An artificial intelligence-based pathological section image recognition system

Abstract

The application discloses a pathological section image recognition system based on artificial intelligence and belongs to the technical field of digital pathology. A parallel processing scheduling engine, a progressive scanning controller, a dynamic region of interest recognition module, a high-resolution focusing and feature enhancement module, an artificial intelligence analysis module and a result integration and report generation module are cooperatively used in the server. The system adopts a parallel architecture with deep coupling of progressive scanning and real-time analysis. In the low-resolution fast scanning process, candidate regions are dynamically identified, and a scanner is immediately dispatched for high-resolution focusing scanning and image enhancement. Then, an artificial intelligence model is used for analysis and output of structured results containing a certainty score. The method embeds the analysis process into the scanning process, greatly shortens the intraoperative diagnosis waiting time, improves the recognition accuracy of low-quality images such as frozen sections through multi-layer focusing fusion and special enhancement processing, and enhances the reliability of the diagnosis results through the built-in certainty evaluation and feedback mechanism.

Description

Technical Field

[0001] This invention relates to the fields of digital pathology and artificial intelligence, and in particular to an artificial intelligence-based pathological slide image recognition system. Background Technology

[0002] Digital pathology lays the foundation for remote and intelligent pathological diagnosis by generating high-resolution digital images through whole-slice scanning of traditional glass pathology slides. In existing technologies, AI-based pathology slide image recognition systems typically employ the following process: First, a digital pathology scanner performs a full-field-of-view, line-by-line scan of the entire glass slide, generating a complete ultra-high-resolution digital image (usually reaching billions of pixels, i.e., a WSI image); then, this complete WSI image is input into a pre-trained deep learning model (such as a convolutional neural network CNN) for analysis, and the model outputs segmentation, classification, or grading results for the lesion area.

[0003] However, when applied to specific time-sensitive scenarios such as rapid intraoperative pathological diagnosis, the aforementioned existing technologies reveal the following significant drawbacks:

[0004] First, existing technologies suffer from inherent time delays. Intraoperative rapid pathological diagnosis requires pathologists to complete the entire process—from tissue extraction and frozen section preparation to staining and slide interpretation—within a very short time (typically 15-30 minutes). Current technologies require waiting for the entire slide scan to be completed before AI analysis can begin. Even with a high-speed scanner, a typical WSI image takes several minutes to complete. This significant waiting time is unacceptable in intraoperative scenarios, severely limiting the practicality of AI-assisted diagnosis.

[0005] Secondly, existing technologies suffer from a mismatch between computational resources and diagnostic needs. To achieve comprehensiveness and accuracy in analysis, current systems typically require analysis of the entire WSI image. However, in intraoperative rapid pathology scenarios, surgeons' most pressing needs are to determine critical issues such as whether the surgical margins are clean and whether lymph node metastasis has occurred. The answers to these questions often depend only on the characteristics of specific regions within the slice (such as the surgical margins or hilum). Undifferentiated analysis of the entire slice consumes a significant amount of unnecessary computational resources and prolongs the overall processing time.

[0006] Furthermore, existing technologies suffer from decreased accuracy when faced with the unique technical limitations of intraoperative frozen sections. Due to the rapid preparation process, intraoperative frozen sections often suffer from poor cell morphology visualization, unclear tissue layering, uneven section thickness, and interference from ice crystal artifacts, resulting in image quality far lower than that of conventional paraffin sections. Most existing AI models are trained on high-quality conventional pathological slide images. When directly applied to low-quality frozen section images, the model's ability to extract key features such as cell boundaries and nuclear details decreases, leading to increased false negative or false positive rates. Particularly in areas with dense cell density and significant overlap (such as when identifying lymph node micrometastases), the model is prone to misjudgment due to feature blurring.

[0007] Finally, the decision-making process of existing technologies lacks effective handling of the model's own uncertainties. Deep learning models experience a decrease in prediction confidence when faced with blurred images or regions with atypical features. Existing systems typically simply output a classification or segmentation result without effectively feeding this uncertainty information back to the system or the surgeon. In the stressful intraoperative environment, a low-confidence positive or negative result could mislead the surgeon, and the system itself cannot proactively indicate that the result needs further verification.

[0008] Therefore, there is an urgent need for a pathological slide image recognition system optimized for time-sensitive, target-specific, and image-quality-challenging scenarios such as rapid intraoperative pathological diagnosis, in order to solve the aforementioned technical problems of time delay, computational redundancy, poor adaptability to low-quality images, and insufficient uncertainty management.

[0009] Digital pathology lays the foundation for remote and intelligent pathological diagnosis by generating high-resolution digital images through whole-slice scanning of traditional glass pathology slides. In existing technologies, AI-based pathology slide image recognition systems typically employ the following process: First, a digital pathology scanner performs a full-field-of-view, line-by-line scan of the entire glass slide, generating a complete ultra-high-resolution digital image (usually reaching billions of pixels, i.e., a WSI image); then, this complete WSI image is input into a pre-trained deep learning model (such as a convolutional neural network CNN) for analysis, and the model outputs segmentation, classification, or grading results for the lesion area.

[0010] However, when applied to specific time-sensitive scenarios such as rapid intraoperative pathological diagnosis, the aforementioned existing technologies reveal the following significant drawbacks:

[0011] First, existing technologies suffer from inherent time delays. Intraoperative rapid pathological diagnosis requires pathologists to complete the entire process—from tissue extraction and frozen section preparation to staining and slide interpretation—within a very short time (typically 15-30 minutes). Current technologies require waiting for the entire slide scan to be completed before AI analysis can begin. Even with a high-speed scanner, a typical WSI image takes several minutes to complete. This significant waiting time is unacceptable in intraoperative scenarios, severely limiting the practicality of AI-assisted diagnosis.

[0012] Secondly, existing technologies suffer from a mismatch between computational resources and diagnostic needs. To achieve comprehensiveness and accuracy in analysis, current systems typically require analysis of the entire WSI image. However, in intraoperative rapid pathology scenarios, surgeons' most pressing needs are to determine critical issues such as whether the surgical margins are clean and whether lymph node metastasis has occurred. The answers to these questions often depend only on the characteristics of specific regions within the slice (such as the surgical margins or hilum). Undifferentiated analysis of the entire slice consumes a significant amount of unnecessary computational resources and prolongs the overall processing time.

[0013] Furthermore, existing technologies suffer from decreased accuracy when faced with the unique technical limitations of intraoperative frozen sections. Due to the rapid preparation process, intraoperative frozen sections often suffer from poor cell morphology visualization, unclear tissue layering, uneven section thickness, and interference from ice crystal artifacts, resulting in image quality far lower than that of conventional paraffin sections. Most existing AI models are trained on high-quality conventional pathological slide images. When directly applied to low-quality frozen section images, the model's ability to extract key features such as cell boundaries and nuclear details decreases, leading to increased false negative or false positive rates. Particularly in areas with dense cell density and significant overlap (such as when identifying lymph node micrometastases), the model is prone to misjudgment due to feature blurring.

[0014] Finally, the decision-making process of existing technologies lacks effective handling of the model's own uncertainties. Deep learning models experience a decrease in prediction confidence when faced with blurred images or regions with atypical features. Existing systems typically simply output a classification or segmentation result without effectively feeding this uncertainty information back to the system or the surgeon. In the stressful intraoperative environment, a low-confidence positive or negative result could mislead the surgeon, and the system itself cannot proactively indicate that the result needs further verification.

[0015] Therefore, there is an urgent need for a pathological slide image recognition system optimized for time-sensitive, target-specific, and image-quality-challenging scenarios such as rapid intraoperative pathological diagnosis, in order to solve the aforementioned technical problems of time delay, computational redundancy, poor adaptability to low-quality images, and insufficient uncertainty management. Summary of the Invention

[0016] To achieve the above objectives, the present invention provides an artificial intelligence-based pathological slide image recognition system, the system comprising a digital pathology scanner and a computing server;

[0017] The digital pathology scanner is used to scan pathological slides to obtain digital images;

[0018] The computing server is communicatively connected to the digital pathology scanner. The computing server is equipped with a parallel processing scheduling engine, a progressive scanning controller, a dynamic region of interest identification module, a high-resolution focusing and feature enhancement module, an artificial intelligence analysis module, and a result integration and report generation module.

[0019] The parallel processing scheduling engine is used to coordinate the scanning process of the digital pathology scanner and the analysis process of the computing server.

[0020] The progressive scan controller is used to control the working mode of the digital pathology scanner according to the instructions of the parallel processing scheduling engine.

[0021] The dynamic region of interest identification module is used to analyze the continuously acquired image data in real time during the scanning process to identify candidate regions;

[0022] The high-resolution focusing and feature enhancement module is used to perform focused scanning and image quality enhancement on the candidate region;

[0023] The artificial intelligence analysis module is used to analyze the enhanced image and output structured results containing deterministic scores;

[0024] The results integration and report generation module is used to integrate all structured results and generate diagnostic reports.

[0025] Preferably, the specific steps for real-time analysis by the dynamic region of interest identification module include:

[0026] The image acquisition and preprocessing module receives the image block data stream scanned and output by the digital pathology scanner at a first preset resolution in real time, and stitches the image blocks in memory to form a continuously expanding low-resolution panoramic preview image.

[0027] The dynamic region of interest identification module traverses the latest expanded region of the low-resolution panoramic preview image using a sliding window of a preset size, and uses a lightweight convolutional neural network to perform preliminary analysis on the image content within each sliding window.

[0028] The lightweight convolutional neural network is pre-trained to identify regions with abnormal cell density, irregular tissue boundaries, and texture differences from the surrounding background in an image, and outputs the probability value of each sliding window belonging to a candidate region.

[0029] When the probability value output by the lightweight convolutional neural network exceeds a preset activation threshold, the dynamic region of interest identification module determines the area covered by the current sliding window as a candidate region and immediately calculates the image coordinates of the candidate region in the low-resolution panoramic preview image.

[0030] The dynamic region of interest identification module sends the image coordinates to the parallel processing scheduling engine;

[0031] The parallel processing scheduling engine converts the image coordinates into the physical coordinates corresponding to the stage of the digital pathology scanner and generates scanning control commands.

[0032] Preferably, the specific steps for the high-resolution focusing and feature enhancement module to perform focused scanning and image quality enhancement include:

[0033] The parallel processing scheduling engine sends instructions to the progressive scan controller to control the digital pathology scanner to switch the scanning mode from the first scanning mode to the second scanning mode and drive the scanning head to move to the physical coordinates corresponding to the candidate region.

[0034] In the second scanning mode, the digital pathology scanner acquires images of the candidate region at a second preset resolution, which is higher than the first preset resolution.

[0035] During image acquisition, the high-resolution focusing and feature enhancement module controls the autofocus system of the digital pathology scanner to acquire a high-resolution image of the candidate region at multiple preset focal plane positions along the vertical direction of the optical axis, thereby obtaining an image sequence composed of multiple images acquired at different focal planes.

[0036] The high-resolution focusing and feature enhancement module performs a depth-of-field fusion algorithm on the image sequence. The process of the depth-of-field fusion algorithm is as follows: First, the sharpness metric value of each pixel position in each image in the image sequence is calculated. The sharpness metric value is obtained by calculating the sum of the gradient magnitudes in the pixel's neighborhood. Then, for each target pixel position in the fused output image, the pixel value with the highest sharpness metric value at that pixel position is selected from all images in the image sequence, and this pixel value is assigned to the target pixel position in the output image.

[0037] After the depth-of-field fusion is completed, the high-resolution focusing and feature enhancement module further applies an artifact removal filter based on generative adversarial network training to process the fused image. The artifact removal filter is specifically used to suppress bright spot artifacts caused by ice crystals and color differences caused by uneven staining in intraoperative frozen section images, so as to enhance the contrast between cell nuclei and cytoplasm.

[0038] Preferably, the specific steps for the artificial intelligence analysis module to analyze and output structured results include:

[0039] The artificial intelligence analysis module receives high-quality candidate region images that have been enhanced from the output of the high-resolution focusing and feature enhancement module;

[0040] The artificial intelligence analysis module uses a deep convolutional neural network to analyze high-quality images. The deep convolutional neural network includes a feature extraction backbone network and multiple parallel output heads.

[0041] The first output head is used for lesion classification, outputting a classification probability distribution vector, where each element of the vector represents the probability that the image belongs to a preset lesion category;

[0042] The second output head is used to output a deterministic score. The process of generating the deterministic score is as follows: an uncertainty estimation subnetwork inside the deep convolutional neural network receives the intermediate feature map from the feature extraction backbone network. The uncertainty estimation subnetwork outputs a scalar value, which is normalized by the sigmoid function and mapped to a deterministic score value between 0 and 1.

[0043] The structured result is a dataset, which includes at least: the lesion type label determined by the first output head, the highest probability value corresponding to the lesion type label, the position coordinates of the candidate region in the whole film, and the deterministic score value generated by the second output head.

[0044] Preferably, the specific steps for the result integration and report generation module to integrate results and generate reports include:

[0045] The results integration and report generation module receives structured results for different candidate regions from the artificial intelligence analysis module in real time, and caches each structured result in a list associated with a unique slice identifier.

[0046] When caching each newly received structured result, the result integration and report generation module checks whether the candidate region coordinates corresponding to the result have been covered by other results that were previously cached and have a higher deterministic score. If so, the new result is marked with a lower weight.

[0047] Once the digital pathology scanner completes the first preset resolution scan of the entire slide, and the parallel processing scheduling engine confirms that all identified candidate regions have completed high-resolution analysis and result reporting, the result integration and report generation module triggers the final report generation process.

[0048] The final report generation process includes: generating a low-resolution panoramic image of the entire slice as the report base map; drawing the positions of all cached, unweighted candidate regions on the report base map with visually distinguishable bounding boxes and highlight colors; generating a detailed entry for each drawn candidate region, listing the lesion type, probability, deterministic score, and a link to a thumbnail of the enhanced image.

[0049] The results integration and report generation module summarizes the entries in all candidate regions based on the clinical concerns information pre-entered by the operator, generates a summary text description of the key diagnostic questions, and combines the report base map, detailed entries, and summary text into a complete diagnostic report document for output.

[0050] Preferably, the specific method for determining the first preset resolution includes:

[0051] The first preset resolution refers to the number of pixels acquired per unit physical length by the digital pathology scanner in the first scanning mode;

[0052] The first preset resolution is selected so that the lightweight convolutional neural network can achieve a recall rate of no less than 99% for abnormal areas that need further attention when analyzing low-resolution panoramic preview images, while ensuring that the total time required to complete the rapid traversal scan of the entire pathological slide does not exceed 2 minutes.

[0053] Before system deployment, the specific value of the first preset resolution is determined by the following steps: collect a training set containing multiple intraoperative frozen sections, and label all abnormal regions in the training set at the highest resolution; downsample the highest resolution image to a series of different lower resolutions; at each lower resolution, use a lightweight convolutional neural network for detection and calculate its recall rate; select the resolution with the lowest value among all resolutions that meet the requirement of a recall rate of not less than 99% as the first preset resolution.

[0054] Preferably, the specific method for determining the second preset resolution includes:

[0055] The second preset resolution refers to the number of pixels acquired per unit physical length by the digital pathology scanner in the second scanning mode.

[0056] The second preset resolution is selected based on the minimum pixel scale required for the deep convolutional neural network to reliably analyze the morphological features of the cell nucleus. The morphological features include at least the outline of the cell nucleus, the area ratio of the cell nucleus to the cytoplasm, and the distribution texture of the chromatin.

[0057] The definition of the minimum pixel scale is: in a digital image, the number of pixels occupied by the nucleus diameter of a typical lymphocyte is no less than 8 pixels;

[0058] Before system deployment, the specific value of the second preset resolution is determined through the following steps: a standard micrometer ruler with a known physical size is calibrated under a microscope, and scanned using a digital pathology scanner at different resolution settings; the number of pixels in the image corresponding to the known length on the standard micrometer ruler at each resolution is measured, thereby establishing the conversion relationship between physical size and the number of pixels; based on the physical diameter range of typical lymphocyte nuclei, the number of pixels per micrometer required to satisfy an imaging diameter of not less than 8 pixels is calculated, and this value is set as the second preset resolution.

[0059] Preferably, the deterministic score is used to trigger a secondary analysis process:

[0060] The results integration and report generation module has a preset deterministic scoring threshold;

[0061] When the deterministic score value in the structured result output by the artificial intelligence analysis module for a candidate region is lower than the deterministic score threshold, the result integration and report generation module marks the candidate region as "needs review" and feeds this status back to the parallel processing scheduling engine.

[0062] After receiving the "requires review" status feedback, the parallel processing scheduling engine generates a secondary analysis instruction and sends it to the progressive scan controller;

[0063] The progressive scan controller controls the digital pathology scanner to scan the candidate areas marked as "needs review" again. This scan uses a higher resolution than the second scan mode, or uses more focal planes than the first focused scan, to obtain higher quality or more dimensional image data.

[0064] The high-resolution focusing and feature enhancement module and the artificial intelligence analysis module perform the same enhancement and analysis process on the data obtained from the second scan as the first analysis, and output the second structured results;

[0065] The results integration and report generation module compares the second set of structured results with the first set of results, and selects the result with the higher certainty score as the final analysis result for the candidate region for integration and reporting.

[0066] Preferably, the parallel processing scheduling engine performs model parameter preloading during the system initialization phase:

[0067] Before the scan begins, the operator inputs the clinical focus information for this scan into the system through the interactive interface. The clinical focus information includes at least the tissue type and key diagnostic questions.

[0068] The parallel processing scheduling engine maintains a configuration database that stores the mapping relationship between different combinations of clinical concern information and a set of artificial intelligence model parameters;

[0069] The parallel processing scheduling engine queries the configuration database based on the received clinical concern information and loads the corresponding set of artificial intelligence model parameters;

[0070] The loaded artificial intelligence model parameters include at least: the weight parameters of the lightweight convolutional neural network in the dynamic region of interest identification module, the weight parameters of the deep convolutional neural network in the artificial intelligence analysis module, and the coefficient parameters of the artifact removal filter.

[0071] The parallel processing scheduling engine distributes the loaded weight parameters and coefficient parameters to the corresponding modules to complete the preparatory work before analysis.

[0072] Preferably, the system further includes an interactive interface connected to a computing server, the interface being used for information input and report presentation:

[0073] The interactive interface provides an information input area for operators to enter or select clinical concern information before the scan begins;

[0074] The interactive interface provides a real-time display area that continuously expands the low-resolution panoramic preview image during the scanning and analysis process, and marks the candidate regions that have just been discovered by the dynamic region of interest identification module on the preview image with dynamically flashing boxes.

[0075] The interactive interface provides an intermediate results area for real-time scrolling display of the latest structured results summary of candidate regions for which the artificial intelligence analysis module has completed the analysis;

[0076] The interactive interface provides a final report area. After the results integration and report generation module completes the final diagnostic report, the report base map, detailed item list and summary text are displayed in the final report area in a column format.

[0077] The interactive interface provides an interactive control that allows pathologists to click on any candidate region entry in the report to view a full-resolution enhanced image of that region after processing by the high-resolution focusing and feature enhancement module in a pop-up window.

[0078] The beneficial effects of this invention are:

[0079] 1. Through a parallel architecture of "scanning and analyzing simultaneously," the system begins preliminary analysis during the low-resolution scanning phase and can output high-confidence results for some key areas before the scan is completed. Pathologists do not need to wait for the entire scan to finish before reviewing the AI ​​report, advancing the start time of AI-assisted diagnosis from after the scan ends to shortly after the scan begins, effectively meeting the urgent requirement of rapid intraoperative diagnosis.

[0080] 2. By dynamically identifying regions of interest (ROIs), the system concentrates computationally expensive, detailed analysis on a small subset of areas most likely to have problems, avoiding the enormous computational waste associated with indiscriminate analysis of the entire image in traditional solutions. This not only improves analysis speed but also reduces the excessive demands on computing server hardware performance.

[0081] 3. By introducing multi-layer focusing scanning and depth-of-field fusion technology, the local blurring problem caused by uneven slice thickness in single-layer scanning was solved; and a dedicated artifact removal filter effectively suppressed noise such as ice crystal artifacts. These image enhancement steps, as preprocessing for AI analysis, directly improved the data quality of the input model, thereby enhancing the model's feature extraction capability and final judgment accuracy in challenging scenarios.

[0082] 4. By outputting a "certainty score," the system can self-assess the reliability of its judgments and quantify the uncertainty feedback to the doctor. For areas with low certainty scores, the system can proactively prompt the doctor to review them carefully and trigger more detailed secondary scan analysis, forming a self-optimization closed loop of "perception-feedback-enhancement," reducing the risk of misdiagnosis that may be caused by the model's overconfidence. Attached Figure Description

[0083] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0084] Figure 1 This is a structural block diagram of the system of the present invention. Detailed Implementation

[0085] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0086] Please see Figure 1This invention provides an artificial intelligence-based pathological slide image recognition system, comprising a digital pathology scanner and a computing server. The computing server and the digital pathology scanner communicate via a network. The computing server internally includes multiple functional modules, including a parallel processing scheduling engine, a progressive scan controller, a dynamic region of interest identification module, a high-resolution focusing and feature enhancement module, an artificial intelligence analysis module, and a result integration and report generation module.

[0087] The digital pathology scanner is used to carry glass pathology slides and perform optical imaging and digitization. The scanner has at least two programmable scanning modes: one for rapid, low-resolution panoramic scanning, and the other for fine, high-resolution scanning of a specified coordinate area. The parallel processing scheduling engine, as the system's central coordinating unit, manages the timing and resource allocation between the scanning and analysis processes. The progressive scan controller receives instructions from the parallel processing scheduling engine and translates these instructions into specific control signals that control the digital pathology scanner's motion mechanisms and imaging parameters.

[0088] The dynamic region of interest (ROI) identification module runs continuously during scanning, performing real-time analysis on the constantly acquired image data. The high-resolution focusing and feature enhancement module specifically processes image data marked as candidate regions. The artificial intelligence analysis module receives the enhanced image data and performs deep analysis. The results integration and report generation module summarizes all analysis results and generates a final report.

[0089] In one possible implementation, the image acquisition and preprocessing module continuously receives image patch data streams generated by scanning at a first preset resolution from the digital pathology scanner. These image patches are transmitted in real time to a designated buffer in memory. The image acquisition and preprocessing module invokes an image stitching algorithm to stitch the received image patches onto a virtual panoramic canvas based on their coordinate information. This panoramic canvas continuously expands to the right and downwards as the scanning progresses, forming a low-resolution panoramic preview image.

[0090] During the stitching process, each image patch undergoes automatic contrast adjustment and mild noise reduction preprocessing. The dynamic region of interest (ROI) identification module initiates a separate analysis thread that continuously monitors the update status of the low-resolution panoramic preview image. Once a new image region is stitched, the module immediately processes this newly expanded region. During processing, a fixed-size rectangular sliding window is used, traversing the expanded region from left to right and top to bottom according to a set step size. For each sub-region covered by the sliding window, the module invokes a pre-loaded lightweight convolutional neural network for fast inference. This lightweight convolutional neural network is specially designed with shallow layers and few parameters to achieve millisecond-level inference speeds. During the training phase, the network uses a large amount of labeled data to learn to identify preliminary visual patterns related to lesions, such as abnormal cell aggregation morphology, tissue structure disorder, or abnormal regions with specific staining intensities.

[0091] The network outputs a probability value between 0 and 1 for each sliding window, representing the likelihood that the window's content belongs to a candidate region requiring further attention. The system sets an activation threshold, for example, 0.65. When the probability output value of a sliding window exceeds this activation threshold, the dynamic region of interest (ROI) identification module determines that window region as a candidate region and immediately records the precise pixel coordinates of that window in the low-resolution panoramic preview image, including the coordinates of the top-left and bottom-right corners.

[0092] Subsequently, the module encapsulates this coordinate information into a message and sends it to the parallel processing scheduling engine. After receiving the coordinate message, the decoding unit of the parallel processing scheduling engine calculates the coordinate transformation based on the pre-stored mapping relationship between the scanner coordinate system and the image pixel coordinate system, converting the image pixel coordinates into the actual physical coordinates of the digital pathology scanner stage in three-dimensional space, and generating scanning control commands containing these physical coordinates.

[0093] In one possible implementation, after the parallel processing scheduling engine generates scan control commands, it sends them to the progressive scan controller. The progressive scan controller parses the commands and first sends a command to the digital pathology scanner to switch the current scan mode from a first scan mode to a second scan mode. The resolution in the second scan mode is higher than that in the first scan mode.

[0094] Next, the controller drives the scanner's stepper motor, controlling the precise movement of the stage to align the optical center of the scanning head with the physical coordinates corresponding to the candidate area. In the second scanning mode, the digital pathology scanner acquires images of a predetermined-size rectangular area centered at these physical coordinates at a second preset resolution. During acquisition, the high-resolution focusing and feature enhancement module intervenes. This module, through the scanner's application programming interface, invokes the extended functions of its autofocus system. The module instructs the autofocus system not to be fixed on a single focal plane, but rather to focus and acquire images along the optical axis at multiple different height positions preset by the software.

[0095] For example, the system can select two offset points upwards and two downwards from the theoretically optimal focal plane, acquiring a total of five images of the same region at five different focal planes, forming an image sequence. After completing the multi-focal-plane image acquisition, the high-resolution focusing and feature enhancement module executes a depth-of-field fusion algorithm. This algorithm first needs to evaluate the sharpness of each pixel location in each image of the image sequence. Sharpness is typically measured by calculating the image gradient intensity within a small neighborhood around the pixel; a larger gradient intensity generally indicates a sharper point. The algorithm iterates through each pixel location of the final output image, comparing the sharpness metric values ​​of all images in the image sequence at that same pixel location.

[0096] The algorithm selects the source image with the highest sharpness metric and assigns the pixel value of the corresponding pixel location in the source image to the same pixel location in the output image. By traversing all pixels, a high-quality image is finally synthesized in which all objects within the entire target area are sharp.

[0097] The module then applies a dedicated artifact removal filter trained using a generative adversarial network (GAN) to process the fused image. This GAN is a deep learning model that trains using pairs of frozen section images with ice crystal artifacts and expert-processed images with fewer artifacts. After training, its generator effectively identifies and reduces bright spot artifacts formed by ice crystals in the image, while also compensating for uneven staining caused by the rapid staining process. This significantly enhances the contrast between the cell nucleus and cytoplasm, making key morphological features more prominent.

[0098] In one possible implementation, the AI ​​analysis module receives a high-quality image of candidate regions, processed by depth-of-field fusion and artifact removal, from the output of the high-resolution focusing and feature enhancement module. At the core of this module is a deep convolutional neural network, comprising a backbone network for feature extraction and multiple parallel output heads. The feature extraction backbone network typically consists of multiple convolutional layers, pooling layers, and residual connections, responsible for extracting multi-level, abstract feature maps from the input image. The first output head is a lesion classification head, which receives the high-level features extracted by the backbone network and, through structures such as fully connected layers, outputs a probability distribution vector. The dimension of this vector equals the preset number of lesion categories, and the value of each dimension represents the probability that the input image belongs to the corresponding category; the sum of the probabilities of all dimensions is 1.

[0099] The second output head is the deterministic scoring head, which is essentially an uncertainty estimation subnetwork. This subnetwork also receives intermediate feature maps from the backbone network as input. These feature maps contain essential information about the image. The uncertainty estimation subnetwork learns to infer the model's confidence in the current prediction from these features through several additional convolutional and fully connected layers. This subnetwork outputs a raw scalar value, which is normalized by a sigmoid activation function and mapped to the range of 0 to 1, serving as the final deterministic score. The closer the score is to 1, the more confident the model is in the classification result given by the first output head; the closer it is to 0, the more the model considers the current input image features atypical, and the uncertainty of the prediction result is high.

[0100] Ultimately, the structured result output by the artificial intelligence analysis module is a well-encapsulated data object, which contains at least the following fields: the name of the lesion type with the highest probability determined by the first output head; the specific probability value corresponding to the lesion type; the position coordinates of the candidate region in the entire pathological slice panorama, usually represented in the form of bounding box coordinates; and a deterministic score value between 0 and 1 generated by the second output head.

[0101] In one possible implementation, the results integration and report generation module continuously listens for and receives structured results for different candidate regions asynchronously sent by the artificial intelligence analysis module. Each time a new structured result is received, the module adds it to a dynamic list associated with a unique identifier for the currently processed pathological slide. While caching new results, the module performs a deduplication check: calculating the overlap between the coordinates of the candidate region corresponding to the new result and the coordinates of existing results in the list. If the overlap exceeds a preset threshold, such as 50%, and the certainty score of the existing result is higher than that of the new result, the new result is marked as "redundant" or "de-weighted" and may not be displayed independently in the final report, or may only be shown as a note.

[0102] When the system detects that both completion conditions are met—that is, the digital pathology scanner has completed a panoramic scan of the entire slide at the first preset resolution, and the parallel processing scheduling engine has confirmed that all sent candidate region analysis tasks have received results—the result integration and report generation module initiates the final report generation process. This process first calls an image processing function to generate a complete, low-resolution panoramic image as the report's visual base map. Subsequently, the module iterates through the filtered, unweighted structured result list, drawing bounding boxes on the report base map based on the location coordinates of each result, and highlighting the bounding boxes with different colors, such as red for high-risk lesions and yellow for areas awaiting confirmation.

[0103] For each drawn bounding box, the module generates a corresponding detailed text entry. This entry systematically lists the analysis conclusions for that region, including lesion type, diagnostic probability, and deterministic score, and provides an interactive link to a locally stored high-resolution image file of that region after enhancement. Finally, based on the clinical concerns input by the operator during system initialization, such as "margin status assessment," the module logically summarizes all detailed entries to generate a concluding text description that directly answers clinical questions. The report base map, the list of all detailed entries, and the concluding text are combined and packaged into a complete diagnostic report document, which is displayed through the system's interactive interface.

[0104] In one possible implementation, the first preset resolution is a key parameter that defines the number of image pixels per unit physical length for the digital pathology scanner in the first scanning mode, typically measured in micrometers per pixel. The selection of this parameter follows two core principles: the primary principle is to ensure that the lightweight convolutional neural network in the dynamic region of interest (ROI) recognition module has a sufficiently high lesion detection capability, i.e., a recall rate of no less than 99%; the secondary principle is to ensure that the total time required to complete a rapid traversal scan of the entire pathology slide at this resolution is controlled within 2 minutes, to meet the timeliness requirements of rapid intraoperative diagnosis. Before the system is formally deployed, an offline calibration and verification process is needed to determine the specific value of this parameter.

[0105] The specific steps include: collecting a representative training image set containing multiple intraoperative frozen sections, and having a pathologist precisely label all relevant abnormal regions in the image set at the highest available resolution. These high-resolution images are then processed using downsampling techniques in digital image processing to generate a series of image copies with different lower resolutions. For each lower-resolution copy, a trained lightweight convolutional neural network is used to automatically detect abnormalities in the entire image set, and the detection results are compared with the expert-labeled real regions to calculate the recall rate at that resolution. Among all resolution values ​​that meet the recall requirement of at least 99%, the lowest value is selected—the "coarsest" resolution that still meets the detection rate requirement—and this is formally set as the first preset resolution. This selection method maximizes scanning speed while ensuring detection performance.

[0106] In one possible implementation, the second preset resolution defines the image precision of the digital pathology scanner during fine scanning in the second scanning mode, also measured in micrometers per pixel. This parameter is selected to meet the minimum pixel scale required for reliable and accurate analysis of microscopic morphological features at the cell nucleus level by the deep convolutional neural network in the artificial intelligence analysis module. These key morphological features include at least the smoothness of the cell nucleus outline, the integrity of the nuclear membrane, the area ratio of the cell nucleus to the surrounding cytoplasm, and the distribution texture of chromatin granules within the cell nucleus.

[0107] The specific definition of the minimum pixel scale can be quantified as follows: in the final digital image, the diameter of a typical target cell, such as the nucleus of a lymphocyte, should occupy no fewer than 8 pixels. Before system deployment, determining the specific value of this parameter requires a device calibration process. A standard microscopic ruler with known physical dimensions, such as a ruler with 2-micrometer intervals, is used as the calibration object. This ruler is placed on the scanner stage, and multiple different resolution options are set in the scanner software, followed by scanning. At each resolution setting, the number of pixels corresponding to the known physical length on the ruler in the scanned image is measured, thereby accurately calculating the "micrometer per pixel" value for each resolution setting.

[0108] Subsequently, based on medical knowledge, the physical diameter range of typical target cell nuclei in the tissue to be tested is determined, for example, 5 to 10 micrometers. Using established conversion relationships, the lower limit of the number of pixels per micrometer required to ensure that this diameter is represented by at least 8 pixels in the image is calculated. This lower limit is set as the second preset resolution. For example, if the cell nucleus diameter is 8 micrometers and an image of at least 8 pixels is required, the resolution must be no coarser than 1 micrometer / pixel, and a higher resolution, such as 0.25 micrometers / pixel, is usually chosen to ensure analytical redundancy.

[0109] In one possible implementation, the result integration and report generation module has a preset numerical threshold, such as 0.7, as the boundary for judging the certainty score. When the certainty score in the structured result output by the artificial intelligence analysis module for a candidate region is lower than the preset certainty score threshold, the result integration and report generation module will perform a specific operation. First, the module appends a "requires review" status flag to the metadata of the result.

[0110] Simultaneously, the module generates a status feedback message containing the identifier of the candidate region and its "requires review" status, and sends it back to the parallel processing scheduling engine. Upon receiving such a feedback message, the parallel processing scheduling engine initiates a higher-priority secondary analysis task. The engine generates new, more refined scan instructions that require a rescan of the original candidate region, but with more stringent scan parameters. For example, the instructions might require a third resolution scan than the second scan mode, or sampling across a greater number of focal planes to obtain richer depth information.

[0111] Upon receiving the secondary analysis command, the progressive scan controller instructs the digital pathology scanner to perform a new, more demanding scan of the target area. The high-resolution focusing and feature enhancement module and the artificial intelligence analysis module then perform the same processing and analysis procedures on the newly acquired image data as the initial analysis, outputting a second structured result containing a new, typically higher, certainty score. Finally, the results integration and report generation module compares the first and second structured results for the same candidate region. Based on a set of predefined rules, such as prioritizing the result with the higher certainty score, the selected result is used as the final analysis conclusion for that region, for subsequent integration and final report generation.

[0112] In one possible implementation, before the digital pathology scanner begins scanning, the operator needs to input clinical concerns relevant to the diagnostic task through the system's graphical user interface. This information is typically collected via drop-down menus, checkboxes, or text input boxes, and includes at least two parts: tissue type and key diagnostic questions. The parallel processing scheduling engine internally maintains a structured configuration database. This database stores a large number of mapping entries, each associating a specific combination of clinical concerns with a specific set of artificial intelligence model parameter file paths.

[0113] For example, when the tissue type is "breast" and the key diagnostic question is "margin assessment," one set of parameters corresponds to it; when the tissue type is "lymph node" and the key diagnostic question is "metastasis detection," another set of parameters corresponds to it. After receiving the clinical focus information from the interactive interface, the engine immediately uses it as the query condition to perform a search and matching in the configuration database. Upon successful matching, the engine loads the corresponding set of AI model parameter files from a storage device, such as a hard drive or solid-state drive. These parameter files typically include: all weight parameters of the lightweight convolutional neural network in the dynamic region of interest identification module; all weight parameters of the deep convolutional neural network in the AI ​​analysis module, including its feature extraction backbone and uncertainty estimation subnetwork; and all coefficient parameters of the artifact removal filter used in the high-resolution focusing and feature enhancement module.

[0114] After the loading process is completed, the parallel processing scheduling engine distributes and injects these parameters into the corresponding functional modules through the system's internal communication mechanism. This ensures that these modules are configured with analysis capabilities optimized for the current specific task before the scan begins, completing the preparatory work before analysis.

[0115] In one possible implementation, the interface is a software application window running on a computing server or standalone workstation, tightly connected to the various modules within the computing server. The interface provides a clear information input area containing form controls, allowing the operator to manually enter or select clinical interest information from a predefined list before initiating the scanning process. The interface includes a real-time display area that continuously updates during scanning and analysis. Its main content is a dynamically growing low-resolution panoramic preview. Visual elements, such as dynamically flashing rectangles, are overlaid on the preview to highlight candidate regions recently discovered and reported by the dynamic region of interest (ROI) identification module, providing the operator with intuitive process feedback.

[0116] The interface also includes an intermediate results area, typically presented as a list or table. This area displays a real-time scrolling summary of the latest structured results from the AI ​​analysis module for the candidate regions that have undergone analysis. The summary may only contain key information such as location, preliminary classification, and certainty score. Once the results integration and report generation module completes the synthesis of the final diagnostic report, the final report area of ​​the interface is activated and updated. This area typically uses a column or tab layout, systematically displaying the report's base map, a detailed list of all candidate regions, and the final summary text. Furthermore, the interface provides important interactive controls, such as clickable links or buttons.

[0117] When reviewing a report, pathologists can trigger a pop-up window by clicking on an item in the detailed item list or a highlighted box on the report's background image. This window will display a full-resolution enhanced image of the specific candidate region after processing by the high-resolution focusing and feature enhancement module, allowing pathologists to verify and interpret the details.

[0118] Example;

[0119] This embodiment demonstrates the specific application of the system of the present invention in a rapid pathological diagnosis scenario during breast-conserving surgery for breast tumors. In this scenario, the surgeon sends the removed breast tissue specimen to the pathology department, where pathologists immediately prepare frozen sections of the surgical margins and stain them. The system of the present invention is used to analyze these frozen sections in a very short time to determine whether there are residual cancer cells at the surgical margins, thus guiding the surgeon to decide whether an extended resection is necessary.

[0120] 1. System configuration and initialization;

[0121] The system hardware in this embodiment includes a digital pathology scanner (such as a model equipped with a 40x objective lens and a motorized stage) that supports high-speed scanning and high-precision stage positioning, and a computing server configured with a high-performance GPU (such as an NVIDIA RTX A6000). The scanner and the server are connected via a high-speed local area network.

[0122] In terms of software, the various modules described in this invention are deployed on a computing server. The pathology technician places the prepared frozen sections of breast tissue (hematoxylin-eosin stained) onto the scanner stage. Subsequently, the technician enters the clinical focus information for this scan on the interactive interface connected to the computing server: selecting "breast tissue" in the "tissue type" drop-down menu and checking "margin status assessment" in "key diagnostic questions".

[0123] The interactive interface sends the input clinical concern information to the parallel processing scheduling engine. The configuration database maintained within the parallel processing scheduling engine pre-stores configurations for various scenarios. Based on the combination of "breast tissue" and "surgical margin status assessment," the engine loads a corresponding set of optimized AI model parameters. These parameters include:

[0124] Weights for a lightweight convolutional neural network (named "Fast ScreenNet") used in the dynamic region of interest (ROI) identification module. This network employs a streamlined MobileNetV2 architecture and is specifically trained for the rapid identification of cellularly dense regions and atypical cell clusters in breast tissue on low-resolution images.

[0125] A weight file for a deep convolutional neural network (named "Deep Dx Net") used in the artificial intelligence analysis module. This network uses ResNet-50 as the feature extraction backbone and adds an output head optimized for the classification task of benign and malignant breast epithelial cells.

[0126] This is a file of artifact removal filter coefficients specifically optimized for ice crystal artifacts in frozen breast sections, used in the high-resolution focusing and feature enhancement module.

[0127] The parallel processing scheduling engine injects these parameters into the corresponding modules. At this point, the system initialization is complete and ready.

[0128] 2. Progressive scanning and real-time analysis workflow;

[0129] The technician clicks the "Start Scan" button on the interactive interface. The parallel processing scheduling engine instructs the progressive scan controller to start the digital pathology scanner.

[0130] 2.1 First round of low-resolution rapid scanning and real-time candidate region discovery;

[0131] The progressive scan controller commands the digital pathology scanner to adopt the first scanning mode. In this embodiment, after prior calibration, the first preset resolution is set to 0.4 μm / pixel. This resolution is determined as follows: 500 frozen breast slides with known diagnostic results were collected as a training set, and all cancer cell regions were labeled within them. These 500 slides were then downsampled sequentially at the highest resolution (0.1 μm / pixel) to 0.2 μm / pixel, 0.3 μm / pixel, 0.4 μm / pixel, and 0.5 μm / pixel. Fast Screen Net was used to perform detection experiments at each resolution. The results show that at a resolution of 0.4 μm / pixel, Fast Screen Net achieved a recall rate of 99.2% for cancer cell regions, while the panoramic scan time for a single slide was approximately 1 minute and 50 seconds, meeting the requirement of "no more than 2 minutes". Therefore, 0.4 μm / pixel was selected as the first preset resolution.

[0132] The scanner scans line by line from the top left corner of the slice at a resolution of 0.4 μm / pixel. The image acquisition and preprocessing module receives image blocks of 1024x1024 pixels in real time and stitches them together in memory in a "tile" manner to form a low-resolution panoramic preview image that expands continuously towards the bottom right corner. Simultaneously with the stitching, the module performs automatic white balance and a light Gaussian filter (σ=0.5) preprocessing on each image block to improve the overall visual effect.

[0133] The dynamic region of interest (ROI) identification module works synchronously. In a separate computation thread, it analyzes the latest three rows of image patches (i.e., the most recently scanned areas) of the low-resolution panoramic preview. It uses a 256x256 pixel sliding window with a step size of 128 pixels to traverse this new region. For each window, FastScreenNet analyzes it and outputs a value between 0 and 1, representing the probability that the image within that window contains a "suspicious region." In this embodiment, the activation threshold is set to 0.6. This threshold was determined by adjusting it on another validation set to balance the initial detection sensitivity with unnecessary subsequent focusing scans. When the probability value of a window is greater than 0.6, for example, a window located at a tissue cutting edge with a probability value of 0.85, the module immediately records the coordinates of this window (e.g., the starting pixel coordinates in the panoramic image). ), and then send this coordinate to the parallel processing scheduling engine.

[0134] 2.2 Dynamic scheduling and high-resolution focusing scanning;

[0135] The parallel processing scheduling engine receives coordinates Then, the known scanner parameters (such as objective magnification and camera pixel size) are first converted into the physical coordinates of the stage (e.g., X=12.5mm, Y=7.8mm). The engine checks that the current scanning head is scanning the 5th line, and its corresponding physical coordinates do not overlap with (12.5mm, 7.8mm). Therefore, the engine immediately sends a priority command to the progressive scan controller.

[0136] Upon receiving the instruction, the progressive scan controller pauses the preset low-resolution panoramic scan plan after the scanner completes the scan of the current 5th row. It then rapidly moves the scanning head to the physical coordinates (12.5mm, 7.8mm) and switches the scanning mode to the second scan mode. In this embodiment, the second preset resolution is set to 0.1μm / pixel. This resolution is determined by measuring under a microscope; the average physical diameter of the nucleus of a typical breast ductal carcinoma cell is approximately 10μm. At a resolution of 0.1μm / pixel, the diameter of this nucleus in the image is approximately 100 pixels, far exceeding the minimum requirement of "no less than 8 pixels," which is sufficient for DeepDxNet to clearly distinguish key morphological features such as irregular nuclear membranes and coarse chromatin granules.

[0137] 2.3 Multi-slice focusing scanning and image enhancement;

[0138] The digital pathology scanner performs a fine scan of a 512x512 pixel area centered at (12.5 mm, 7.8 mm) (corresponding to physical dimensions of 51.2 μm x 51.2 μm) at a resolution of 0.1 μm / pixel. A high-resolution focusing and feature enhancement module intervenes to control the autofocus system. Due to the uneven thickness of the frozen sections, the module's instruction system shifts upwards and downwards by 1 μm and 2 μm respectively from the optimal focal plane, acquiring a high-resolution image of the area at five different focal planes (Z = -2 μm, -1 μm, 0 μm, +1 μm, +2 μm), resulting in a sequence of five images.

[0139] The module then executes a depth-of-field fusion algorithm. The algorithm first calculates the sharpness metric for each pixel location in each image. This embodiment uses a gradient-based sharpness evaluation function:

[0140]

[0141] in, This represents the grayscale value of a pixel. The formula calculates the pixel value. The sum of gradient magnitudes of all pixels within a 3x3 neighborhood; a larger value indicates a sharper pixel. This is applied to each target pixel location in the final fused image. The algorithm compares the sharpness metrics of five source images at this location. ,in Choose to make The biggest one The corresponding pixel values ​​in the source image As a fused image in The pixel value at that location. By iterating through all pixels, a fully sharp, blended image is generated.

[0142] Next, an artifact removal filter processes the fused image. This filter is the generator part of a U-Net-structured generative adversarial network (GAN), which learns a mapping from "frozen section image with ice crystal artifacts" to "approximate paraffin section image without ice crystal artifacts" during training. During processing, the filter effectively identifies and reduces ice crystal artifacts that appear as small, irregular bright spots in the image. At the same time, it balances color differences caused by uneven staining through local histogram matching, making the boundary contrast between cell nuclei (stained blue-purple by hematoxylin) and cytoplasm / interstitium (stained red-pink by eosin) more distinct.

[0143] 2.4 Artificial Intelligence Analysis of Deterministic Perception;

[0144] The high-resolution focusing and feature enhancement module sends the enhanced 512x512 pixel region image to the artificial intelligence analysis module. DeepDxNet analyzes this image. Its network structure is modified at the end of the standard ResNet-50: after the global average pooling layer, the features are fed into two parallel fully connected layer branches.

[0145] First output head (classification head): Outputs a 3D vector, for example , representing the probability that the image region is "benign / normal breast tissue", "atypical hyperplasia", or "invasive carcinoma", respectively. In this case, the probability of "invasive carcinoma" is the highest (0.85).

[0146] The second output head (deterministic scoring head): This is an uncertainty estimation subnetwork. It receives intermediate feature maps (16x16x2048 pixels) from the last convolutional layer of the ResNet-50. This subnetwork consists of two convolutional layers and a global average pooling layer, ultimately outputting a scalar value. This scalar value, after being normalized by the sigmoid function, becomes the deterministic score. This score reflects the model's confidence in the current prediction. The higher the score, the more typical the features of the input image are considered by the model, and the more reliable the prediction result. Its training process uses a combined loss function:

[0147]

[0148] in, It is the standard cross-entropy classification loss. It's a real label. It is a predicted probability. This is a regularization term that encourages uncertainty; it's designed to penalize models that are overly confident in making incorrect predictions (i.e., misclassifying with a high probability). Hyperparameters Controlling the regularization strength. Through this training, the model learns to output a lower deterministic score when features are ambiguous or atypical.

[0149] In this example, the structured results output by the artificial intelligence analysis module are as follows:

[0150] Lesion type label: Invasive carcinoma;

[0151] Probability value: 0.85;

[0152] Location coordinates: (Coordinates of the low-resolution panoramic image);

[0153] Certainty score: 0.91;

[0154] 2.5 Results integration and iterative advancement;

[0155] The results integration and report generation module receives and caches the above results. Simultaneously, upon receiving a signal indicating the completion of analysis for that region, the parallel processing scheduling engine commands the scanning head to return to the interrupted position and resume rapid panoramic scanning at 0.4 μm / pixel. The dynamic region of interest identification module continues to analyze subsequently scanned regions. Throughout the scanning process of the entire slice, this closed loop of "rapid scanning - candidate region discovery - scheduling - focused scanning - enhancement - analysis" iterates continuously. For example, the system subsequently discovered two more suspicious regions, with coordinates [missing information]. and The same high-resolution analysis procedure was performed on it, and the results were "benign, probability 0.95, certainty score 0.98" and "atypical hyperplasia, probability 0.70, certainty score 0.65" respectively.

[0156] 3. Report generation and output;

[0157] Once the low-resolution scan of the entire slice is complete and the analysis of all three candidate regions is finished, the results integration and report generation module begins generating the final report. The module first checks the cached results list and identifies the coordinates. The certainty score for the result was 0.65. The preset certainty score threshold in this embodiment is 0.7. Therefore, the result was automatically marked as "requires review" by the system.

[0158] Module generates report:

[0159] 1. Report base image: A low-resolution (0.4μm / pixel) thumbnail of the entire slice.

[0160] 2. Area Marking: Mark the area on the base map with a green box. The area (invasive carcinoma) is marked with a blue box. Areas (benign) are marked with a yellow flashing box. Region (atypical hyperplasia, requires verification).

[0161] 3. Detailed entries:

[0162] Area 1 (green box): Lesion type - invasive carcinoma; probability -0.85; certainty score -0.91; conclusion - positive surgical margin (cancer cells found).

[0163] Area 2 (blue box): Lesion type - benign tissue; probability -0.95; certainty score -0.98; conclusion - negative surgical margin.

[0164] Area 3 (yellow box): Lesion type - atypical hyperplasia; probability -0.70; certainty score -0.65; conclusion - manual review required.

[0165] 4. Summary Text: Based on the initial input question "Evaluation of resection margin status", the system automatically generates a summary: "A highly definitive (0.91) invasive carcinoma lesion was found at the resection margin. Another low-definition area is recommended for review. Overall, a positive resection margin is highly probable."

[0166] The report appeared in the "Final Report Area" of the interactive interface approximately 4 minutes after the scan began. Simultaneously, because region 3 was marked "Requires Review," the parallel processing scheduling engine automatically triggered a secondary analysis process. It commanded the scanner to rescan and enhance the region at a higher resolution of 0.05 μm / pixel, across seven focal planes. After the secondary analysis, DeepDxNet output a new result: "Invasive carcinoma, probability 0.82, certainty score 0.88." The system automatically updated the report with this higher certainty result, changed the box color of region 3 to red, and updated the summary text to "Two invasive carcinoma lesions were found at the resection margin. Clearly indicating a positive resection margin."

[0167] Pathologists can click on any region box on the interactive interface, and a pop-up window will display a full-resolution image of that region after enhancement, allowing for intuitive verification.

[0168] To demonstrate the effectiveness of the system in this embodiment, the following comparative experiment was designed:

[0169] Comparative Example 1 (Traditional Method): Using the same digital pathology scanner, a complete scan of the same breast resection margin frozen section was performed at a resolution of 0.1 μm / pixel, taking 8 minutes. After scanning, the entire WSI image was input into a standard DeepDxNet network (without deterministic scoring output) that was not optimized for frozen sections for analysis, taking 3 minutes. A cancer heatmap of the entire section was obtained after a total of 11 minutes.

[0170] Comparative Example 2 (Existing Improved Method): A low-to-high scanning strategy was adopted, but in serial mode. First, the entire image was scanned at 0.4 μm / pixel (2 minutes). Then, a simple image processing algorithm (such as Otsu thresholding) was used to identify densely celled areas on the panoramic image. These areas were then scanned at high resolution (2 minutes) and analyzed (2 minutes), for a total of 6 minutes.

[0171] The method of this invention embodiment is as described above.

[0172] A test set containing 100 frozen sections of breast resection margins (all confirmed by three senior pathologists) was used for testing. The results are shown in the table below:

[0173] Evaluation indicators Comparative Example 1 (Traditional Method) Comparative Example 2 (Existing Partial Improvements) Embodiments of the present invention Average total time (minutes) 11.0 6.0 4.2 Cancer lesion detection rate (recall rate) 95% 92% 99% Accuracy of edge condition judgment 88% 90% 96% Low-certainty area alert rate 0% (This feature is not available) 0% (This feature is not available) 8% Robustness to low-quality sections (with many ice crystals) Poor (accuracy drops to 75%) Average (accuracy drops to 82%) Excellent (accuracy rate maintained at 93%) Computing resource consumption (peak GPU memory) High (12GB) Medium(6GB) Low (3GB)

[0174] Results analysis:

[0175] 1. Timeliness: By using parallel processing, this invention significantly reduces the total time from 11 minutes in Comparative Example 1 and 6 minutes in Comparative Example 2 to 4.2 minutes, demonstrating a clear advantage and meeting the needs for rapid intraoperative diagnosis.

[0176] 2. Accuracy: This invention boasts the highest detection rate and final judgment accuracy. This is attributed to its targeted image enhancement (ice crystal removal, depth-of-field fusion) and model optimized for frozen sections.

[0177] 3. Reliability: The invention’s unique “deterministic scoring” and “secondary analysis” mechanisms can identify and handle 8% of uncertainties and improve the reliability of the final conclusion through feedback loops. This is a function that the comparative model does not have at all.

[0178] 4. Robustness and Efficiency: When faced with low-quality slices, the accuracy of this invention decreases minimally, demonstrating the strongest robustness. Simultaneously, its dynamic focusing strategy minimizes computational resource consumption and achieves higher efficiency.

[0179] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0180] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A pathological slide image recognition system based on artificial intelligence, characterized in that, The system includes a digital pathology scanner and a computing server; The digital pathology scanner is used to scan pathological slides to obtain digital images; The computing server is communicatively connected to the digital pathology scanner. The computing server is equipped with a parallel processing scheduling engine, a progressive scanning controller, a dynamic region of interest identification module, a high-resolution focusing and feature enhancement module, an artificial intelligence analysis module, and a result integration and report generation module. The parallel processing scheduling engine is used to coordinate the scanning process of the digital pathology scanner and the analysis process of the computing server. The progressive scan controller is used to control the working mode of the digital pathology scanner according to the instructions of the parallel processing scheduling engine. The dynamic region of interest identification module is used to analyze the continuously acquired image data in real time during the scanning process to identify candidate regions; The high-resolution focusing and feature enhancement module is used to perform focused scanning and image quality enhancement on the candidate region; The artificial intelligence analysis module is used to analyze the enhanced image and output structured results containing deterministic scores; The results integration and report generation module is used to integrate all structured results and generate diagnostic reports.

2. The pathological slide image recognition system based on artificial intelligence according to claim 1, characterized in that, The specific steps for real-time analysis by the dynamic region of interest identification module include: The image acquisition and preprocessing module receives the image block data stream scanned and output by the digital pathology scanner at a first preset resolution in real time, and stitches the image blocks in memory to form a continuously expanding low-resolution panoramic preview image. The dynamic region of interest identification module traverses the latest expanded region of the low-resolution panoramic preview image using a sliding window of a preset size, and uses a lightweight convolutional neural network to perform preliminary analysis on the image content within each sliding window. The lightweight convolutional neural network is pre-trained to identify regions with abnormal cell density, irregular tissue boundaries, and texture differences from the surrounding background in an image, and outputs the probability value of each sliding window belonging to a candidate region. When the probability value output by the lightweight convolutional neural network exceeds a preset activation threshold, the dynamic region of interest identification module determines the area covered by the current sliding window as a candidate region and immediately calculates the image coordinates of the candidate region in the low-resolution panoramic preview image. The dynamic region of interest identification module sends the image coordinates to the parallel processing scheduling engine; The parallel processing scheduling engine converts the image coordinates into the physical coordinates corresponding to the stage of the digital pathology scanner and generates scanning control commands.

3. The pathological slide image recognition system based on artificial intelligence according to claim 1, characterized in that, The specific steps by which the high-resolution focusing and feature enhancement module performs focused scanning and image quality enhancement include: The parallel processing scheduling engine sends instructions to the progressive scan controller to control the digital pathology scanner to switch the scanning mode from the first scanning mode to the second scanning mode and drive the scanning head to move to the physical coordinates corresponding to the candidate region. In the second scanning mode, the digital pathology scanner acquires images of the candidate region at a second preset resolution, which is higher than the first preset resolution. During image acquisition, the high-resolution focusing and feature enhancement module controls the autofocus system of the digital pathology scanner to acquire a high-resolution image of the candidate region at multiple preset focal plane positions along the vertical direction of the optical axis, thereby obtaining an image sequence composed of multiple images acquired at different focal planes. The high-resolution focusing and feature enhancement module performs a depth-of-field fusion algorithm on the image sequence. The process of the depth-of-field fusion algorithm is as follows: First, the sharpness metric value of each pixel position in each image in the image sequence is calculated. The sharpness metric value is obtained by calculating the sum of the gradient magnitudes in the pixel's neighborhood. Then, for each target pixel position in the fused output image, the pixel value with the highest sharpness metric value at that pixel position is selected from all images in the image sequence, and this pixel value is assigned to the target pixel position in the output image. After the depth-of-field fusion is completed, the high-resolution focusing and feature enhancement module further applies an artifact removal filter based on generative adversarial network training to process the fused image. The artifact removal filter is specifically used to suppress bright spot artifacts caused by ice crystals and color differences caused by uneven staining in intraoperative frozen section images, so as to enhance the contrast between cell nuclei and cytoplasm.

4. The pathological slide image recognition system based on artificial intelligence according to claim 1, characterized in that, The specific steps by which the artificial intelligence analysis module analyzes and outputs structured results include: The artificial intelligence analysis module receives high-quality candidate region images that have been enhanced from the output of the high-resolution focusing and feature enhancement module; The artificial intelligence analysis module uses a deep convolutional neural network to analyze high-quality images. The deep convolutional neural network includes a feature extraction backbone network and multiple parallel output heads. The first output head is used for lesion classification, outputting a classification probability distribution vector, where each element of the vector represents the probability that the image belongs to a preset lesion category; The second output head is used to output a deterministic score. The process of generating the deterministic score is as follows: an uncertainty estimation subnetwork inside the deep convolutional neural network receives the intermediate feature map from the feature extraction backbone network. The uncertainty estimation subnetwork outputs a scalar value, which is normalized by the sigmoid function and mapped to a deterministic score value between 0 and 1. The structured result is a dataset, which includes at least: the lesion type label determined by the first output head, the highest probability value corresponding to the lesion type label, the position coordinates of the candidate region in the whole film, and the deterministic score value generated by the second output head.

5. The pathological slide image recognition system based on artificial intelligence according to claim 1, characterized in that, The specific steps for the result integration and report generation module to integrate results and generate reports include: The results integration and report generation module receives structured results for different candidate regions from the artificial intelligence analysis module in real time, and caches each structured result in a list associated with a unique slice identifier. When caching each newly received structured result, the result integration and report generation module checks whether the candidate region coordinates corresponding to the result have been covered by other results that were previously cached and have a higher deterministic score. If so, the new result is marked with a lower weight. Once the digital pathology scanner completes the first preset resolution scan of the entire slide, and the parallel processing scheduling engine confirms that all identified candidate regions have completed high-resolution analysis and result reporting, the result integration and report generation module triggers the final report generation process. The final report generation process includes: generating a low-resolution panoramic image of the entire slice as the report base map; drawing the positions of all cached, unweighted candidate regions on the report base map with visually distinguishable bounding boxes and highlight colors; generating a detailed entry for each drawn candidate region, listing the lesion type, probability, deterministic score, and a link to a thumbnail of the enhanced image. The results integration and report generation module summarizes the entries in all candidate regions based on the clinical concerns information pre-entered by the operator, generates a summary text description of the key diagnostic questions, and combines the report base map, detailed entries, and summary text into a complete diagnostic report document for output.

6. The pathological slide image recognition system based on artificial intelligence according to claim 2, characterized in that, The specific method for determining the first preset resolution includes: The first preset resolution refers to the number of pixels acquired per unit physical length by the digital pathology scanner in the first scanning mode; The first preset resolution is selected so that the lightweight convolutional neural network can achieve a recall rate of no less than 99% for abnormal areas that need further attention when analyzing low-resolution panoramic preview images, while ensuring that the total time required to complete the rapid traversal scan of the entire pathological slide does not exceed 2 minutes. Before system deployment, the specific value of the first preset resolution is determined by the following steps: collect a training set containing multiple intraoperative frozen sections, and label all abnormal regions in the training set at the highest resolution; downsample the highest resolution image to a series of different lower resolutions; at each lower resolution, use a lightweight convolutional neural network for detection and calculate its recall rate; select the resolution with the lowest value among all resolutions that meet the requirement of a recall rate of not less than 99% as the first preset resolution.

7. The pathological slide image recognition system based on artificial intelligence according to claim 3, characterized in that, The specific method for determining the second preset resolution includes: The second preset resolution refers to the number of pixels acquired per unit physical length by the digital pathology scanner in the second scanning mode. The second preset resolution is selected based on the minimum pixel scale required for the deep convolutional neural network to reliably analyze the morphological features of the cell nucleus. The morphological features include at least the outline of the cell nucleus, the area ratio of the cell nucleus to the cytoplasm, and the distribution texture of the chromatin. The definition of the minimum pixel scale is: in a digital image, the number of pixels occupied by the nucleus diameter of a typical lymphocyte is no less than 8 pixels; Before system deployment, the specific value of the second preset resolution is determined through the following steps: a standard micrometer ruler with a known physical size is calibrated under a microscope, and scanned using a digital pathology scanner at different resolution settings; the number of pixels in the image corresponding to the known length on the standard micrometer ruler at each resolution is measured, thereby establishing the conversion relationship between physical size and the number of pixels; based on the physical diameter range of typical lymphocyte nuclei, the number of pixels per micrometer required to satisfy an imaging diameter of not less than 8 pixels is calculated, and this value is set as the second preset resolution.

8. The pathological slide image recognition system based on artificial intelligence according to claim 4, characterized in that, The deterministic score is used to trigger the secondary analysis process: The results integration and report generation module has a preset deterministic scoring threshold; When the deterministic score value in the structured result output by the artificial intelligence analysis module for a candidate region is lower than the deterministic score threshold, the result integration and report generation module marks the candidate region as "needs review" and feeds this status back to the parallel processing scheduling engine. After receiving the "requires review" status feedback, the parallel processing scheduling engine generates a secondary analysis instruction and sends it to the progressive scan controller. The progressive scan controller controls the digital pathology scanner to scan the candidate areas marked as "needs review" again. This scan uses a higher resolution than the second scan mode, or uses more focal planes than the first focused scan, to obtain higher quality or more dimensional image data. The high-resolution focusing and feature enhancement module and the artificial intelligence analysis module perform the same enhancement and analysis process on the data obtained from the second scan as the first analysis, and output the second structured results; The results integration and report generation module compares the second set of structured results with the first set of results, and selects the result with the higher certainty score as the final analysis result for the candidate region for integration and reporting.

9. The pathological slide image recognition system based on artificial intelligence according to claim 1, characterized in that, The parallel processing scheduling engine performs model parameter preloading during the system initialization phase: Before the scan begins, the operator inputs the clinical focus information for this scan into the system through the interactive interface. The clinical focus information includes at least the tissue type and key diagnostic questions. The parallel processing scheduling engine maintains a configuration database that stores the mapping relationship between different combinations of clinical concern information and a set of artificial intelligence model parameters; The parallel processing scheduling engine queries the configuration database based on the received clinical concern information and loads the corresponding set of artificial intelligence model parameters; The loaded artificial intelligence model parameters include at least: the weight parameters of the lightweight convolutional neural network in the dynamic region of interest identification module, the weight parameters of the deep convolutional neural network in the artificial intelligence analysis module, and the coefficient parameters of the artifact removal filter. The parallel processing scheduling engine distributes the loaded weight parameters and coefficient parameters to the corresponding modules to complete the preparatory work before analysis.

10. The pathological slide image recognition system based on artificial intelligence according to claim 1, characterized in that, The system also includes an interactive interface connected to a computing server, which is used for information input and report presentation. The interactive interface provides an information input area for operators to enter or select clinical concern information before the scan begins; The interactive interface provides a real-time display area that continuously expands the low-resolution panoramic preview image during the scanning and analysis process, and marks the candidate regions that have just been discovered by the dynamic region of interest identification module on the preview image with dynamically flashing boxes. The interactive interface provides an intermediate results area for real-time scrolling display of the latest structured results summary of candidate regions for which the artificial intelligence analysis module has completed the analysis; The interactive interface provides a final report area. After the results integration and report generation module completes the final diagnostic report, the report base map, detailed item list and summary text are displayed in the final report area in a column format. The interactive interface provides an interactive control that allows pathologists to click on any candidate region entry in the report to view a full-resolution enhanced image of that region after processing by the high-resolution focusing and feature enhancement module in a pop-up window.

Images