Panoramic image-based interactive scan image intelligent labeling method and device

Abstract

The application discloses an interactive scanning image intelligent labeling method and device based on a panoramic image, and the method comprises the following steps: splicing all sub-pictures into a high-resolution panoramic slide image based on metadata; performing batch inference on the sub-pictures in the effective area corresponding to the panoramic slide image by using a pre-trained segmentation model, performing semantic coding on a labeling category, and mapping target labeling information to a global coordinate system of the panoramic slide image to obtain a labeling layer; and performing dynamic rendering on the panoramic slide image by using a pre-constructed double-layer rendering architecture, wherein the double-layer rendering architecture comprises a labeling overlay layer, the labeling overlay layer adopts an encoding mode of decoupling an RGB channel from an Alpha channel, real-time interactive adjustment of global transparency of the labeling overlay layer is realized through a front-end control, pre-labeling results are checked and corrected in the panoramic view through an intelligent interaction tool, and all labeling operation sequences are recorded to obtain a structured labeling data file. The efficiency of labeling is improved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and intelligent image processing, and specifically relates to an interactive scanning image intelligent annotation method and device based on panoramic images. Background Technology

[0002] In the field of high-precision slide scanning imaging, the current image annotation process has the following significant drawbacks:

[0003] 1. Damage to image information integrity: During the scanning and acquisition process, images are usually cropped, which causes the permanent loss of sample texture information in the edge areas, making it impossible to fully annotate the area.

[0004] 2. Errors introduced during the scanning process: The unavoidable vibrations of mechanical scanning equipment can cause image distortion, such as blurred edges and misalignment of adjacent images, which further reduces the accuracy and reliability of the labeled data.

[0005] 3. Disconnect between annotation perspective and panoramic perspective: Under current technology, annotation can only be performed based on discrete single scanned images. Operators cannot annotate on the overall panoramic view of the slide, making it difficult to establish a spatial relationship between local annotations and the overall sample structure. This results in low annotation efficiency and an easy omission of edge areas.

[0006] 4. Data Redundancy and Annotation Conflicts: If an overlapping scanning strategy is used to preserve edge information, it will result in adjacent images storing a large amount of redundant information. During annotation, the same target may be repeatedly annotated in multiple images, causing data inconsistency issues.

[0007] 5. Low efficiency of annotation interaction: For irregularly shaped targets, the outline is completely drawn manually, which is cumbersome, time-consuming and labor-intensive, and the annotation accuracy is highly dependent on the operator's proficiency.

[0008] Therefore, there is an urgent need in this field for a technical solution that can preserve image information, provide a panoramic interactive perspective, and improve annotation efficiency. Summary of the Invention

[0009] To address this, embodiments of the present invention provide an interactive scanning image intelligent annotation method and apparatus based on panoramic images, thereby solving at least one technical problem existing in the prior art.

[0010] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions: This invention provides an interactive intelligent annotation method for scanned images based on panoramic images, the method comprising: Based on a predetermined overlap rate, the target slide is subjected to a sequential redundant scan using a scanner, and the metadata of each sub-image obtained from the scan is recorded. Based on the metadata, all sub-images are stitched together into a high-resolution panoramic slide image. The sub-images within the effective region corresponding to the panoramic slide image are batch-inferred using a pre-trained segmentation model to obtain target annotation information; the annotation categories are semantically encoded, and the target annotation information is mapped to the global coordinate system of the panoramic slide image to obtain the annotation layer; The pre-built dual-layer rendering architecture is used to dynamically render panoramic slide images. The dual-layer rendering architecture includes a label overlay layer, which adopts an encoding method that decouples RGB and Alpha channels. Real-time interactive adjustment of the global transparency of the label overlay layer is achieved through front-end controls. In the panoramic view, the pre-labeling results are verified and corrected through intelligent interactive tools, and all labeling operation sequences are recorded to obtain a structured labeling data file. After optimizing and verifying the final structured annotation data file, it is exported as a standard format file containing target category information and accurate vector contours in the panoramic coordinate system.

[0011] In some embodiments, based on a predetermined overlap rate, a sequence of redundant scans of the target slide is performed using a scanner, specifically including: The scanner expands the actual capture area of ​​a single scan and sets overlapping areas in both the horizontal and vertical directions. The scanner automatically performs sequential scanning with the upper left corner of the target slide as the origin and a preset serpentine path. During the scanning process, each sub-image undergoes real-time image registration, overlapping area verification, and scanning position calibration. After verification, the standard valid area is cropped out, and the metadata is recorded in the structured index.

[0012] In some embodiments, real-time image registration is performed for each sub-image, specifically including: Extract key points and feature descriptors from adjacent subgraphs; Calculate the Euclidean distance between feature descriptors to obtain matching point pairs; Filter out mismatches and fit subpixel-level transformation relationships to parse out the horizontal and vertical pixel offsets.

[0013] In some embodiments, the network architecture of the segmentation model includes a feature extraction backbone network, a region proposal network, and parallel classification, bounding box regression, and mask prediction heads.

[0014] In some embodiments, the intelligent interaction tool includes: The intelligent brush tool receives rough brush strokes from the user on the panoramic image. The background image segmentation algorithm identifies the foreground / background regions based on the brush stroke information, calculates and generates a pixel-level precision target contour mask in real time, and updates it to the annotation mask layer. The precision removal tool includes point removal, box selection removal, and smear removal, which correspond to the annotation tool to enable precise modification of annotations; A smart point selection tool that automatically identifies and selects the connected region to which the coordinate point belongs when the user clicks on the target area; The intelligent selection tool automatically identifies and selects multiple independent targets within a rectangle by having the user drag it.

[0015] In some embodiments, a seamless transition from completely transparent to completely opaque overlay is achieved by dragging a slider.

[0016] In some embodiments, the standard format file is in COCOJSON or GeoJSON format so that it can be directly used for quantitative analysis, statistical reporting, or training of deep learning models.

[0017] The present invention also provides an interactive scanning image intelligent annotation device based on panoramic images, for implementing the method described above, the system comprising: The scanning control module is used to control the scanner to perform a sequence of redundant scans with a predetermined overlap rate, acquire scanned sub-images in real time, perform image registration, overlap area verification and scan position calibration for each sub-image, crop the standard effective area and record the metadata of each sub-image to a structured index. The panoramic construction module receives metadata and effective area sub-images from the scanning control module. Based on the metadata, it uses image registration and fusion algorithms to stitch the sub-images into a high-resolution seamless panoramic digital slide image. The pre-annotation module is used to load the pre-trained segmentation model, perform batch inference on the effective region sub-images corresponding to the panoramic image output by the panoramic construction module to generate target category labels, bounding boxes and binary masks, perform semantic encoding on the categories, map the target masks to the panoramic global coordinate system and synthesize logical annotation layers, and finally generate structured annotation files. The annotation processing module is an interactive annotation processing module that provides intelligent interactive tools, receives annotation operation instructions from users, verifies, modifies and updates the structured annotation files generated by the pre-annotation module, records all annotation operation sequences, and supports multi-step undo and redo functions. The visualization rendering module is used to dynamically display panoramic images and annotation information using a dual-layer rendering architecture. The bottom layer renders panoramic digital slide images, and the upper layer parses structured annotation files and generates annotation overlays based on the visible area. The annotation overlays are rendered using an encoding method that decouples RGB and Alpha channels, and a front-end interactive control is provided to adjust the global transparency of the annotation overlays in real time. The results output module is used to optimize and verify the structured annotation data file updated by the annotation processing module, and export it as a standard format annotation result file.

[0018] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0019] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0020] The interactive scanning image intelligent annotation method and system based on panoramic images provided by this invention actively performs computationally overlapping scans and simultaneously records the precise location and metadata of each sub-image, providing a foundation for subsequent high-precision, lossless image stitching driven by metadata, fundamentally ensuring the integrity and accuracy of the source data used for annotation. By changing the annotation reference system, all annotations are directly anchored to the coordinate system of the seamless panoramic image, rather than discrete sub-images, thus solving the problems of edge omission and duplicate annotation. By freeing users from tedious precise outlining, they can input with their rough strokes, and the background image segmentation algorithm will output pixel-level precision target masks in real time, improving the annotation efficiency of irregular targets. Dual-layer rendering is adopted, separating color coding semantics (category) from layer visibility control (transparency), and providing real-time adjustment capabilities through a front-end slider, achieving a dynamic balance optimization between annotation display and original image viewing. AI automatic annotation, artificial intelligence revision tools (such as intelligent brushes), and fine-grained editing functions (such as point selection, box selection, and smearing / clearing) are integrated into a unified panoramic view to achieve efficient collaboration. Attached Figure Description

[0021] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0022] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0023] Figure 1 A flowchart of the interactive scanned image intelligent annotation method based on panoramic images provided by the present invention; Figure 2 This is a structural block diagram of the interactive scanning image intelligent annotation device based on panoramic images provided by the present invention; Figures 3-7 This is a simulation diagram of the effect in one embodiment of the present invention; Figure 8 This is a structural block diagram of a computer device provided by the present invention. Detailed Implementation

[0024] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] To address the shortcomings of the existing technologies, the present invention aims to provide a method and system for pre-annotation and manual annotation of slide scan images. This method and system are particularly suitable for applications requiring high-resolution panoramic display and precise annotation, such as digital pathological slides and geological sample slides. It can solve the problem of image information loss caused by edge cropping and scanning misalignment, ensuring the integrity of sample information and breaking the limitations of discrete image annotation. At the same time, by providing a panoramic annotation environment that supports arbitrary scaling and translation, it achieves the unification of annotation perspective and observation perspective. By providing intelligent annotation interaction tools, it reduces the operational complexity of annotating irregular targets and significantly improves annotation efficiency and accuracy.

[0026] In one specific implementation, such as Figure 1 As shown, the interactive scanned image intelligent annotation method based on panoramic images provided by this invention includes the following steps: S110: Based on a predetermined overlap rate, the target slide is subjected to a sequential redundant scan by a scanner, and the metadata of each sub-image obtained by the scan is recorded. Based on the metadata, all sub-images are stitched together into a high-resolution panoramic slide image. That is, by controlling the scanner to perform a sequential scan with a predetermined overlap rate (i.e., redundant scan), and synchronously recording the coordinate offset and size of each sub-image, the image stitching algorithm is driven based on this metadata to generate a seamless panoramic base map for annotation.

[0027] Specifically, step S110 provides a method for constructing and representing panoramic images. Through collaborative optimization of hardware and software, a lossless panoramic digital slide image is constructed. During overlapping scanning and metadata recording, the scanner is controlled to perform redundant scanning of overlapping areas and accurately record metadata such as coordinate offset and rotation angle of each sub-image. Based on the recorded metadata, existing image registration and fusion algorithms are used to seamlessly stitch all sub-images into a complete, high-resolution panoramic digital slide image. This panoramic image supports a smooth zooming and panning browsing experience.

[0028] S120: Batch reasoning is performed on sub-images within the effective area corresponding to the panoramic slide image using a pre-trained segmentation model to obtain target annotation information; semantic encoding is performed on the annotation categories, and the target annotation information is mapped to the global coordinate system of the panoramic slide image to obtain the annotation layer; S130: Utilizing a pre-built dual-layer rendering architecture, dynamic rendering of the panoramic slide image is performed. This dual-layer rendering architecture includes a label overlay layer, which employs a decoupled encoding method for RGB and Alpha channels. Real-time interactive adjustment of the global transparency of the label overlay layer is achieved through front-end controls. In the panoramic view, intelligent interactive tools verify and correct the pre-labeling results and record all labeling operation sequences to obtain a structured labeling data file. In other words, a dual-layer rendering technology is used, where the labeling layer uses a decoupled encoding method for color and transparency (RGB color encoding categories, independent Alpha channels controlling global transparency), and user interface controls (such as sliders) are provided to achieve real-time interactive adjustment of the overall transparency of the labeling layer. Steps S120 and S130 of the method provided by this invention offer an interactive labeling method based on a panoramic image, providing a flexible and efficient labeling interaction mechanism on top of the panoramic image. Specifically, in the labeling within the panoramic context, users operate directly on the seamless panoramic image, fundamentally avoiding edge omissions and repeated labeling problems caused by image discretization. The labeling data is bound to specific coordinate positions in the panoramic image. The diverse annotation tools include intelligent point selection annotation, which allows users to click on a target area and the algorithm automatically identifies and selects the connected region to which the coordinate point belongs; intelligent box selection annotation, which allows users to drag a rectangle and the system automatically identifies and selects multiple independent targets within the box; and brush smudge annotation, which assists users in roughly marking targets by smudges.

[0029] Furthermore, this invention provides a visualization method based on interactive dual-layer RGBA fusion. Specifically, to achieve optimal fusion and display of annotation results with the original image, this invention employs an innovative dual-layer rendering architecture. The system maintains an original image layer and an annotation mask layer on the front end. Annotation information on the mask layer corresponds to different categories with different RGB color values. During decoupled visual encoding, color (RGB) and transparency (Alpha) are decoupled. RGB values ​​are dedicated to encoding category semantic information, ensuring high contrast distinction between categories; the Alpha channel value independently controls the overall transparency of all annotation areas. In the dynamically adjustable user interface, the front end provides real-time transparency adjustment controls (such as sliders). Users can dynamically adjust the Alpha value through interactive operations, achieving a seamless transition of the annotation overlay layer from completely transparent to completely opaque. This allows users to personalize the balance between image detail viewing and annotation area visibility according to their current observation needs, fundamentally solving the problem that static overlay layers cannot adapt to different verification scenarios. Regarding operation history and undo, the system automatically records the user's complete annotation operation sequence from start to finish (such as adding, modifying, and deleting). This mechanism supports multi-step undo and redo functions, providing users with a very high operational fault tolerance rate, allowing them to easily revert to any historical state, effectively improving the reliability and user experience of complex annotation tasks. For precise clearing, it provides clearing tools corresponding to the annotation tools (such as point clearing, box selection clearing, and smear clearing) to achieve precise annotation modifications.

[0030] S140: After optimizing and verifying the final structured annotation data file, export it as a standard format file containing target category information and accurate vector contours in the panoramic coordinate system.

[0031] In general, the system used in the method provided by this invention includes a scanning control module for redundant scanning and metadata recording, a panoramic construction module for generating panoramic images, an annotation processing module for implementing interactive functions such as intelligent brushes, and a visualization rendering module for dynamic display. During operation, the panoramic image is first automatically pre-annotated. Then, using an interactive interface integrating intelligent brushes and dynamic display functions, the pre-annotation results are verified and corrected within the panoramic context. Through a pre-trained segmentation model, the slides are automatically pre-annotated, obtaining initial annotation results. Users only need to perform verification and correction, significantly reducing manual workload. All annotation results are integrated into the panoramic view, allowing users to freely zoom and pan the panoramic image, much like operating a digital microscope, to intuitively view and evaluate the overall annotation effect.

[0032] Thus, the method provided by this invention effectively solves the problems of limited field of view, stitching gaps, and insufficient resolution in single sub-image scanning by using redundant scanning with a predetermined overlap rate and synchronous recording of metadata. This ensures that the stitched panoramic digital slide image has high resolution and seamless connection, fully preserving all target details on the slide and providing a high-quality image foundation for subsequent accurate annotation. The application of batch inference of the pre-trained instance segmentation model greatly improves annotation efficiency and avoids the tedious operation of manual region-by-region annotation. Semantic encoding and global coordinate system mapping ensure the accuracy and consistency of annotation information, and the structured annotation file facilitates data storage, retrieval, and subsequent processing. The dual-layer rendering architecture and decoupled alpha channel encoding enable clear distinction between panoramic images and annotation overlays. Real-time transparency adjustment allows users to flexibly switch between viewing original image details and annotation information. Intelligent interactive tools and operation sequence recording reduce the difficulty of manual verification and correction, minimizing annotation errors. Multi-step undo and redo functions enhance the flexibility and convenience of annotation operations. Finally, the export of standard format files ensures the universality of annotation data, directly adapting to various subsequent application scenarios such as quantitative analysis, statistical report generation, and deep learning model training. Overall, this improves the efficiency, accuracy, and practicality of image annotation, while reducing the cost and barrier to manual annotation.

[0033] In some embodiments, based on a predetermined overlap rate, a sequence of redundant scans of the target slide is performed using a scanner, specifically including: The scanner expands the actual capture area of ​​a single scan and sets overlapping areas in both the horizontal and vertical directions. The scanner automatically performs sequential scanning with the upper left corner of the target slide as the origin and a preset serpentine path. During the scanning process, each sub-image undergoes real-time image registration, overlapping area verification, and scanning position calibration. After verification, the standard valid area is cropped out, and the metadata is recorded in the structured index.

[0034] In this way, overlapping areas are set both horizontally and vertically during the scanning process, providing a sufficient matching basis for subsequent sub-image stitching, effectively avoiding problems such as gaps and misalignments during stitching, and improving the stitching accuracy and integrity of the panoramic image; the automatic sequential scanning of the serpentine path, compared with other scanning paths, can maximize the use of the scanning area, reduce scanning blind spots and redundant scanning, and improve scanning efficiency. At the same time, setting the upper left corner of the slide as the origin provides a unified benchmark for the unified recording of sub-image metadata and coordinate system calibration, ensuring the accuracy of the position information of each sub-image; real-time image registration, overlapping area verification and scanning during the scanning process Position calibration can promptly detect and correct problems such as positional offset and non-compliance rate during scanning, avoiding the failure of the entire panoramic image stitching or the decrease in accuracy due to the scanning deviation of a single sub-image, thus improving the stability and reliability of the scanning process. The cropping of the standard effective area can remove invalid areas at the edges of the sub-image (such as scanning noise, slide edge impurities, etc.), ensuring the effectiveness and consistency of each sub-image and providing high-quality image input for subsequent annotation inference. The structured index record of metadata facilitates subsequent sub-image stitching, annotation information mapping and data traceability, further improving the standardization and operability of the entire annotation method.

[0035] Furthermore, real-time image registration is performed on each sub-image, specifically including: Extract key points and feature descriptors from adjacent subgraphs; for example, the SIFT algorithm can be used to extract key points and feature descriptors. The Euclidean distance between feature descriptors can be calculated using a brute-force matching algorithm to obtain matching point pairs; for example, the Euclidean distance can be calculated using a brute-force matching algorithm. Filter out mismatches and fit subpixel-level transformation relationships to parse out the horizontal and vertical pixel offsets; for example, the RANSAC algorithm can be used for error filtering.

[0036] It should be understood that the SIFT algorithm possesses strong scale invariance and rotation invariance, enabling it to stably extract representative key points and feature descriptors from sub-images under different lighting and scaling ratios. This ensures the accuracy and stability of feature matching between adjacent sub-images, solving the problems of low matching accuracy and poor robustness of traditional registration algorithms in complex image scenes. The brute-force matching algorithm obtains matching point pairs by calculating the Euclidean distance between descriptors. Its simple logic and high execution efficiency allow for rapid matching of feature descriptors from adjacent sub-images, meeting the needs of real-time registration and avoiding the decrease in scanning efficiency caused by complex matching algorithms. R The ANSAC algorithm effectively filters out mismatched point pairs generated during the matching process, eliminates interference from noise, image impurities, and other factors on the registration results, and improves the purity of the matching point pairs. At the same time, it fits sub-pixel level transformation relationships and resolves accurate pixel offsets, improving the registration accuracy to the sub-pixel level, which is far higher than the traditional pixel-level registration accuracy. It effectively solves problems such as misalignment and ghosting that occur during the stitching of adjacent sub-images, ensuring seamless connection and high-resolution quality of panoramic digital slide images. It provides a reliable positional reference for the accurate mapping of subsequent annotation information in the global coordinate system, further improving the annotation accuracy of the entire annotation method.

[0037] The segmentation model is Mask R-CNN, whose network architecture includes a feature extraction backbone network, a region proposal network, and parallel classification, bounding box regression, and mask prediction heads. The structured annotation file is in JSON or GeoJSON format. Choosing Mask R-CNN as the segmentation model leverages its feature extraction backbone network to effectively extract deep features from images, improving the accuracy of target recognition. The region proposal network quickly generates potential target regions, reducing unnecessary computation and improving inference efficiency. The parallel classification, bounding box regression, and mask prediction heads simultaneously perform target category identification, precise target bounding box localization, and pixel-level segmentation of target contours. Compared to traditional segmentation models, this significantly improves annotation accuracy and efficiency, accurately identifying different categories of targets in panoramic images and obtaining precise target contours, thus solving the problems of inaccurate target localization and blurred contours in traditional annotation methods. Furthermore, the structured annotation files adopt JSON or GeoJSON format. Both formats are universal lightweight data exchange formats with strong readability, good scalability, and cross-platform compatibility. They facilitate the storage, transmission, and parsing of annotation data. JSON format is suitable for storing structured annotation information, which facilitates the quick retrieval and modification of subsequent data. GeoJSON format is better adapted to the storage of spatial coordinate information, ensuring the accuracy of the position of annotation information in the panoramic global coordinate system. The choice of the two formats can meet the needs of different subsequent application scenarios, improve the universality and practicality of annotation data, and provide convenience for subsequent data verification, export, and application.

[0038] Specifically, the intelligent interaction tools may include an intelligent brush tool, a precise cleanup tool, an intelligent point selection tool, and an intelligent bounding box selection tool. The intelligent brush tool receives rough brushstrokes from the user on the panoramic image. A background image segmentation algorithm identifies foreground / background regions based on the brushstroke information, calculates and generates pixel-level precision target contour masks in real time, and updates the annotation mask layer. The background algorithm automatically identifies the foreground / background regions indicated by the brushstrokes and calculates and generates pixel-level precision target contour masks in real time. The precise cleanup tool includes point-based cleanup, bounding box cleanup, and brush-based cleanup, corresponding to the annotation tools to achieve precise modification of annotations. The intelligent point selection tool automatically identifies and selects the connected region to which the user clicks on the target area. The intelligent bounding box selection tool automatically identifies and selects multiple independent targets within a bounding box by having the user drag a rectangle.

[0039] In this way, the intelligent brush tool eliminates the need for users to perform detailed outlining; it generates pixel-level precision target contour masks simply by roughly applying the brush, significantly reducing the difficulty of manual annotation and saving annotation time. Simultaneously, it updates the annotation mask layer in real time, ensuring users can promptly view the annotation results, thus improving the convenience and efficiency of annotation operations. The precise clearing tool offers multiple clearing methods: point clearing precisely removes individual annotation details, box selection clearing quickly removes annotations within a specified area, and smear clearing flexibly removes annotations in irregular areas. Complementing the annotation tools, it enables precise annotation modifications, effectively solving the problem of difficult-to-correct mis-annotations and redundant annotations, reducing annotation errors, and improving annotation accuracy. The point selection tool allows users to select connected regions of a target with a simple click, avoiding the tedious manual selection or shading process. It is especially suitable for quickly selecting small or scattered targets, improving operational efficiency. The intelligent bounding box tool can select multiple independent targets within a bounding box, suitable for quickly selecting and batching large-scale, centrally distributed targets. This significantly saves users time compared to selecting targets one by one, improving the efficiency of annotation operations. The combination of various intelligent interactive tools covers the entire process of annotation, modification, and selection, adapting to different types and distributions of target annotation scenarios. It balances ease of operation and annotation accuracy, further lowering the threshold for manual verification and correction, and improving the efficiency and user experience of the entire annotation process.

[0040] Optionally, the control for adjusting the transparency of the annotation overlay is a slider, which allows for a seamless transition from completely transparent to completely opaque by dragging the slider. The standard format file is in COCOJSON or GeoJSON format, so that the standard format file can be directly used for quantitative analysis, statistical reporting, or training of deep learning models.

[0041] Theoretically, using a slider as the transparency adjustment control offers a simple and intuitive operation logic. Users can quickly adjust the transparency of the annotation overlay by dragging the slider, eliminating the need for complex parameter settings, thus lowering the user's operational threshold and improving the interactive experience. Achieving a seamless transition from completely transparent to completely opaque annotation overlays meets the needs of users in different viewing scenarios: when focusing on viewing the original details of the panoramic image and verifying the accuracy of annotation positions, the transparency can be adjusted to a lower level (close to complete transparency) to prevent the annotation overlay from obscuring the original image; when focusing on viewing annotation information and verifying the accuracy of annotation outlines, the transparency can be adjusted to a higher level (close to complete opacity) to clearly display annotation details; when viewing the original image and annotation information simultaneously for comparison and verification, the transparency can be adjusted to a suitable intermediate level to achieve a clear overlay display of the original image and annotation information. This seamless adjustment method ensures smooth transparency changes, avoiding visual discomfort caused by abrupt adjustments, while also allowing users to accurately find the appropriate transparency level, improving the convenience and accuracy of annotation verification, and indirectly improving the efficiency of the entire annotation process.

[0042] Understandably, COCOJSON is a common annotation data format in the field of deep learning, widely supporting various deep learning frameworks (such as TensorFlow and PyTorch). Exporting annotation data to COCOJSON format allows for direct use in the training, validation, and testing of deep learning models without format conversion, saving time and cost in data preprocessing and improving data usability and adaptability. GeoJSON format possesses excellent spatial coordinate information storage and parsing capabilities, accurately preserving the position and contour information of annotation information in a panoramic coordinate system. It is suitable for quantitative analysis of annotation data (such as calculation of parameters like target area, quantity, and distribution density) and statistical report generation, ensuring the accuracy of quantitative analysis and statistical results. The choice of these two standard formats covers various subsequent application scenarios such as deep learning training, quantitative analysis, and statistical reporting, improving the universality and reusability of annotation data, avoiding the problem of data being unusable due to format incompatibility, and expanding the application scope of the entire annotation method. At the same time, the standard format file contains target category information and precise vector contours in a panoramic coordinate system, ensuring the integrity and accuracy of data in subsequent applications, further enhancing the practical value of the entire annotation method.

[0043] In the above specific embodiments, the interactive scanning image intelligent annotation method based on panoramic images provided by this invention actively performs computationally overlapping scans and simultaneously records the precise location and metadata of each sub-image, providing a foundation for subsequent high-precision, lossless image stitching driven by metadata, fundamentally ensuring the integrity and accuracy of the source data used for annotation. By changing the annotation reference system, all annotations are directly anchored to the coordinate system of the seamless panoramic image, rather than discrete sub-images, to solve the problems of edge omission and duplicate annotation. By freeing users from tedious precise outlining, and allowing them to input with their rough strokes, the background image segmentation algorithm outputs pixel-level precision target masks in real time, improving the annotation efficiency of irregular targets. Dual-layer rendering is adopted, separating color coding semantics (category) from layer visibility control (transparency), and providing real-time adjustment capabilities through a front-end slider, achieving a dynamic balance optimization between annotation display and original image viewing. AI automatic annotation, artificial intelligence revision tools (such as intelligent brushes), and refined editing functions (such as point selection, box selection, and smearing / clearing) are integrated into a unified panoramic view to achieve efficient collaboration.

[0044] In addition to the methods described above, this invention also provides an interactive intelligent annotation device for scanned images based on panoramic images, such as... Figure 2 As shown, the system includes: The scanning control module 210 is used to control the scanner to perform a sequence of redundant scans with a predetermined overlap rate, acquire scanned sub-images in real time, perform image registration, overlap area verification and scan position calibration for each sub-image, crop the standard effective area and record the metadata of each sub-image to a structured index. The panoramic construction module 220 is used to receive metadata and effective area sub-images from the scanning control module, and stitch the sub-images into a high-resolution seamless panoramic digital slide image based on the metadata using image registration and fusion algorithms. The pre-annotation module 230 is used to load the pre-trained instance segmentation model, perform batch inference on the effective region sub-images corresponding to the panoramic image output by the panoramic construction module to generate target category labels, bounding boxes and binary masks, perform semantic encoding on the categories, map the target masks to the panoramic global coordinate system and synthesize logical annotation layers, and finally generate structured annotation files. The annotation processing module 240 is an interactive annotation processing module used to provide intelligent interactive tools, receive annotation operation instructions from users, verify, modify and update the structured annotation files generated by the AI ​​pre-annotation module, record all annotation operation sequences, and support multi-step undo and redo functions. The visualization rendering module 250 is used to realize the dynamic display of panoramic images and annotation information using a dual-layer rendering architecture. The bottom layer renders panoramic digital slide images, and the upper layer parses structured annotation files and generates annotation overlays based on the visible area. The annotation overlays are rendered using an encoding method that decouples RGB and Alpha channels, and a front-end interactive control is provided to realize real-time adjustment of the global transparency of the annotation overlays. The output module 260 is used to optimize and verify the structured annotation data file updated by the annotation processing module, and export it as a standard format annotation result file.

[0045] In some embodiments, performing a sequence redundancy scan with a predetermined overlap rate on the slide specifically includes: The actual capture area of ​​a single scan by the scanner is expanded to set overlapping areas in both the horizontal and vertical directions. The scanner performs automatic sequential scanning with the upper left corner of the slide as the origin and a preset serpentine path. During the scanning process, each sub-image undergoes real-time image registration, overlapping area verification, and scanning position calibration. After verification, the standard valid area is cropped out, and the metadata is recorded in the structured index.

[0046] In some embodiments, real-time image registration is performed for each sub-image, specifically including: The SIFT algorithm is used to extract key points and feature descriptors of adjacent subgraphs; Matching point pairs are obtained by calculating the Euclidean distance between descriptors using a brute-force matching algorithm; The RANSAC algorithm is used to filter out mismatches and fit sub-pixel level transformation relationships to parse out the horizontal and vertical pixel offsets.

[0047] In some embodiments, the instance segmentation model is a Mask R-CNN model, the network architecture of which includes a feature extraction backbone network, a region proposal network, and parallel classification, bounding box regression, and mask prediction heads; the structured annotation file is in JSON or GeoJSON format.

[0048] In some embodiments, the intelligent interaction tool includes: The intelligent brush tool receives rough brush strokes from the user on the panoramic image. The background image segmentation algorithm identifies the foreground / background regions based on the brush stroke information, calculates and generates a pixel-level precision target contour mask in real time, and updates it to the annotation mask layer. The precision removal tool includes point removal, box selection removal, and smear removal, which correspond to the annotation tool to enable precise modification of annotations; A smart point selection tool that automatically identifies and selects the connected region to which the coordinate point belongs when the user clicks on the target area; The intelligent selection tool automatically identifies and selects multiple independent targets within a rectangle by having the user drag it.

[0049] In some embodiments, the control for adjusting the transparency of the annotation overlay is a slider, which allows for a seamless transition of the annotation overlay from completely transparent to completely opaque by dragging the slider.

[0050] In some embodiments, the standard format file is in COCOJSON or GeoJSON format so that it can be directly used for quantitative analysis, statistical reporting, or training of deep learning models.

[0051] The interactive scanning image intelligent annotation device based on panoramic images provided by this invention actively performs computationally overlapping scans and simultaneously records the precise location and metadata of each sub-image, providing a foundation for subsequent high-precision, lossless image stitching driven by metadata, fundamentally ensuring the integrity and accuracy of the source data used for annotation. By changing the annotation reference system, all annotations are directly anchored to the coordinate system of the seamless panoramic image, rather than discrete sub-images, thus solving the problems of edge omission and duplicate annotation. By freeing users from tedious precise outlining, and allowing them to input with their rough strokes, the background image segmentation algorithm outputs pixel-level precision target masks in real time, improving the annotation efficiency of irregular targets. Dual-layer rendering is employed, separating color coding semantics (category) from layer visibility control (transparency), and providing real-time adjustment capabilities through a front-end slider, achieving a dynamic balance optimization between annotation display and original image viewing. AI automatic annotation, artificial intelligence revision tools (such as intelligent brushes), and refined editing functions (such as point selection, box selection, and smearing / clearing) are integrated into a unified panoramic view to achieve efficient collaboration.

[0052] To facilitate understanding, the following will be combined with... Figure 3-7 The overall technical solution and technical effects of the method provided by the present invention will be described by way of example.

[0053] This embodiment uses the panoramic display and annotation of digital images of microbial slides as an example to explain in detail the implementation process of the present invention. Figure 3 A simulation interface diagram for pre-labeling "Gram-negative bacteria-bacilli-long bacilli" in the algorithm; Figure 4 A simulation interface diagram for intelligent selection and labeling of "white blood cells"; Figure 5 A simulation interface diagram showing the completion of the intelligent selection annotation "Mycobacteria"; Figure 6 A simulated interface for using a paintbrush to paint over areas marked "fungus".

[0054] The system provided by this invention mainly includes a scanning control module, a pre-annotation module, a panoramic image construction module, an interactive annotation processing module, and a visualization rendering module. These modules work together to form a complete processing loop from data acquisition to result output. The specific steps are as follows: S1: Panoramic map construction based on overlapping scans and metadata records. This step is executed collaboratively by the scan control module and the panoramic image construction module. It aims to generate a high-precision, seamless panoramic base map for annotation and its accurate spatial index through sequential scans with preset redundant regions.

[0055] S101: Overlap Scanning Strategy and Execution; Drives the scanner hardware, employing a strategy of actively expanding the scanning area. The actual capture area of ​​a single scan is expanded from the standard effective area (1080×1800 pixels) to (1216×1936 pixels), adding an overlap area of ​​136 pixels both horizontally and vertically. The system establishes a global coordinate system with the upper left corner of the slide as the origin, controlling the scanner to automatically scan sequentially along a preset serpentine path, ensuring sufficient physical overlap between adjacent scan sites to provide data redundancy for subsequent image registration and stitching.

[0056] S102: Real-time image registration, verification, and metadata recording; During the scanning process, the system processes each newly acquired sub-image in real time to ensure data quality and provide accurate positioning information for panorama construction. The core process is as follows: 1. Precise Offset Calculation: A feature-based image registration algorithm is used to calculate the relative positions between adjacent sub-images. The specific process is as follows: For two adjacent original scanned images (1216×1936 pixels), the SIFT algorithm is used to extract scale-invariant keypoints and feature descriptors; subsequently, a brute-force matching algorithm is used to calculate the Euclidean distance between the descriptors to find matching point pairs; finally, the RANSAC algorithm is used for robust estimation, filtering out false matches, and fitting the sub-pixel-level transformation relationship between the two images to parse the precise horizontal and vertical pixel offsets (offset_x, offset_y).

[0057] 2. Overlapping Area Verification and Scanning Position Calibration: Based on the calculated offset, it is determined whether the actual overlapping area reaches the preset threshold. If it does, the cropping process begins; if not, the error is immediately reported, and the scanner is controlled to fine-tune its mechanical position for local rescanning, ensuring the geometric continuity of the global stitching from the source.

[0058] 3. Effective Region Cropping and Metadata Storage: For images that pass verification, a standard effective region (1080×1800 pixels) is cropped from the original image based on its precise offset relationship. Simultaneously, key metadata such as the effective region file path, its logical position (row and column number) in the global coordinate system, physical coordinates, size, and offset are recorded in a structured index, laying the foundation for subsequent panoramic spatial mapping.

[0059] S2: AI Pre-annotation and Annotation Layer Initialization. S2 details how to use the Mask R-CNN instance segmentation algorithm to automatically pre-annotate scanned sub-images and generate an annotation mask layer aligned with the panoramic image coordinates.

[0060] S201: Model Loading and Inference; This step is performed by the AI ​​pre-annotation module, aiming to utilize advanced deep learning models for pixel-level recognition of sub-images. Its core process is as follows: Model Loading: The system loads a pre-trained Mask R-CNN model. This model typically includes a feature extraction backbone network, a region proposal network, and parallel classification, bounding box regression, and mask prediction heads.

[0061] Batch Inference and Instance Segmentation: The model can perform batch inference on each valid region sub-image obtained in step S102. For each sub-image, Mask R-CNN operates according to the following process: Region proposal: First, the region proposal network generates multiple candidate regions in the image that may contain the target.

[0062] Target identification and localization: Then, the model classifies each candidate region (determining whether it is "background" or a specific microbial category, such as "Gram-negative bacteria", "white blood cells", etc.) and performs bounding box regression (refining the location of the target).

[0063] Pixel-level mask generation: The most crucial step is that the mask prediction head generates a binary pixel-level mask for each confirmed target. This mask delineates the irregular contours of the target with extremely high precision, achieving instance-level segmentation.

[0064] S202: Data Structuring, Encoding, and Coordinate Mapping; This step converts the Mask R-CNN model output into structured annotation information that is usable by the system and bound to panoramic coordinates. The core process includes three sub-steps: 1. Instance Information Extraction and Semantic Encoding: The AI ​​pre-annotation module receives the model output (category label, bounding box, and binary mask for each target). The system assigns a unique semantic ID and a corresponding display color value to each predefined category (e.g., the ID for the category "Mycobacterium" is 1, displayed in red [255,0,0]; the ID for the category "Fungus" is 2, displayed in green [0,255,0]). This step completes the mapping from text categories to computer-processable codes.

[0065] 2. Coordinate Mapping and Logical Layer Composition: Using the absolute coordinates of each sub-image recorded in S102 within the panorama, the binary mask of each target instance is mapped onto the global coordinate system of the panorama. At this point, the system logically "combines" the masks of all instances according to their coordinates onto a virtual "logical annotation layer" the same size as the panorama image. The value of each pixel position is determined by the semantic color value of its corresponding instance. However, please note that this "logical layer" is not the final storage format.

[0066] 3. Structured Annotation File Generation: To facilitate efficient storage and subsequent interaction, the system converts the aforementioned "logical layer" information into a lightweight structured format. Specifically, the binary mask of each target is converted into a contour polygon, and along with the target's semantic ID, panoramic coordinates, and other attributes, it is stored as a record in a structured annotation file (such as JSON / GeoJSON). This file is essentially an equivalent, optimized, and operable data representation of the "panoramic annotation layer."

[0067] S3: Interactive Panoramic Annotation and Visualization. This step is the core of user interaction and is jointly implemented by the interactive annotation processing module and the visualization rendering module on the front-end interface.

[0068] S301: Dual-layer rendering and dynamic display; the visualization rendering module implements innovative dual-layer dynamic rendering on the front end: Low-level rendering: Dynamically loads and renders the panoramic image generated by S1.

[0069] Upper-layer rendering: Instead of loading a pre-generated giant mask image, the system parses the structured annotation file generated by S2 in real time. Based on the current visible area, the system retrieves relevant target instances from the file in real time and draws their polygonal outlines onto the upper-layer canvas according to the colors corresponding to their categories, thereby dynamically generating an "annotation overlay layer".

[0070] The system uses the RGBA color model to render the annotation layer. Its innovation lies in decoupling color (RGB) from transparency (Alpha): RGB values ​​are strictly used to encode category semantic information, while the Alpha channel value uniformly controls the transparency of the entire annotation layer.

[0071] The front-end interface provides a transparency adjustment control (such as a slider). By dragging the slider, users can adjust the global alpha value of the annotation layer in real time and continuously, achieving a seamless transition from completely transparent (for easy viewing of image details) to completely opaque (for easy focusing on the annotation area).

[0072] S302: Manual Verification and Precise Correction. Users can use the interactive tools provided by the system to verify and correct the AI ​​pre-annotation results in the panoramic view.

[0073] Smart Brush Tool: For irregular targets that are missed or mislabeled by the AI, users can select the smart brush. Users simply use the mouse to roughly paint over the target area and its surroundings, indicating the foreground (target) and background. The background image segmentation algorithm will calculate and generate the target outline with pixel-level precision in real time based on this sparse painting information and update it to the annotation mask layer.

[0074] Precise removal tools: For AI over-labeling or erroneous annotations, users can use corresponding removal tools (such as point removal, box selection removal, and smear removal) to precisely erase them.

[0075] Operation history management: The system automatically records every marked operation (addition, modification, deletion) performed by the user. Users can easily revert to the state before any operation using the "undo" and "redo" functions, greatly improving the fault tolerance of operations.

[0076] S303: Annotation Management in a Panoramic Context. Because all editing operations are performed within a unified "panoramic global coordinate system" and directly modify a unified structured annotation data source, the annotation status seen by the user remains consistent throughout any zooming and panning process. This fundamentally eliminates the problems common in traditional sub-map annotation modes, such as fragmented targets at the edges of sub-maps, duplicate annotations of the same target in adjacent areas, or omissions in annotation.

[0077] S4: Annotation Result Output. After the user completes interactive verification and correction, the system outputs the final annotation result.

[0078] Data integration: The system optimizes and verifies the final structured labeled data file.

[0079] Export Format: This file can be directly exported to a standard format (such as COCO JSON, GeoJSON, etc.), which includes the category information of each target instance and its precise vector contour in the panoramic coordinate system. This file can be directly used for subsequent quantitative analysis, statistical reporting, or training of deep learning models.

[0080] Through S1-S4 above, this invention constructs a closed-loop system based on panoramic coordinates. S1 establishes a high-precision panoramic base map and coordinate index through overlapping scanning and real-time registration; S2 utilizes AI for batch pre-annotation and unifies the results to the panoramic space through coordinate mapping, generating structured annotation data; S3 uses dynamic rendering and intelligent interactive tools to enable efficient and accurate manual verification within a unified panoramic context; finally, S4 outputs standardized vector annotation results. This system effectively solves the core problems of low efficiency, cumbersome operation, and inconsistent results in the annotation of large-size glass slide images.

[0081] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and model predictions. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The model predictions of the computer device store static and dynamic information data. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0082] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0083] Corresponding to the above embodiments, this invention also provides a computer storage medium containing one or more program instructions. These one or more program instructions are used to execute the method described above.

[0084] The present invention also provides a computer program product, the computer program product including a computer program, the computer program being stored on a non-transitory computer-readable storage medium, and the computer being able to perform the above-described method when the computer program is executed by a processor.

[0085] In this embodiment of the invention, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0086] The various methods, steps, and logic diagrams disclosed in the embodiments of this invention can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor reads information from the storage medium and, in conjunction with its hardware, completes the steps of the above methods.

[0087] The storage medium can be memory, such as volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0088] Among them, non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory.

[0089] Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (Synchlink DRAM, SLDRAM), and direct memory bus RAM (DRRAM).

[0090] The storage media described in the embodiments of the present invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0091] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using a combination of hardware and software. When applied as software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of computer programs from one place to another. Storage media can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0092] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. An interactive intelligent annotation method for scanned images based on panoramic images, characterized in that, The method includes: Based on a predetermined overlap rate, the target slide is subjected to a sequential redundant scan using a scanner, and the metadata of each sub-image obtained from the scan is recorded. Based on the metadata, all sub-images are stitched together into a high-resolution panoramic slide image. The sub-images within the effective region corresponding to the panoramic slide image are batch-inferred using a pre-trained segmentation model to obtain target annotation information; the annotation categories are semantically encoded, and the target annotation information is mapped to the global coordinate system of the panoramic slide image to obtain the annotation layer; The pre-built dual-layer rendering architecture is used to dynamically render panoramic slide images. The dual-layer rendering architecture includes a label overlay layer, which adopts an encoding method that decouples RGB and Alpha channels. Real-time interactive adjustment of the global transparency of the label overlay layer is achieved through front-end controls. In the panoramic view, the pre-labeling results are verified and corrected through intelligent interactive tools, and all labeling operation sequences are recorded to obtain a structured labeling data file. After optimizing and validating the final structured annotation data file, it is exported as a standard format file containing target category information and accurate vector contours in the panoramic coordinate system.

2. The method according to claim 1, characterized in that, Based on a predetermined overlap rate, the target slide is subjected to sequential redundant scanning using a scanner, specifically including: The scanner expands the actual capture area of ​​a single scan and sets overlapping areas in both the horizontal and vertical directions. The scanner automatically performs sequential scanning with the upper left corner of the target slide as the origin and a preset serpentine path. During the scanning process, each sub-image undergoes real-time image registration, overlapping area verification, and scanning position calibration. After verification, the standard valid area is cropped out, and the metadata is recorded in the structured index.

3. The method according to claim 2, characterized in that, Real-time image registration is performed for each sub-image, specifically including: Extract key points and feature descriptors from adjacent subgraphs; Calculate the Euclidean distance between feature descriptors to obtain matching point pairs; Filter out mismatches and fit subpixel-level transformation relationships to parse out the horizontal and vertical pixel offsets.

4. The method according to claim 1, characterized in that, The network architecture of the segmentation model includes a feature extraction backbone network, a region proposal network, and parallel classification, bounding box regression, and mask prediction heads.

5. The method according to claim 1, characterized in that, The intelligent interaction tools include: The intelligent brush tool receives rough brush strokes from the user on the panoramic image. The background image segmentation algorithm identifies the foreground / background regions based on the brush stroke information, calculates and generates a pixel-level precision target contour mask in real time, and updates it to the annotation mask layer. The precision removal tool includes point removal, box selection removal, and smear removal, which correspond to the annotation tool to enable precise modification of annotations; A smart point selection tool that automatically identifies and selects the connected region to which the coordinate point belongs when the user clicks on the target area; The intelligent selection tool automatically identifies and selects multiple independent targets within a rectangle by having the user drag it.

6. The method according to claim 1, characterized in that, Achieve a seamless transition from completely transparent to completely opaque annotation overlay by dragging the slider.

7. The method according to claim 1, characterized in that, The standard format file is in COCOJSON or GeoJSON format, so that it can be directly used for quantitative analysis, statistical reporting, or training of deep learning models.

8. An interactive scanned image intelligent annotation device based on panoramic images, used to implement the method described in any one of claims 1-7, characterized in that, The system includes: The scanning control module is used to control the scanner to perform a sequence of redundant scans with a predetermined overlap rate, acquire scanned sub-images in real time, perform image registration, overlap area verification and scan position calibration for each sub-image, crop the standard effective area and record the metadata of each sub-image to a structured index. The panoramic construction module receives metadata and effective area sub-images from the scanning control module. Based on the metadata, it uses image registration and fusion algorithms to stitch the sub-images into a high-resolution seamless panoramic digital slide image. The pre-annotation module is used to load the pre-trained segmentation model, perform batch inference on the effective region sub-images corresponding to the panoramic image output by the panoramic construction module to generate target category labels, bounding boxes and binary masks, perform semantic encoding on the categories, map the target masks to the panoramic global coordinate system and synthesize logical annotation layers, and finally generate structured annotation files. The annotation processing module is an interactive annotation processing module that provides intelligent interactive tools, receives annotation operation instructions from users, verifies, modifies and updates the structured annotation files generated by the pre-annotation module, records all annotation operation sequences, and supports multi-step undo and redo functions. The visualization rendering module is used to dynamically display panoramic images and annotation information using a dual-layer rendering architecture. The bottom layer renders panoramic digital slide images, and the upper layer parses structured annotation files and generates annotation overlays based on the visible area. The annotation overlays are rendered using an encoding method that decouples RGB and Alpha channels, and a front-end interactive control is provided to adjust the global transparency of the annotation overlays in real time. The results output module is used to optimize and verify the structured annotation data file updated by the annotation processing module, and export it as a standard format annotation result file.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images