultimately outputting the optimal stitching solution. Step 5: reconstruct the full-slice image of the pathological tissue based on the optimal splicing solution.

2. The method for reconstructing ultra-high resolution pathological full sections according to claim 1, wherein: The specific steps of step 1 are as follows: Step 101: Obtain and use the pyramid-level downsampling factor of the pathological tissue fragment image to establish a coordinate transformation relationship from the original resolution level to the target resolution level image. Use the contour scaling function to achieve geometric coordinate transformation of different resolution levels to ensure the topological structure consistency of the tissue region across resolutions. Step 102: After downsampling the pathological tissue fragment image to the target resolution level, perform HSV color space conversion and extract the saturation channel. A median filter is used to eliminate the granular noise generated during the slide preparation process. An adaptive Otsu threshold segmentation algorithm is used to generate an initialized binary mask. Morphological closing operation is used for edge smoothing. Connected component analysis is combined to filter out discrete background noise to obtain an optimized Otsu threshold segmentation mask. Step 103: Extract the effective tissue region mask of the target resolution hierarchical image based on the preprocessing method of the multi-instance learning framework CLAM, perform vector graphics processing using OpenCV, map the edges of the tissue region and the hole into vector polygons, and convert the tissue region into positive fill polygons and the inner contour of the hole into negative space fill polygons using hierarchical fill rules to generate a topological mask of the pathological tissue fragment; In step 104, the Otsu threshold segmentation mask and the pathological tissue fragment topology mask are pixel-wise fused to obtain a fused mask, and then the effective tissue area confirmed by the two masks is retained through a matrix dot multiplication operation. Based on the fused mask and the coordinate transformation relationship, geometric alignment and synchronous cropping are performed on the pathological tissue fragment image to obtain an effective tissue area image of the pathological tissue fragment.

3. The method for reconstructing ultra-high resolution pathological full sections according to claim 1, wherein: The specific steps of step 2 are as follows: Step 201: Determine the principal axis direction of the effective tissue region image by extracting the contour of the fused mask and analyzing the minimum bounding rectangle (MBR). That is, calculate the angle between the major axis or minor axis of the MBR and the horizontal axis of the effective tissue region image. Then, construct an affine transformation matrix with the center of mass of the effective tissue region as the rotation center to synchronously correct the rotation directions of the fused mask and the effective tissue region image. Step 202: Use EfficientNet-B0 as the basic network architecture to build a feature extractor for splicing edge classification, remove the original classification layer of the basic network architecture, and retain the original convolution layer to obtain the backbone network EfficientNet Backbone , used to perform feature encoding on the corrected effective tissue area image to obtain a corresponding feature map; Step 203: Design an adaptive classification head, compress the feature map through a global average pooling layer, introduce a batch normalization layer and a 0.5 dropout layer for regularization, and use a fully connected layer with softmax activation to output the category probability distribution to obtain the original predicted label of the effective splicing edge; Step 204: Using a virtual integrated image enhancement strategy, 16 geometrically transformed versions of the effective tissue region image are generated. The geometric transformations include any combination of image rotation transformations and mirror flip transformations to form a pseudo image set of the effective tissue region image. The feature extractor is then used to obtain the corresponding transformed image predicted pseudo labels in the pseudo image set. At the same time, a label space inverse transform decoder is designed to map the transformed image predicted pseudo labels back to the original predicted labels, thereby achieving multi-view decision fusion of a single model and improving the confidence of the classification results. Step 205, based on the edge classification model and the virtual integrated image enhancement strategy, a majority voting mechanism is used to process the classification prediction results of the 16 image enhancements, a prediction label frequency statistical histogram is constructed, the highest frequency prediction label is selected as the final output label, and the effective splicing edge of the effective tissue area image of the corresponding pathological tissue fragment is obtained.

4. The method for reconstructing ultra-high resolution pathological full sections according to claim 1, wherein: The specific steps of step 3 are as follows: Step 301: Use an edge detection algorithm to obtain the effective splicing edge contour coordinate information of each pathological tissue fragment in a single full-slice sample, and generate all possible tissue fragment pair combinations through Cartesian product operations to ensure that all potential tissue fragment matching possibilities are traversed; Step 302 , using the anti-outlier feature of the Theil-Sen regression algorithm to perform anti-noise modeling of effective splicing edge features, constructing an alignment transformation matrix by calculating the centroid of the splicing edge line and the relative components of the rotation angle, achieving geometric alignment of the splicing edges of the tissue fragments, and verifying the overlap of the tissue fragments to ensure the splicing quality; In step 303, a multi-dimensional matching scoring system is designed to calculate the stitching edge contour length difference matching score and the fusion mask overlap matching score, and the stitching edge contour similarity matching score is calculated using the Hausdorff distance algorithm. The fitness matching score is obtained through linear combination and dynamic truncation. At the same time, a low-resolution automatic stitching network based on deep learning is used to further evaluate the stitching quality, and finally a stitching solution with the top-N matches is obtained.

5. The method for reconstructing ultra-high resolution pathological whole sections according to claim 4, characterized in that: Step 302 includes the following steps: Step 30201: After obtaining all matching combinations of the pathological tissue fragments, a Theil-Sen regression algorithm is used to establish models for the horizontal edge and the vertical edge of each splicing edge line; Step 30202: The centroid of each splicing edge line is calculated using the endpoint midpoint instead of the traditional centroid. The rotation angle of each splicing edge line is determined by the difference between the regression line endpoints. The relative translation component of any two splicing edges is calculated to compensate for the coordinate offset caused by the initial position difference and rotation. The alignment coordinate transformation matrix is ​​constructed to achieve geometric correction of the splicing edges of the tissue fragments. In step 30203, a mismatch detection criterion based on fusion mask area statistics is constructed. When the fusion mask overlap rate of the tissue fragment image is detected to be greater than 20%, the 180° rotation compensation mechanism is triggered and the corresponding translation relative component is updated. The alignment coordinate transformation matrix is ​​recalculated to resolve the mirror matching error caused by contour symmetry.

6. The method for reconstructing ultra-high resolution pathological whole sections according to claim 4, characterized in that: Step 303 includes the following steps: Step 30301, using the splicing edge length consistency assessment, set the splicing edge line lengths of any two valid tissue area images to be L1 and L2 respectively, and then the splicing edge contour length difference matching score fitness between the two len for: If the lengths of the splicing edge lines are similar, the score is higher; if the lengths differ greatly, the score is lower; Step 30302: Use spatial exclusive constraints to construct an exponential fusion mask overlap matching score. Set the fusion mask overlap rate of any two valid tissue region images to ρ, and the fusion mask overlap matching score fitness ρ for: fitness ρ =100(1-ρ) 6 The larger the fusion mask overlap ratio ρ, the lower the score; the smaller the fusion mask overlap ratio ρ, the higher the score; Step 30303: Use contour similarity for measurement, establish an inverse value mapping based on the bidirectional Hausdorff distance, and set the Hausdorff distance between the edge lines of any two effective tissue area images A and B as d H (A→B),d H (B→A), the contour similarity matching score fitness based on Hausdorff distance Hausdorff for: If the shapes of the two splicing edges are similar, the smaller the Hausdorff distance is, the higher the score is, and vice versa; In step 30304, the edge contour length difference, fusion mask overlap, and contour similarity matching scores are comprehensively scored. The robustness of the evaluation is ensured through linear combination and dynamic truncation to obtain the fitness matching score. The formula is as follows: fitness=clip(fitness len +fitness p +fitness Hausdorff ,0,400) The clipping operation limits the impact of outliers, and the 400-point threshold is designed so that the ideal match is fitness. hausdorff >250 and fitness len >250; qualified match is fitness>180; invalid match is fitness<100; Step 30305: Using a data storage format compatible with the JigsawNet framework, save the pairing information of all pathological tissue fragment images, including node index, matching score, transformation matrix, and stitching edge classification results. Simultaneously, calculate the dilated mask intersection of any two pathological tissue fragment images, skeletonize the fusion mask to create a pseudo stitching line representing the optimal stitching path, and convert the skeletonized coordinates to the original image coordinates for visualization. In step 30306, the low-resolution automatic stitching network JigsawNetWithROI pre-trained with the traditional fragment stitching dataset is used to further evaluate the image edge stitching quality of the tissue fragment pairs, improve the confidence of the fitness matching score, and finally obtain the top-N matching stitching solutions.

7. The method for reconstructing ultra-high resolution pathological whole sections according to claim 1, characterized in that: The specific steps of step 4 are as follows: Step 401: construct a four-level resolution pyramid using macro-topology, regional structure, local features, and sub-pixel downsampling factors. Global topology optimization is performed at the low-resolution level of the image, while sub-pixel fine-tuning is performed at the high-resolution level, effectively balancing computational efficiency and stitching accuracy. Step 402 , obtaining the stitching information of the multi-resolution tissue, including the coordinate transformation relationship of the multi-resolution, the corner points in the tissue fragments, and the transformation matrix, and extracting the Theil-Sen lines of the relevant edges as robust approximations of the stitching edges to initialize the tissue fragment pairing transformation; In step 403, the quality of the paired tissue fragment splicing is initialized using the multi-dimensional matching scoring system described in step 3, and individuals with high fitness are selected as parents for crossover mutation. The optimal solution offspring are retained for directed mutation to achieve the iterative suturing process of pathological tissue fragments and ultimately output the optimal splicing solution.

8. The method for reconstructing ultra-high resolution pathological whole sections according to claim 7, characterized in that: Step 403 includes the following steps: Step 40301: Initialize the number of iterations, fitness matching score, and crossover mutation probability, use the top-N splicing schemes as the initial population, and use the splicing information parameters as the gene values, and then run the genetic algorithm for iteration. Step 40302, repeating the iterative process of evaluating all solutions in each generation until a predefined maximum number of iterations is reached, or the best solution fails to improve for 50 consecutive generations; In step 40303, a four-level resolution layer-by-layer optimization strategy is adopted to complete the main image registration at the lowest resolution level, and then the preliminary registration results are fine-tuned using the high-resolution level to finally output the optimal stitching solution.

9. The method for reconstructing ultra-high resolution pathological whole sections according to claim 1, characterized in that: The step 5 also includes the following optimization steps: Step 501: Divide the high-resolution pathological tissue fragment image into 64×64 pixel basic image block units, and independently perform multi-resolution optimization on each image block. The obtained rigid transformation parameters include relative rotation components, relative translation components, and transformation matrices. By converting the high-resolution pathological tissue fragment image into calculations for each image block, computational overhead is reduced during the reconstruction process. Step 502 : The overlapping areas of the adjacent pathological tissue fragment images are fused based on bilinear interpolation to make the splicing edges between the pathological tissue fragments more visually attractive, and finally reconstruct a full-slice image with the best fitting effect.

Images