A pathological whole section diagnosis method and system based on multi-magnification deep learning

Abstract

The application relates to a pathological whole-section diagnosis method and system based on multi-magnification deep learning. The method comprises the following steps: acquiring a pathological whole-section digital image file, and performing background filtering on the pathological whole-section digital image file to obtain effective tissue region coordinates; inputting the effective tissue region coordinates into a low-magnification screening model to obtain a lesion probability heat map; based on a low-magnification lesion probability value corresponding to each basic coordinate point in the probability heat map, performing local peak value sorting to obtain a dynamic scanning path queue; based on the dynamic scanning path queue, performing lesion dynamic scanning to obtain a scanned point set; integrating the scanned point set to obtain a probability comprehensive heat map; based on the probability comprehensive heat map, calculating a key diagnosis index, and based on the key diagnosis index, generating a diagnosis report to obtain a diagnosis report document. The method can dynamically plan a scanning path and improve the reliability of diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of image analysis, and in particular relates to a pathological whole-section diagnosis method and system based on multi-magnification deep learning. Background Technology

[0002] With the widespread adoption of digital pathology and high-resolution whole-slice scanning technology, artificial intelligence-assisted diagnosis has become a reality. AI technologies, represented by deep learning, can automatically learn complex features from massive amounts of images, providing a powerful tool for pathological analysis and giving rise to computer vision-based automated scanning and diagnostic methods.

[0003] In traditional automated analysis workflows, systems typically rely on a pre-defined, linear analysis strategy. For example, a low-magnification model is first used to perform a rapid but coarse screening of the entire slice, generating a static heatmap of suspected areas. Then, according to a fixed priority list in this heatmap, a high-magnification model is guided to verify predetermined coordinate points one by one. This approach simplifies the dynamic medical image analysis process into a rigid set of sequentially executed instructions.

[0004] However, this current static sequential processing method has inherent limitations. It cannot adjust decisions based on real-time high-magnification analysis results during the scan, and its scan path is open-loop and pre-set. When dealing with complex lesions such as invasive carcinoma with indistinct boundaries and diffuse growth, the fixed path is prone to missing trace or scattered lesion cell clusters that actually exist outside the peak area of ​​the thermal map, leading to a significantly increased risk of false negatives and directly affecting the comprehensiveness of the diagnosis and the accuracy of subsequent treatment decisions. While intensive scanning of the entire high-resolution area can avoid this risk, it incurs an unacceptable time cost, making it difficult to meet the timeliness requirements of real-time clinical diagnosis. Summary of the Invention

[0005] Therefore, it is necessary to provide a pathological whole-slice diagnostic method and system based on multi-magnification deep learning that can dynamically plan the scanning path and improve the reliability of lesion diagnosis, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a pathological whole-section diagnostic method based on multi-magnification deep learning, including:

[0007] Obtain digital image files of whole pathological slides, and perform background filtering on the digital image files of whole pathological slides to obtain the coordinates of the effective tissue regions;

[0008] The effective tissue region coordinates are input into the low-magnification screening model for rapid analysis to obtain a lesion probability heatmap. Based on the low-magnification lesion probability value corresponding to each basic coordinate point in the probability heatmap, local peaks are sorted to obtain a dynamic scanning path queue.

[0009] Based on a dynamic scan path queue, dynamic scanning of lesions is performed to obtain a set of scanned points;

[0010] The scan results of the scanned point set are integrated to obtain a probability-integrated heatmap.

[0011] Based on the probabilistic comprehensive heatmap, key diagnostic indicators are calculated, and based on these indicators, a diagnostic report is generated, resulting in a diagnostic report document.

[0012] Furthermore, based on the dynamic scan path queue, dynamic scanning of the lesion is performed to obtain a set of scanned points, including:

[0013] Extract at least one coordinate point with the highest probability of low-magnification lesions from the dynamic scan path queue to obtain high-magnification scan coordinate points;

[0014] Based on high-magnification scan coordinate points, image blocks corresponding to high-magnification scan coordinate points are extracted from the digital image files of whole pathological sections, and the image blocks are input into the high-magnification lesion analysis model to perform lesion probability diagnosis and obtain high-magnification lesion probability values.

[0015] By comparing the high-magnification lesion probability value with the lesion threshold, the lesion conclusion corresponding to the high-magnification scan coordinate point is obtained. The high-magnification scan coordinate point, the high-magnification lesion probability value and the lesion conclusion are integrated to obtain the set of scanned points.

[0016] Furthermore, the lesion conclusions include positive, negative, and ambiguous;

[0017] After comparing the high-magnification lesion probability value with the lesion threshold to obtain the lesion conclusion at the corresponding high-magnification scan coordinate point, the following steps are also included:

[0018] If the lesion conclusion corresponding to the high-magnification scan coordinate point is positive, then coordinate points to be explored are generated within the target radius with the high-magnification scan coordinate point as the center;

[0019] Remove the high-magnification scan coordinates from the dynamic scan path queue to obtain the updated dynamic path queue, and add the coordinates to be explored to the updated dynamic path queue to obtain the path to be explored queue.

[0020] Based on the step size threshold, the queue of paths to be explored is deduplicated to obtain a deduplicated path queue. Based on the priority score, the coordinates in the deduplicated path queue are sorted to obtain a priority path queue.

[0021] Based on pruning rules, the priority path queue is pruned to obtain the pruned path queue; the pruned path queue is used to update the dynamic scan path queue.

[0022] Furthermore, the formula for calculating the priority score is as follows:

[0023]

[0024] in, Let i be the priority score for coordinate point i. Let i be the low-magnification lesion probability value at coordinate point i. Let be the high-probability lesion probability value of the parent node of node i. For spatially consistent components, , , These are the weighting coefficients;

[0025] The formula for calculating the spatial consistency component is as follows:

[0026]

[0027] in, Let i be a circular neighborhood with radius R centered at coordinate point i. This represents the number of points within the circular neighborhood that include the already scanned set of points. denoted as the high-magnification lesion probability value of scan coordinate point j within the circular neighborhood.

[0028] Furthermore, the low-magnification screening model was obtained through the following method:

[0029] Obtain the historical annotation dataset, and based on the historical annotation dataset, perform downsampling to obtain low-magnification images, and based on the low-magnification images, perform image segmentation to obtain low-magnification image patches;

[0030] Based on low-magnification image patches, the proportion of pixel area in the covered tissue region that belongs to the lesion annotation region is calculated to obtain soft labels. The low-magnification image patches and soft labels are then integrated to obtain a training sample set.

[0031] Based on the training sample set, low-magnification image patches with soft labels greater than the sample threshold are marked as suspected positive, and the remaining low-magnification image patches are marked as confirmed negative, thus obtaining positive and negative samples;

[0032] The positive and negative samples are divided to obtain the model training dataset. Based on the model training dataset, a lightweight convolutional neural network is trained to obtain a low-ratio screening model.

[0033] Secondly, this application also provides a pathological whole-slice diagnostic system based on multi-magnification deep learning, including:

[0034] The filtering module is used to acquire digital image files of whole pathological slides and perform background filtering on the digital image files of whole pathological slides to obtain the coordinates of the effective tissue regions;

[0035] The path module is used to input the coordinates of the effective tissue area into the low-magnification screening model for rapid analysis, obtain a lesion probability heatmap, and sort the local peaks based on the low-magnification lesion probability value corresponding to each basic coordinate point in the probability heatmap to obtain a dynamic scanning path queue.

[0036] The scanning module is used to perform dynamic scanning of lesions based on a dynamic scanning path queue to obtain a set of scanned points;

[0037] The integration module is used to integrate the scan results of the scanned point set to obtain a probabilistic comprehensive heatmap.

[0038] The diagnostic module is used to calculate key diagnostic indicators based on a probabilistic comprehensive heatmap, and generate a diagnostic report document based on the key diagnostic indicators.

[0039] Thirdly, this application also provides a computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any step of the method provided in the first aspect of this application.

[0040] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any step of the method provided in the first aspect of this application.

[0041] The aforementioned pathological whole-slice diagnostic method and system based on multi-magnification deep learning acquires digital image files of pathological whole-slices and performs background filtering to obtain effective tissue region coordinates. These coordinates are then input into a low-magnification screening model for rapid analysis, generating a lesion probability heatmap. Based on the low-magnification lesion probability value corresponding to each basic coordinate point in the probability heatmap, local peak values ​​are sorted to obtain a dynamic scanning path queue. Dynamic lesion scanning is performed based on this queue, resulting in a set of scanned points. The scan results are then integrated to obtain a comprehensive probability heatmap. Key diagnostic indicators are calculated based on this comprehensive heatmap, and a diagnostic report is generated based on these indicators. This system dynamically adjusts the priority of scanned coordinate points and performs real-time path planning, considering the global suspected probability provided by the low-magnification model while incorporating the local accurate diagnostic results provided by the high-magnification model. This achieves optimal allocation of search resources, ensuring scanning efficiency and significantly improving the detection rate and diagnostic accuracy of diffuse and multifocal lesions within a limited time. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a schematic diagram of the process of a pathological whole-section diagnosis method based on multi-magnification deep learning provided in an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the structure of a pathological whole-slice diagnostic system based on multi-magnification deep learning, provided in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] In one embodiment, such as Figure 1 As shown, a pathological whole-section diagnosis method based on multi-magnification deep learning is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0047] Step 101: Obtain the digital image file of the whole pathological slide and perform background filtering on the digital image file of the whole pathological slide to obtain the coordinates of the effective tissue area.

[0048] Among them, the digital image file of whole pathological slides is a digital image file generated by scanning stained pathological glass slides at high speed and high resolution using a whole slide scanner. It is usually stored in a pyramid model and contains multi-level image data from low magnification to high magnification for subsequent computer-aided analysis and diagnosis.

[0049] Background filtering is an image preprocessing operation that analyzes features such as color, texture, and brightness of an image to automatically identify and remove non-organic regions that have no diagnostic value, thereby concentrating computational resources on meaningful tissue regions.

[0050] Effective tissue region coordinates are usually in the form of a coordinate list, bounding box set, or polygon mask, which accurately marks the spatial location and extent of all pixels containing biological tissue in a whole slice image, and is used to define the region for all subsequent image analysis.

[0051] The terminal loads the digital image file of the whole pathological slide. Based on color space conversion and tissue-specific color features, it can utilize the specific color range of eosinophilic / basophilic components in H&E staining (Hematoxylin and Eosin) to generate a binary mask using threshold segmentation, clustering, or machine learning methods. Morphological operations are performed on the mask to remove noise and small debris and fill in the voids inside the tissue. The contours or minimum bounding rectangles of connected regions are extracted from the processed mask, and their coordinate sets are calculated to obtain the coordinates of the effective tissue region.

[0052] Step 102: Input the effective tissue region coordinates into the low-magnification screening model for rapid analysis to obtain a lesion probability heatmap. Based on the low-magnification lesion probability value corresponding to each basic coordinate point in the probability heatmap, sort the local peak values ​​to obtain a dynamic scanning path queue.

[0053] Specifically, the low-magnification screening model is a trained, lightweight convolutional neural network, for example, a MobileNet, ShuffleNet variant, or a small U-Net, designed for fast processing of downsampled full-slice images. It takes a low-resolution image patch as input and outputs a scalar value between 0 and 1, representing the probability that the image patch contains lesion tissue.

[0054] A lesion probability heatmap is a two-dimensional probability matrix or image generated by a low-magnification screening model after traversing the entire effective tissue area. Its spatial resolution is lower than the original image, but the value of each pixel corresponds to the preliminary probability of a lesion in the corresponding area in the original image. After visualization, heatmap color mapping is often used to intuitively display the distribution of suspicious areas.

[0055] The dynamic scan path queue is an ordered data structure generated based on the lesion probability heatmap. Each element in the queue contains the coordinates of a point to be scanned and its initial priority score, which can be directly taken from the probability value of the heatmap. This queue determines the initial access order of high-magnification scans, ensuring that high-probability areas are processed first.

[0056] The terminal reads the corresponding low-magnification image data in the pyramid hierarchy of the full-slice image based on the coordinates of the effective tissue region. It traverses the entire effective tissue region using a sliding window with a fixed step size and window size. For each image patch captured by a window, the terminal inputs it into a low-magnification screening model to obtain its lesion probability value. The terminal summarizes the center coordinates of all windows and their corresponding probability values, generating a lesion probability heatmap covering the entire tissue through interpolation or direct mapping. The terminal performs local peak detection on this heatmap, using a non-maximum suppression algorithm to find all coordinate points with significantly higher probabilities than their immediate neighbors. The peak points are then sorted from highest to lowest probability, and a dynamic scan path queue is initialized accordingly.

[0057] Step 103: Based on the dynamic scan path queue, perform dynamic scanning of the lesion to obtain the set of scanned points.

[0058] Specifically, dynamic lesion scanning is an iterative, sequential, and sophisticated diagnostic process. Following the order set in the dynamic scanning path queue, high-resolution images are extracted and advanced analyses are performed on the coordinate points in the queue in turn. During this process, the queue status may be updated according to a preset strategy until all predetermined points are scanned or the stopping conditions are met.

[0059] The scanned point set is the result dataset generated after the dynamic scanning process of the lesion is completed. It is a structured collection of records, usually stored in the form of a list or database. Each record contains at least the coordinates of the scanned point, features extracted from the high-magnification image patch, specific probability values ​​obtained from high-magnification model analysis, classification labels derived based on thresholds, and metadata such as timestamps.

[0060] The terminal initiates a scanning loop. In each iteration, the terminal retrieves the highest-priority coordinate point or batch from the head of the dynamic scan path queue. Based on the coordinate point, the terminal backtracks to the original whole-slice digital image file and extracts a high-magnification image patch of fixed size centered on that point at its highest resolution level. This high-magnification image patch is input into a dedicated, deeper, or more complex high-magnification lesion analysis model for detailed analysis to obtain its high-magnification lesion probability value. The terminal compares this probability value with a preset global or adaptive threshold to generate a preliminary lesion conclusion for that point. The terminal adds all information for that point as a complete record to the scanned point set. The loop continues until the dynamic scan path queue is empty, or the preset maximum number of scanned points, maximum time budget, or other stopping conditions are reached. At this point, the loop ends, and the complete scanned point set is output.

[0061] Step 104: Integrate the scanning results of the scanned point set to obtain a probability-integrated heatmap.

[0062] The integration of scan results is a process of data fusion and spatial reconstruction. Data from a discrete set of scanned points, which may be spatially unevenly distributed, is transformed into a continuous, complete diagnostic expression covering the entire tissue slice region in a unified format through techniques such as spatial interpolation, probabilistic fusion, or semantic segmentation post-processing.

[0063] A probabilistic heatmap is a two-dimensional image spatially aligned with the entire tissue section, where the value of each pixel represents the overall probability, inferred from all available high-magnification scan evidence, that the location ultimately belongs to a lesion. The probabilistic heatmap is a core visualization tool for generating diagnostic reports.

[0064] The terminal takes the set of scanned points as input. Due to the influence of dynamic paths on the distribution of scanned points, the terminal spatially aligns the high-magnification analysis results of all scanned points. A spatial interpolation algorithm is employed, which can be radial basis function interpolation, kriging interpolation, or Gaussian process-based regression, to spread the probability values ​​of discrete points onto a two-dimensional grid defined by the coordinates of the entire effective tissue region, generating a preliminary continuous probability field. The terminal can utilize a low-magnification lesion probability heatmap as a prior, or introduce spatial smoothing constraints, to optimize the preliminary probability field, filling in probability estimates for unscanned areas and reducing noise, ensuring spatial consistency of the results. The terminal outputs a high-resolution, smooth probability comprehensive heatmap. Alternatively, the terminal can use superpixel or image patch embedding methods to propagate the features of high-magnification points to their similar low-magnification regions, and then upsample through a lightweight decoder to generate a heatmap.

[0065] Step 105: Based on the probability comprehensive heat map, calculate the key diagnostic indicators, and generate a diagnostic report based on the key diagnostic indicators to obtain the diagnostic report document.

[0066] Key diagnostic indicators (CDIs) are numerical or categorical indicators that are quantitatively calculated from probability-based heatmaps and have clear guiding significance for clinical decision-making. CDIs transform the probability distribution on the image into an objective, quantifiable medical description and are the core data component of the diagnostic report.

[0067] The diagnostic report is a structured electronic document. It integrates key diagnostic indicators, visual heatmaps, screenshots of potential key areas, and text descriptions, aiming to present it to pathologists in a clear and standardized format as an important reference for their final diagnosis.

[0068] The terminal applies one or more diagnostic thresholds to the heatmap, which will determine the continuous probability. Figure 2The data is binarized into clearly defined lesion and non-lesion areas. Based on the binarization results, a series of key diagnostic indicators are calculated. The terminal automatically populates a predefined structured report template with the key diagnostic indicators, thumbnails of the probability comprehensive heatmap or high-magnification screenshots of key areas, and the patient's basic information. The report can include patient information, sample information, analysis summary, a list and values ​​of key diagnostic indicators, visualization results, technical notes, and possible diagnostic hints or classification suggestions, generating a standardized, richly illustrated diagnostic report document.

[0069] For example, key diagnostic indicators include quantitative indicators, such as the percentage of the total area of ​​the lesion region to the total tissue area, the number of lesion regions, the size of the largest lesion, and the morphological descriptor of the lesion region; spatial distribution indicators, such as the spatial distribution pattern of the lesion region and the proximity analysis of whether it invades key structures; and probabilistic statistical indicators, such as the average probability of the whole slice, the probability distribution histogram, and the proportion of high probability regions.

[0070] This embodiment provides a pathological whole-slice diagnostic method based on multi-magnification deep learning. It acquires digital image files of pathological whole-slices and performs background filtering to obtain the coordinates of effective tissue regions. These coordinates are then input into a low-magnification screening model for rapid analysis, resulting in a lesion probability heatmap. Based on the low-magnification lesion probability value corresponding to each basic coordinate point in the probability heatmap, local peak values ​​are sorted to obtain a dynamic scanning path queue. Dynamic scanning of lesions is performed based on this queue, yielding a set of scanned points. The scan results are then integrated to obtain a comprehensive probability heatmap. Key diagnostic indicators are calculated based on this comprehensive heatmap, and a diagnostic report is generated based on these indicators. This method dynamically adjusts the priority of scanned coordinate points and performs real-time path planning, considering the global suspected probability provided by the low-magnification model while incorporating the local accurate diagnostic results provided by the high-magnification model. This achieves optimal allocation of search resources, ensuring scanning efficiency and significantly improving the detection rate and diagnostic accuracy of diffuse and multifocal lesions within a limited time.

[0071] In one embodiment, dynamic scanning of the lesion is performed based on a dynamic scan path queue to obtain a set of scanned points, including:

[0072] Step 201: Extract at least one coordinate point with the highest probability of low-magnification lesions from the dynamic scan path queue to obtain the high-magnification scan coordinate point.

[0073] The dynamic scan path queue is an ordered data structure containing coordinate points to be scanned in detail. These coordinate points are sorted according to their initial level of suspicion, i.e., the probability value of low-magnification lesions, with the points at the front of the queue having the highest priority. This queue is used to guide the order of high-magnification scans.

[0074] High-magnification scan coordinates are one or more coordinates extracted from the dynamic scan path queue. High-magnification scan coordinates represent the locations of tissue regions that are determined to be most needed and given the highest priority for high-magnification fine analysis.

[0075] The terminal accesses the dynamic scan path queue and extracts one or more coordinate points from the front of the queue. The extracted coordinate points are defined as the high-magnification scan coordinate points of the current loop, serving as explicit targets for the next high-magnification analysis.

[0076] Step 202: Based on the high-magnification scan coordinate points, extract the image blocks corresponding to the high-magnification scan coordinate points from the digital image file of the whole pathological slide, and input the image blocks into the high-magnification lesion analysis model to perform lesion probability diagnosis and obtain the high-magnification lesion probability value.

[0077] Specifically, pathological whole-slide digital image files are ultra-high-resolution image files obtained by digitizing pathological glass slides using specialized scanning equipment. They are typically stored in a pyramid structure and contain full-scale image information from low to high magnification. In this embodiment, it is the original data source from which high-resolution image patches are extracted.

[0078] An image patch is a fixed-size rectangular region of image extracted from the highest resolution level of a whole pathological slide digital image file, centered on a high-magnification scan coordinate point. The image patch contains cellular-level morphological details for fine diagnosis.

[0079] High-magnification lesion analysis models are trained, typically more complex and deeper deep learning models than low-magnification screening models, specifically designed for fine analysis of high-resolution image patches to determine the presence and malignancy of lesions.

[0080] The high-magnification lesion probability value is a value output by the high-magnification lesion analysis model after analyzing the input image patch. The high-magnification lesion probability value represents the confidence or probability that the model believes that the specific image patch region belongs to lesion tissue, and its accuracy is higher than the preliminary probability given by the low-magnification model.

[0081] Based on the location information of the high-magnification scan coordinates, the terminal locates the highest resolution image layer in the digital image file of the whole pathological slide. Centered on the high-magnification scan coordinates, it extracts a rectangular region of a predetermined size, thus obtaining a high-resolution image patch. The terminal uses this image patch as input and transmits it to the high-magnification lesion analysis model for forward inference calculations. The model performs in-depth analysis of the image patch's features and outputs a scalar value, namely the high-magnification lesion probability value.

[0082] Step 203: Compare the high-magnification lesion probability value with the lesion threshold to obtain the lesion conclusion for the corresponding high-magnification scan coordinate point, and integrate the high-magnification scan coordinate point, the high-magnification lesion probability value and the lesion conclusion to obtain the set of scanned points.

[0083] Specifically, a lesion threshold is one or a set of predefined numerical criteria used to convert continuous high-magnification lesion probability values ​​into discrete diagnostic categories. Lower and upper thresholds can be set to divide different conclusion intervals.

[0084] The lesion conclusion is a classification judgment obtained by comparing the high-magnification lesion probability value with the lesion threshold. The lesion conclusion is divided into at least three categories: positive, which can be judged as a lesion, with a probability value higher than the upper threshold; negative, which can be judged as non-lesion / normal, with a probability value lower than the lower threshold; and ambiguous, which indicates that it is difficult to make a clear judgment, and its probability value is between the upper and lower thresholds.

[0085] The scanned point set is a data collection used to record the location information of all points that have been scanned and analyzed at high magnification. Each record in the set integrates complete diagnostic information of the high magnification scan coordinate points.

[0086] The terminal compares the high-magnification lesion probability value with a preset lesion threshold. Based on whether the probability value falls within the threshold range (above the positive threshold, below the negative threshold, or in between), the terminal assigns a clear lesion conclusion for that point. The terminal then performs data integration, binding the current point's high-magnification scan coordinates, the high-magnification lesion probability value, and its lesion conclusion to form a complete scan record. This record is added to or updated in the scanned point set. The scanned point set accumulates continuously as the scan cycle progresses, containing a complete diagnostic profile of all points analyzed by the high-magnification model.

[0087] In one embodiment, the lesion conclusion includes positive, negative, and ambiguous;

[0088] After comparing the high-magnification lesion probability value with the lesion threshold to obtain the lesion conclusion at the corresponding high-magnification scan coordinate point, the following steps are also included:

[0089] Step 301: If the lesion conclusion corresponding to the high-magnification scan coordinate point is positive, then generate a coordinate point to be explored within the target radius, centered on the high-magnification scan coordinate point.

[0090] A positive lesion conclusion refers to a diagnosis where, in high-magnification scanning, the probability value of the lesion exceeds a preset positive threshold, indicating a high probability or confirmation of the presence of lesion tissue. This is a condition that triggers subsequent exploratory scanning.

[0091] The target radius is a preset distance parameter, corresponding to a number of pixels or micrometers in the image space. It defines the boundary of a circular exploration area centered on a certain point, and the target radius determines the size of the surrounding exploration range after a positive point is discovered.

[0092] The coordinate points to be explored are a new set of coordinate points generated after a high-magnification scan coordinate point is determined to be positive. These coordinate points are evenly distributed or distributed according to specific rules within a circular area centered on the positive point and with a target radius as the radius. The coordinate points to be explored represent potential new targets that need to be added to the scan queue for subsequent high-magnification analysis, used to explore the boundaries and surrounding conditions of the positive lesion.

[0093] When a lesion at a high-magnification scan coordinate point is determined to be positive, the terminal calculates a circular region centered on the positive coordinate point. The radius of this region is determined by a preset target radius value. Within this circular region, a series of new coordinate points are generated according to certain spatial sampling rules. The newly generated set of coordinate points is the coordinate point to be explored.

[0094] For example, spatial sampling can be performed in either polar or Cartesian coordinate systems using gridded sampling.

[0095] Step 302: Remove the high-magnification scan coordinates from the dynamic scan path queue to obtain the updated dynamic path queue, and add the coordinates to be explored to the updated dynamic path queue to obtain the path to be explored queue.

[0096] Specifically, updating the dynamic path queue involves removing processed intermediate state queues (i.e., those high-magnification scan coordinates that have just completed high-magnification scans and yielded positive results) from the current dynamic scan path queue. Consumed task items are also removed.

[0097] The unexplored path queue is a new, temporary list of coordinates created after the set of unexplored coordinates is added to the updated dynamic path queue. It contains both old unscanned points and newly added unexplored points, but has not yet been optimized or organized.

[0098] The terminal finds and removes the high-magnification scan coordinates that have been extracted and identified as positive from the current dynamic scan path queue, and obtains a new queue that no longer contains the point, i.e., the updated dynamic path queue. The set of coordinates to be explored is added to the tail of the updated dynamic path queue, and the new queue obtained after merging is the queue of paths to be explored.

[0099] Step 303: Based on the step size threshold, deduplication is performed on the queue of paths to be explored to obtain a deduplicated path queue. Based on the priority score, the coordinates in the deduplicated path queue are sorted to obtain a priority path queue.

[0100] Specifically, the step size threshold is a preset minimum spatial distance value used to determine whether two coordinate points are too close in the image space, and can be regarded as duplicate or redundant points. Its purpose is to control the density of scan points and avoid repeatedly scanning adjacent areas that are too close.

[0101] The purpose of deduplication is to eliminate redundant items in the list that are too close or identical in space. In this embodiment, it mainly uses a step size threshold to identify and merge coordinate points that are too close in space.

[0102] The deduplication path queue is the result obtained after deduplicating the queue of paths to be explored. In this queue, the spatial distance between any two coordinate points is not less than the step size threshold, thereby eliminating redundancy caused by overlapping exploration areas of different positive points or new points being too close to existing points.

[0103] Priority score is a numerical value used to quantify the urgency or importance of each coordinate point in the queue. The higher the priority score, the higher the priority.

[0104] The priority path queue is the result of sorting all coordinate points in the deduplication path queue from high to low according to their calculated priority scores. The points in the queue have been rearranged according to the updated overall priority.

[0105] The terminal iterates through all coordinate points in the queue of paths to be explored. For each point, the terminal checks its spatial distance to points already in the queue. If the distance between a point and an existing point is less than a preset step size threshold, the point is considered redundant and is either removed or merged with a nearby point. The resulting queue is the deduplication path queue.

[0106] For each coordinate point in the deduplication path queue, the terminal calculates its priority score according to the formula, and sorts all points in descending order based on the score. The sorted list is the priority path queue.

[0107] Step 304: Based on the pruning rules, prune the priority path queue to obtain the pruned path queue; wherein, the pruned path queue is used to update the dynamic scan path queue.

[0108] Pruning rules are a set of preset conditions or strategies used to filter out a portion of relatively low-priority coordinates from the priority path queue when resources are limited, thereby controlling the queue length and improving overall scanning efficiency. Pruning rules may include: setting a maximum queue length limit, retaining only the first N points; setting an absolute priority score threshold, removing points below the threshold; or combining the coverage of the scanned area to remove new points within areas already fully covered by high-probability scanned points.

[0109] The pruned path queue is the result of processing the priority path queue with pruning rules. It is an optimized queue after redundancy removal, priority sorting, and quantity control.

[0110] The terminal applies pruning rules to the priority path queue. Based on preset rules, the terminal removes coordinate points from the queue that do not meet the retention criteria. After this round of filtering, the remaining coordinate points, in their pre-arranged order within the priority path queue, form the pruned path queue. This pruned path queue is used to update the dynamic scan path queue upon which the dynamic lesion scanning cycle depends, thus intelligently guiding the selection of the next scan target.

[0111] For example, pruning rules include:

[0112] The queue length limit, including a preset integer N, specifies the maximum number of coordinate points that the path queue is allowed to retain after pruning. It is the most direct resource control rule.

[0113] Priority score threshold, including a preset numerical threshold T. Any coordinate point with a priority score lower than T will be considered too low in value and removed.

[0114] The spatial coverage redundancy rule states that if the spatial distance between a point to be scanned and a point in the set of scanned points that has a positive lesion conclusion and a high probability value of high magnification lesion is less than a certain coverage radius, then the area is considered to be represented by a high-confidence result, and the point to be scanned can be removed to avoid redundant scanning.

[0115] The exploration budget ratio is a rule used to balance utilization and exploration. It stipulates that a certain proportion of the retained points must come from areas with high probability values ​​of low-magnification lesions but no surrounding positive scan points, to ensure that the exploration does not completely fall into the local area of ​​the current positive lesion, but can continue to explore new suspicious areas.

[0116] Optionally, the terminal determines whether a point should be removed based on preset rules. If the spatial coverage redundancy rule is triggered, the coordinate point is marked for removal. If the priority score of the coordinate point is lower than the priority score threshold T, it is marked for removal. After applying the above removal rules, the terminal checks the number of remaining points. If the number of remaining points still exceeds the queue length limit N, only the top N points with the highest priority scores are retained. When finally retaining the top N points, it checks whether the exploration budget ratio is met. If 20% of the exploration points are required, but the number of exploration points in the current top N points is insufficient, the lower-ranked and redundant utilization points will be replaced with the higher-ranked exploration points.

[0117] In one embodiment, the priority score is calculated using the following formula:

[0118]

[0119] in, Let i be the priority score for coordinate point i. Let i be the low-magnification lesion probability value at coordinate point i. Let be the high-probability lesion probability value of the parent node of node i. For spatially consistent components, , , These are the weighting coefficients.

[0120] The formula for calculating the spatial consistency component is as follows:

[0121]

[0122] in, Let i be a circular neighborhood with radius R centered at coordinate point i. This represents the number of points within the circular neighborhood that include the already scanned set of points. denoted as the high-magnification lesion probability value of scan coordinate point j within the circular neighborhood.

[0123] The priority score is a comprehensive rating calculated for a coordinate point i to be scanned. This score is used to sort the coordinate points in the dynamic scan path queue, determining the order in which each coordinate point is scanned at high magnification; the higher the score, the higher the priority.

[0124] The low-magnification lesion probability value is the preliminary lesion probability value obtained from the low-magnification screening model analysis of coordinate point i. It represents the initial probability of a lesion existing at this location based on a global rapid scan.

[0125] The high-magnification lesion probability value of the parent node is the high-magnification lesion probability value of the parent node of coordinate point i, that is, the high-magnification scan coordinate point that triggered the generation of point i and has been determined to be positive. It represents the confidence level of finding the source region of this scan point. If the parent node is a very confident positive point, then the points to be explored derived from it should also be given more attention. If there is no parent node, the high-magnification lesion probability value of the parent node is a fixed value.

[0126] The spatial consistency component is a measure reflecting the consistency of diagnostic results between coordinate point i and its surrounding scanned areas. Its calculation relies on a defined circular neighborhood. The spatial consistency component quantifies the degree to which the surrounding environment of point i has been confirmed as positive.

[0127] A circular neighborhood is a circular region centered at coordinate point i with a preset value R as its radius.

[0128] The number of scanned points in the neighborhood is the total number of coordinate points located in the set of scanned points within the circular neighborhood.

[0129] The high-magnification lesion probability value is the precise lesion probability value obtained by the high-magnification lesion analysis model for each scanned point j located within the circular neighborhood.

[0130] The weighting coefficients are three preset constants greater than or equal to 0, used to weight and balance the relative importance of the three components in the final priority score. By adjusting the coefficients, the bias of the scanning strategy can be changed.

[0131] By dynamically calculating priorities using this formula, the terminal can allocate limited high-magnification scanning resources preferentially to the most likely, most needed, and information-gaining regions. This avoids ineffective scanning in low-probability areas or areas with already defined characteristics, significantly improving the overall efficiency of whole-slice scanning diagnosis while ensuring that major lesions are not missed.

[0132] In one embodiment, the low-magnification screening model is obtained through the following method:

[0133] Step 501: Obtain the historical annotation dataset, and perform downsampling based on the historical annotation dataset to obtain low-magnification images, and perform image segmentation based on the low-magnification images to obtain low-magnification image patches.

[0134] The historical labeled dataset refers to a set of labeled pathological image data used during the model development phase. Each sample in this dataset typically contains a high-resolution digital image of a whole pathological slide, along with the outline of the lesion area precisely delineated on the image by a pathologist. It serves as the original material and knowledge source for training the model.

[0135] Downsampling is an image processing operation that generates a lower-resolution version of an image by reducing its resolution, i.e., reducing the number of pixels. In digital pathology, it can also refer to reading a lower-resolution image from the pyramid hierarchy of digital pathology slide images than the highest resolution.

[0136] Low-magnification images are obtained by downsampling the original high-resolution whole-slice images from historical labeled datasets. They are smaller in size and less detailed, but retain the overall structure and general morphology of the tissue, and are used to simulate the low-magnification field of view used in rapid screening.

[0137] Image segmentation refers to the operation of cutting a complete low-magnification image into multiple smaller, regularly arranged rectangular sub-images.

[0138] Low-magnification image patches are small, fixed-size image fragments obtained by image segmentation operations from low-magnification images. These low-magnification image patches serve as the basic input units for subsequent annotation and model training.

[0139] The terminal acquires a historical annotation dataset, which contains original high-resolution whole-slice digital images of pathology and their corresponding lesion region annotation masks. Next, it downsamples each high-resolution whole-slice digital image in the dataset to generate a corresponding low-resolution, low-magnification image. On the low-magnification image, the terminal performs sliding window-style image segmentation with a fixed step size and window size, thereby obtaining a large number of uniformly sized low-magnification image patches covering different regions of the image.

[0140] Step 502: Based on the low-magnification image patch, calculate the proportion of pixel area in the covered tissue region that belongs to the lesion annotation region to obtain soft labels, and integrate the low-magnification image patch and soft labels to obtain the training sample set.

[0141] Specifically, a tissue-covered region refers to the pixel area within a low-magnification image patch that actually contains biological tissue. Not all pixels in an image patch are tissue; edge areas may contain a small amount of non-tissue regions remaining from the background filtering process.

[0142] A lesion annotation region refers to a lesion area projected onto a corresponding low-magnification image or low-magnification image patch based on the original high-resolution annotation information through coordinate mapping. The lesion annotation region is used to identify which pixel locations correspond to the lesions annotated by the pathologist at low magnification.

[0143] The pixel area ratio is a calculated value that refers to the ratio of the number of pixels in the area where the covered tissue region intersects with the lesion-labeled region in a low-magnification image patch (i.e., the tissue region belonging to the lesion) to the total number of pixels in the entire covered tissue region of the image patch.

[0144] Soft labels are a direct application of pixel area ratio calculations. A soft label is a continuous value representing the probability or degree to which an image patch contains a lesion. It provides a more precise reflection of the proportion of diseased tissue within an image patch.

[0145] The training sample set is a collection of data pairs that integrates low-magnification image patches and their computed soft labels. It serves as the direct input data for model training.

[0146] The terminal uses a color threshold to distinguish tissue from the background and identify the tissue-covering region within the image patch. Simultaneously, based on the original annotation information, the terminal determines the lesion annotation region mapped to the spatial location of the image patch. The number of pixels in the intersection of the two regions is calculated and then divided by the total number of pixels in the tissue-covering region to obtain the pixel area ratio. This pixel area ratio is assigned as the soft label for the image patch. All low-magnification image patches are paired with their respective calculated soft labels to form a training sample set.

[0147] Step 503: Based on the training sample set, low-magnification image patches with soft labels greater than the sample threshold are marked as suspected positive, and the remaining low-magnification image patches are marked as confirmed negative, thus obtaining positive and negative samples.

[0148] Specifically, the sample threshold is a preset critical value used to discretize continuous soft labels, transforming them into explicit categories for supervised learning.

[0149] A suspected positive result refers to a category label assigned to low-magnification image patches with soft label values ​​greater than the sample threshold. These image patches contain a high percentage of diseased tissue and are considered samples that are very likely to contain lesions.

[0150] A negative confirmation refers to the category label assigned to low-magnification image patches whose soft label values ​​are less than or equal to the sample threshold. These image patches contain a very low or zero percentage of diseased tissue and are considered to be normal tissue samples with virtually no lesions.

[0151] Positive and negative samples are collectively referred to as the set of samples that have been labeled and have clear category information. Positive samples are suspected positive samples, and negative samples are confirmed negative samples.

[0152] The terminal iterates through each sample in the training sample set. It compares the soft label value of each sample with a preset sample threshold. If the soft label is greater than the threshold, the sample is marked as a suspected positive; otherwise, it is marked as a confirmed negative. After iterating through and marking all samples, a set of positive and negative samples consisting of suspected positive and confirmed negative samples is obtained, preparing for subsequent supervised training.

[0153] Step 504: Divide the positive and negative samples to obtain the model training dataset, and train the lightweight convolutional neural network based on the model training dataset to obtain the low-magnification screening model.

[0154] The model training dataset is a combination of data obtained by partitioning the positive and negative sample sets, and is directly used for training and evaluating machine learning models. The model training dataset contains at least a training set and a validation set. The training set is used for learning model parameters, and the validation set is used for monitoring performance and tuning hyperparameters during training. It may also include a test set.

[0155] Lightweight convolutional neural networks (CNNs) are a class of specially designed CNN architectures with low parameter count and computational complexity. Common examples include lightweight versions of MobileNet, ShuffleNet, and EfficientNet. They are characterized by small model size and fast inference speed, making them suitable for deployment in tasks requiring intensive inference across the entire graph and demanding high speed, such as fast analysis tasks.

[0156] The low-magnification screening model is a pre-trained, deployable, lightweight convolutional neural network instance. It learns a mapping from low-magnification image patches to their lesion probabilities, enabling it to quickly output a lesion probability score for new, unseen low-magnification image patches.

[0157] The terminal randomly but hierarchically divides the positive and negative sample sets into non-overlapping subsets according to a certain ratio, thereby forming the model training dataset and ensuring that the model can learn and be evaluated on independent data.

[0158] The terminal trains a lightweight convolutional neural network using a pre-defined model training dataset. The terminal inputs image patches from the training set into the network for forward propagation to obtain predicted values; it then calculates the error between the predicted values ​​and the true labels using a loss function; and updates the network parameters through backpropagation and an optimizer to minimize the loss. This process iterates multiple times on the training set and periodically evaluates performance on the validation set to prevent overfitting. Training stops when the model's performance on the validation set stabilizes or reaches its optimum, and the network parameters at this point are saved, resulting in the final low-magnification screening model.

[0159] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0160] Based on the same inventive concept, this application also provides a pathological whole-slice diagnostic system based on multi-magnification deep learning for implementing the aforementioned pathological whole-slice diagnostic method based on multi-magnification deep learning. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the pathological whole-slice diagnostic system based on multi-magnification deep learning provided below can be found in the limitations of the pathological whole-slice diagnostic method based on multi-magnification deep learning described above, and will not be repeated here.

[0161] In one exemplary embodiment, such as Figure 2 As shown, a pathological whole-slice diagnostic system 600 based on multi-magnification deep learning is provided, including:

[0162] The filtering module 601 is used to acquire digital image files of whole pathological slides and perform background filtering on the digital image files of whole pathological slides to obtain the coordinates of the effective tissue area;

[0163] The path module 602 is used to input the coordinates of the effective tissue area into the low-magnification screening model for rapid analysis, obtain a lesion probability heatmap, and sort the local peaks based on the low-magnification lesion probability value corresponding to each basic coordinate point in the probability heatmap to obtain a dynamic scanning path queue.

[0164] The scanning module 603 is used to perform dynamic scanning of lesions based on a dynamic scanning path queue to obtain a set of scanned points;

[0165] Integration module 604 is used to integrate the scanning results of the scanned point set to obtain a probabilistic comprehensive heat map;

[0166] The diagnostic module 605 is used to calculate key diagnostic indicators based on the probability comprehensive heat map, and generate a diagnostic report based on the key diagnostic indicators to obtain a diagnostic report document.

[0167] Furthermore, the scanning module 603 is used for:

[0168] Extract at least one coordinate point with the highest probability of low-magnification lesions from the dynamic scan path queue to obtain high-magnification scan coordinate points;

[0169] Based on high-magnification scan coordinate points, image blocks corresponding to high-magnification scan coordinate points are extracted from the digital image files of whole pathological sections, and the image blocks are input into the high-magnification lesion analysis model to perform lesion probability diagnosis and obtain high-magnification lesion probability values.

[0170] By comparing the high-magnification lesion probability value with the lesion threshold, the lesion conclusion corresponding to the high-magnification scan coordinate point is obtained. The high-magnification scan coordinate point, the high-magnification lesion probability value and the lesion conclusion are integrated to obtain the set of scanned points.

[0171] Furthermore, the lesion conclusions include positive, negative, and ambiguous;

[0172] The system also includes an update module for:

[0173] If the lesion conclusion corresponding to the high-magnification scan coordinate point is positive, then coordinate points to be explored are generated within the target radius with the high-magnification scan coordinate point as the center;

[0174] Remove the high-magnification scan coordinates from the dynamic scan path queue to obtain the updated dynamic path queue, and add the coordinates to be explored to the updated dynamic path queue to obtain the path to be explored queue.

[0175] Based on the step size threshold, the queue of paths to be explored is deduplicated to obtain a deduplicated path queue. Based on the priority score, the coordinates in the deduplicated path queue are sorted to obtain a priority path queue.

[0176] Based on pruning rules, the priority path queue is pruned to obtain the pruned path queue; the pruned path queue is used to update the dynamic scan path queue.

[0177] Furthermore, the formula for calculating the priority score is as follows:

[0178]

[0179] in, Let i be the priority score for coordinate point i. Let i be the low-magnification lesion probability value at coordinate point i. Let be the high-probability lesion probability value of the parent node of node i. For spatially consistent components, , , These are the weighting coefficients;

[0180] The formula for calculating the spatial consistency component is as follows:

[0181]

[0182] in, Let i be a circular neighborhood with radius R centered at coordinate point i. This represents the number of points within the circular neighborhood that include the already scanned set of points. denoted as the high-magnification lesion probability value of scan coordinate point j within the circular neighborhood.

[0183] Furthermore, the system also includes a training module for:

[0184] Obtain the historical annotation dataset, and based on the historical annotation dataset, perform downsampling to obtain low-magnification images, and based on the low-magnification images, perform image segmentation to obtain low-magnification image patches;

[0185] Based on low-magnification image patches, the proportion of pixel area in the covered tissue region that belongs to the lesion annotation region is calculated to obtain soft labels. The low-magnification image patches and soft labels are then integrated to obtain a training sample set.

[0186] Based on the training sample set, low-magnification image patches with soft labels greater than the sample threshold are marked as suspected positive, and the remaining low-magnification image patches are marked as confirmed negative, thus obtaining positive and negative samples;

[0187] The positive and negative samples are divided to obtain the model training dataset. Based on the model training dataset, a lightweight convolutional neural network is trained to obtain a low-ratio screening model.

[0188] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a pathological whole-section diagnostic method based on multi-magnification deep learning as described above.

[0189] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0190] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0191] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A pathological whole-section diagnostic method based on multi-magnification deep learning, characterized in that, The method includes: Obtain digital image files of whole pathological slides, and perform background filtering on the digital image files of whole pathological slides to obtain the coordinates of the effective tissue regions; The effective tissue region coordinates are input into the low-magnification screening model for rapid analysis to obtain a lesion probability heatmap. Based on the low-magnification lesion probability value corresponding to each basic coordinate point in the probability heatmap, local peaks are sorted to obtain a dynamic scanning path queue. Based on the dynamic scanning path queue, dynamic scanning of the lesion is performed to obtain a set of scanned points; The scan results of the scanned point set are integrated to obtain a probabilistic comprehensive heatmap; Based on the probabilistic comprehensive heatmap, key diagnostic indicators are calculated, and based on the key diagnostic indicators, a diagnostic report is generated to obtain a diagnostic report document.

2. The method according to claim 1, characterized in that, The dynamic scanning of the lesion based on the dynamic scanning path queue, to obtain a set of scanned points, includes: Extract at least one coordinate point with the highest probability of low-magnification lesion from the dynamic scan path queue to obtain the high-magnification scan coordinate point; Based on the high-magnification scan coordinate points, image blocks corresponding to the high-magnification scan coordinate points are extracted from the digital image file of the whole pathological slide, and the image blocks are input into the high-magnification lesion analysis model to perform lesion probability diagnosis and obtain high-magnification lesion probability values. The high-magnification lesion probability value and the lesion threshold are compared to obtain the lesion conclusion corresponding to the high-magnification scan coordinate point. The high-magnification scan coordinate point, the high-magnification lesion probability value and the lesion conclusion are integrated to obtain the scanned point set.

3. The method according to claim 2, characterized in that, The lesion conclusions include positive, negative, and ambiguous; After comparing the high-magnification lesion probability value with the lesion threshold to obtain the lesion conclusion corresponding to the high-magnification scan coordinate point, the method further includes: If the lesion conclusion corresponding to the high-magnification scan coordinate point is positive, then coordinate points to be explored are generated within the target radius with the high-magnification scan coordinate point as the center; Remove the high-magnification scan coordinate points from the dynamic scan path queue to obtain an updated dynamic path queue, and add the coordinate points to be explored to the updated dynamic path queue to obtain a path queue to be explored. Based on the step size threshold, the queue of paths to be explored is deduplicated to obtain a deduplicated path queue. Based on the priority score, the coordinate points in the deduplicated path queue are sorted to obtain a priority path queue. Based on pruning rules, the priority path queue is pruned to obtain a pruned path queue; wherein, the pruned path queue is used to update the dynamic scan path queue.

4. The method according to claim 3, characterized in that, The formula for calculating the priority score is: in, Let i be the priority score for coordinate point i. Let i be the low-magnification lesion probability value at coordinate point i. Let be the high-probability lesion probability value of the parent node of node i. For spatially consistent components, , , These are the weighting coefficients; The formula for calculating the spatial consistency component is as follows: in, Let i be a circular neighborhood with radius R centered at coordinate point i. The number of points in the circular neighborhood that include the scanned point set. The high-magnification lesion probability value is the scan coordinate point j within the circular neighborhood.

5. The method according to claim 1, characterized in that, The low-magnification screening model was obtained through the following method: Obtain a historical annotation dataset, and based on the historical annotation dataset, perform downsampling to obtain a low-magnification image, and based on the low-magnification image, perform image segmentation to obtain low-magnification image patches; Based on the low-magnification image patch, the proportion of pixel area in the covered tissue region that belongs to the lesion annotation region is calculated to obtain a soft label, and the low-magnification image patch and the soft label are integrated to obtain a training sample set; Based on the training sample set, the low-magnification image patches with soft labels greater than the sample threshold are marked as suspected positive, and the remaining low-magnification image patches are marked as confirmed negative, thus obtaining positive and negative samples; The positive and negative samples are divided to obtain a model training dataset. Based on the model training dataset, a lightweight convolutional neural network is trained to obtain the low-magnification screening model.

6. A pathological whole-slice diagnostic system based on multi-magnification deep learning, characterized in that, The system includes: The filtering module is used to acquire digital image files of whole pathological slides and perform background filtering on the digital image files of whole pathological slides to obtain the coordinates of the effective tissue region; The path module is used to input the coordinates of the effective tissue area into the low-magnification screening model for rapid analysis, obtain a lesion probability heatmap, and sort the local peaks based on the low-magnification lesion probability value corresponding to each basic coordinate point in the probability heatmap to obtain a dynamic scanning path queue. The scanning module is used to perform dynamic scanning of lesions based on the dynamic scanning path queue to obtain a set of scanned points; The integration module is used to integrate the scanning results of the scanned point set to obtain a probabilistic comprehensive heat map. The diagnostic module is used to calculate key diagnostic indicators based on the probability comprehensive heatmap, and generate a diagnostic report based on the key diagnostic indicators to obtain a diagnostic report document.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A 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 according to any one of claims 1 to 5.

Images