A case consultation and pathological section automatic diagnosis system based on AI analysis

Abstract

The application relates to the technical field of medical informatization, and discloses a case consultation and pathological section automatic diagnosis system based on AI analysis, which comprises a server and a client; the server generates a lesion binary mask by using a deep learning model, constructs a sparse quadtree index carrying a composite state bit mask, and generates a vector section with the background removed; the client splits resource requests to a multistage queue according to the bit mask in the index, and preferentially loads high-value area images; a three-layer heterogeneous rendering architecture is adopted, and adaptive Gaussian blur is used to process the vector section to realize smooth transition; meanwhile, a behavior audit module collects viewport tracks based on a physical rendering completion state, calculates high-risk area reading coverage, and controls diagnosis report submission permission. The application effectively reduces the transmission bandwidth of gigapixel-level sections, improves loading flicker, and reduces the risk of missed diagnosis caused by network delay or human negligence.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to an AI-based automated diagnosis system for case consultation and pathological slides. Background Technology

[0002] Web-based image slide (WSI) technology has driven the digitization of pathological diagnosis, but its typically gigapixel-level data size poses challenges to network bandwidth and real-time client rendering capabilities. Most existing web-based pathology slide reading systems employ a static slide distribution mechanism based on a pyramid model, where the client directly requests slides based solely on viewport coordinates, lacking semantic awareness of image content. This mechanism results in the system indiscriminately loading large amounts of blank background areas with no diagnostic significance during transmission, consuming limited network resources and causing problems such as high initial loading latency and slow response times.

[0003] In weak network environments such as remote consultations, or when doctors are rapidly dragging and dropping images, transmission delays in high-resolution image data often result in frequent grid-like blank spaces or loading flickering within the viewport, disrupting the visual continuity for doctors observing tissue morphology. Furthermore, existing computer-aided diagnostic systems primarily focus on improving the accuracy of AI algorithms in identifying lesions, neglecting process quality control of the doctor's actual image reading process. Traditional log recording only tracks dwell time, failing to effectively distinguish whether the doctor is observing clear tissue images or waiting for blurry, unloaded images, and making it difficult to enforce preventative measures against missing high-risk areas due to doctor fatigue, distraction, or network latency.

[0004] Therefore, this invention proposes an AI-based automatic diagnosis system for case consultation and pathological slides to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an AI-based automatic diagnosis system for case consultation and pathological slides, which solves the problems of high loading latency caused by invalid data transmission during remote viewing of gigapixel-level fully digital slides, visual discontinuity in viewport rendering, and the risk of missed diagnoses due to the lack of process quality control based on the physical rendering completion status.

[0006] This invention provides an AI-based system for automatic diagnosis of case consultations and pathological slides.

[0007] The system includes server-side devices and client-side devices. The server-side devices are equipped with an image analysis module, an index building module, and a resource generation module; the client-side devices are equipped with a data transmission module, an image rendering module, and a behavior auditing module.

[0008] Preferably, the image analysis module of the server-side device receives the fully digital slice image, calls a deep convolutional neural network model to infer and output a lesion probability distribution map, and performs a morphological dilation operation based on the physical radius on the distribution map. The binary mask generated by the morphological dilation operation can cover the core lesion and the surrounding microenvironment area. The index construction module divides the fully digital slice image into a quadtree grid, and assigns a composite state bitmask to the grid nodes according to the binary mask and the coordinates of the clinical text data. The composite state bitmask includes visual saliency bits and clinical annotation bits. The system performs a bottom-up pruning operation on the quadtree grid, merges child nodes with the same composite state bitmask value and retains the parent node, and serializes the pruned data structure into a binary index stream. This step, through pre-processed image semantic analysis, realizes the structured extraction of spatial distribution features of pathological images.

[0009] Preferably, the resource generation module of the server device generates image slice files from the all-digital slice image and generates vector slice files with background removed. When generating the vector slice files, the system divides the all-digital slice image into patches, calculates the perceived brightness value of pixels within each patch, and filters out a set of valid organized pixels whose perceived brightness value is lower than the background removal threshold. The system calculates the weighted average of the pixel colors in the valid organized pixel set and uses this weighted average as the color data of the vector slice file. If the number of valid organized pixels is lower than a preset value, the corresponding vector slice file is marked as fully transparent.

[0010] Preferably, the data transmission module of the client device receives the binary index stream and builds a local query tree. The data transmission module queries the local query tree based on the viewport coordinates, reads the composite status bitmask of the corresponding node, and distributes resource requests to different priority queues according to the status value. When the visual saliency bit of the composite status bitmask is valid, the request for the corresponding area is allocated to the high-priority queue to load image tile files; when the visual saliency bit is invalid, the request for the corresponding area is allocated to the low-priority queue to load vector tile files. Simultaneously, the system performs time window de-jitter processing on the requests in the low-priority queue. Furthermore, the data transmission module has a proxy routing control function. When the viewport stays in an area with an invalid visual saliency bit for more than a preset threshold or the zoom level exceeds a cell-level threshold, the data transmission module forcibly increases the request priority for that area, requesting image tile files or real-time rendered image streams from the server device to replace the vector tile files.

[0011] Preferably, the client device's image rendering module constructs a full-viewport canvas. The full-viewport canvas employs a three-layer heterogeneous overlay architecture: the bottom layer is a static raster layer used to draw image slice files; the middle layer is a dynamic vector layer used to draw vector slice files; and the top layer is an interactive overlay layer used to draw text labels in spatially labeled slice files. The image rendering module applies adaptive dynamic Gaussian blur processing to the middle layer, setting the blur radius to be inversely proportional to the viewport scaling level. After the image slice files of a specific area at the bottom layer are rendered, the system clears the vector slice files and blur effects of the corresponding area in the middle layer. Before drawing text labels at the top layer, the system samples the average brightness of background pixels within the bounding box area of ​​the text labels. If the contrast between the average brightness of the background pixels and the text color is lower than a preset threshold, the system automatically switches the text color or draws a semi-transparent background panel below the text labels to maintain the visual contrast between the text information and the organizational background.

[0012] Preferably, the behavior auditing module of the client device collects viewport trajectory data and calculates the viewing coverage of high-risk areas by determining whether the viewport state vector at the current moment meets the valid viewing conditions. Valid viewing conditions include: the viewport zoom level being greater than the minimum diagnostic magnification threshold; the viewport movement speed being less than the visual stability threshold; the viewport dwell time in the same area being greater than the cognitive processing time threshold; and the image slice file within the viewport range completing physical rendering loading. The behavior auditing module compares the viewing coverage with a preset safety threshold in real time. When the viewing coverage is lower than the safety threshold, the system locks the diagnostic report submission permission, calculates the geometric center of the uncovered high-risk areas, and renders a dynamic navigation marker pointing to that geometric center on the top layer of the full viewport canvas. The system unlocks the diagnostic report submission permission only when the viewing coverage reaches the safety threshold. This mechanism, combining physical rendering state and visual dwell time characteristics, establishes a quality control approach that constrains the viewing process.

[0013] This invention provides an AI-based automated system for case consultation and pathological slide diagnosis. It offers the following advantages:

[0014] 1. This invention constructs a sparse quadtree index carrying a composite state bitmask, enabling the client to anticipate the content attributes of a slice region before initiating a network request. Combined with a multi-level request queue and de-jitter logic, the system prioritizes loading images of high-value regions containing lesions or clinical annotations, while loading only very lightweight vector slices for large areas of blank background or temporarily ignoring them during rapid movement. This content-aware, on-demand distribution mechanism reduces the amount of invalid data transmission during the browsing of gigapixel-level fully digital slices, solving the problems of high initial screen loading latency and interactive stuttering in weak network environments.

[0015] 2. This invention employs a rendering architecture based on a three-layer heterogeneous overlay and generates vector slices using the weighted average color of the tissue after background removal. While the high-resolution image slices are still being downloaded or decoded, the system uses an adaptive Gaussian blur algorithm to process the vector slices in the intermediate layers, dynamically adjusting the blur radius according to the scaling level to simulate the macroscopic texture of the tissue. This processing method improves upon the grid-like blanking or loading flickering phenomena commonly found in traditional block loading techniques, ensuring visual continuity for doctors when quickly dragging or zooming through images.

[0016] 3. This invention establishes a behavior auditing mechanism based on physical rendering completion confirmation, taking the completion of DOM element loading for image slices as a necessary condition for valid image review, thereby eliminating invalid gaze duration when the image is blurry due to network congestion. By calculating the image coverage rate of high-risk areas in real time and locking the submission permission of diagnostic reports, the system can force doctors to supplement their review of uncovered suspected lesion areas through dynamic navigation markers, effectively avoiding the risk of missed diagnoses due to human negligence, fatigue, or network latency. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0018] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0019] Figure 3 This is a schematic diagram comparing resource scheduling and performance response during the interactive process of the present invention;

[0020] Figure 4 This is a schematic diagram illustrating the correlation between audit coverage and the risk of missed diagnoses in this invention.

[0021] Among them, 100 is the server-side equipment; 110 is the image analysis module; 120 is the index building module; 130 is the resource generation module; 210 is the data transmission module; 220 is the image rendering module; 230 is the behavior auditing module; and 300 is the data network. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1The present invention provides an AI-based case consultation and pathological slide automatic diagnosis system, including a server device 100, a client device 200 and a data network 300.

[0024] Server-side device 100 is configured to perform storage, computation, and distribution of pathological image data, including an image analysis module 110, an index building module 120, and a resource generation module 130. The image analysis module 110 receives fully digital slice images and clinical text data, runs a segmentation algorithm to generate a lesion probability distribution map, performs morphological dilation on the lesion probability distribution map covering a region with a physical radius of 500 micrometers, and outputs a binary mask. The index building module 120 divides the fully digital slice images based on a quadtree structure, assigns a composite state bit mask containing visual saliency bits and clinical annotation bits to the quadtree nodes, and performs bottom-up pruning to generate a binary index stream. The resource generation module 130 performs pyramid slicing on the fully digital slice images to generate image slice files, calculates the tissue-weighted average color after removing background pixels with brightness values ​​greater than 240 to generate vector slice files, and maps clinical feature coordinates to a pixel coordinate system to generate spatially labeled slice files.

[0025] The client device 200 is configured to handle image data requests and rendering, including a data transmission module 210, an image rendering module 220, and a behavior auditing module 230. The data transmission module 210 receives a binary index stream and builds a local query tree. Based on the viewport coordinates, it queries the local query tree to distribute resource requests to different priority queues, performing de-jitter processing on low-priority requests. The image rendering module 220 constructs a full-viewport canvas, overlaying and drawing standard image slices, vector slices, and spatial annotation slices on the full-viewport canvas. It applies dynamic blurring to the vector slice layer and adjusts the display style of the spatial annotation slices according to the background brightness. The behavior auditing module 230 collects viewport trajectory data, calculates the effective image viewing time and image viewing coverage, and controls the submission permissions of diagnostic reports based on the image viewing coverage.

[0026] See attached document Figure 2 This invention provides a method for automatic diagnosis of case consultations and pathological slides based on AI analysis, comprising the following steps:

[0027] S100: The server device receives fully digital slice images, uses the image analysis module to generate lesion probability maps and performs morphological dilation processing, the index building module builds a quadtree index based on the dilated region and encodes the node attributes into a composite state bitmask, and the resource generation module generates image slices, vector slices with highlighted backgrounds removed, and clinical data spatial annotation slices.

[0028] S200: The client device initializes the consultation session, downloads the compressed quadtree index stream, reconstructs the quadtree structure in local memory, and establishes a spatial query interface.

[0029] S300, the client device responds to the viewport change query spatial query interface, the data transmission module allocates the coordinates of the plot within the viewport to a multi-level request queue according to the query results, immediately requests image slices for high-value areas, delays requesting vector slices for low-value areas, and independently requests spatially labeled slices for areas containing clinical annotations.

[0030] S400: The client device performs multi-layer heterogeneous rendering. The image rendering module draws image slices at the bottom layer, vector slices in the middle layer and applies dynamic Gaussian blur, and spatial annotation slices at the top layer and calculates background pixel brightness in real time to dynamically adjust the annotation text color.

[0031] S500: The client device performs proxy routing control. When the interaction triggers certain conditions, it forcibly increases the priority of low-value area requests and downloads image tiles to replace vector tiles.

[0032] S600: The client device performs movie viewing behavior audit. The behavior audit module records viewport coordinates, zoom level and dwell time, calculates the effective movie viewing coverage in high-risk areas, and locks the report submission function and renders navigation icons at the top level if the coverage does not meet the standard.

[0033] To further clarify the implementation of each technical aspect of this embodiment, the following will provide a detailed description of the implementation of each functional module involved above and its internal processing flow.

[0034] See attached document Figure 1 In constructing the system described in this embodiment, considering the special characteristics of the pathological diagnosis scenario, such as huge data throughput and extremely high real-time requirements, a distributed physical architecture with decoupled computing and storage was adopted.

[0035] In the physical topology of this embodiment, the server device 100 is no longer a single server entity, but a distributed cluster composed of compute nodes and storage nodes. To address the numerous matrix convolution operations involved in AI model inference, the compute nodes are equipped with high-performance graphics processing unit (GPU) arrays. These arrays are specifically configured to support mixed FP16 half-precision and FP32 single-precision floating-point operations to improve computational throughput while ensuring inference accuracy. Specifically, to accommodate images typically reaching gigapixel levels in all-digital slicing, the video memory capacity of each compute node is set to be no less than 1.5 times the video memory required for input tensor batch processing. This redundancy design aims to prevent inference interruptions caused by video memory overflow during concurrent input of high-resolution image blocks. Simultaneously, the storage nodes employ disk arrays composed of non-volatile memory, configured in RAID 0+1 mode, thereby providing sufficient IOPS support for high-concurrency random access of image pyramid files.

[0036] Corresponding to the server, the client device 200 serves as the terminal for doctors to conduct diagnostic interactions. Its form factor includes, but is not limited to, desktop graphics workstations, portable tablets, or mobile smart terminals. Regardless of the hardware form factor, the client device 200 runs in a standard browser environment that supports the Canvas drawing interface and WebGL / WebGPU rendering engines. In this environment, the client's central processing unit (CPU) is primarily responsible for the main thread's logical control and network I / O scheduling, while the graphics processing unit (GPU) is invoked to execute image compositing algorithms in the fragment shader. Hardware-accelerated bilinear interpolation and color space conversion ensure visual smoothness during scaling and panning operations.

[0037] Data network 300 serves as the system's communication infrastructure, connecting server device 100 and client device 200. Considering the large number of small file requests generated during pathology slide browsing, this embodiment preferably uses HTTP / 2 or HTTP / 3 protocols as the transmission medium. The system can transmit multiple image slices and vector data streams in parallel over a single TCP / QUIC connection. This mechanism improves upon the head-of-line blocking phenomenon commonly found in the traditional HTTP / 1.1 protocol, reducing first-screen loading latency in weak network environments.

[0038] Furthermore, to meet the stringent requirements of multi-party collaboration in remote consultation scenarios, this embodiment establishes an independent WebSocket long-lived connection channel between the server device 100 and the client device 200, in addition to the basic static resource distribution channel. This channel is not used for transmitting image data, but is specifically used for full-duplex transmission of lightweight viewport state vectors and control signaling. To accurately describe the viewing field of view at any given time and achieve multi-device synchronization, this embodiment defines a viewport state vector. :

[0039] ;

[0040] In the formula: This represents the x and y coordinates of the current viewport center point in the absolute pixel coordinate system of the all-digital slice. Its data type is defined as a double-precision floating-point number, and its value range is as follows: and (in The purpose of using double precision instead of integer (for the physical width and height of the slice) is to ensure sub-pixel level alignment accuracy at high magnification (such as 40x and above) and avoid image jitter during synchronization. This represents the current continuous scaling factor, with a value range of 1. This corresponds to the equivalent magnification of the microscope, and this parameter supports stepless scaling rather than being limited to discrete pyramid levels. The UNIX timestamp representing the state was generated, with a precision of milliseconds (ms). This is a 128-bit unique identifier (UUID) for the current consultation session.

[0041] Based on the above vector definition, the client device has built-in strict timing filtering logic to handle network jitter. Let the timestamp of the latest locally processed frame be... The current state vector timestamp received via WebSocket is The system will execute the following judgment logic: only when Only when the frame is considered a valid new state and a view update is triggered, and the view is updated accordingly. Conversely, if If the frame is determined to be an out-of-order or expired frame due to network routing delay, the system will discard the data packet without rendering it. This logic ensures that, in complex public network environments, the multi-party consultation view always maintains a monotonous forward-moving timeline, avoiding view rollback or jumps caused by out-of-order data.

[0042] See attached document Figure 1 The image analysis module 110 is configured to convert fully digital slice images into spatial index base data with semantic information. To construct a deep learning model capable of recognizing specific pathological features, this embodiment first establishes a training dataset. Specifically, historically diagnosed pathological slice data is collected, and pathologists annotate cancerous areas and microenvironments at the pixel level, generating a binary mask as the gold standard. The slices are cropped into 512×512 pixel sample blocks, and positive and negative sample sets are constructed at a 1:3 ratio. Data augmentation operations such as rotation, flipping, and color jitter are performed on the samples. A deep convolutional neural network model is trained based on this dataset until the loss value on the validation set converges.

[0043] During the actual inference phase, the image analysis module 110 receives fully digital slice images as model input, feeding only the image data into the trained deep convolutional neural network model. It should be noted that the received clinical text data (such as immunohistochemistry reports) does not participate in the inference of the aforementioned deep learning model; instead, it is directly transmitted to the subsequent index building module 120 to assist in generating clinical annotation bits in the composite state bitmask.

[0044] As the starting point of the entire automated diagnostic pipeline, the image analysis module 110 does not attempt... Figure 1 Loading gigapixel-level fully digital slices (WSI) into video memory in a single operation is impractical under current hardware conditions. Instead, this embodiment employs a memory-mapped sliding window mechanism to handle these pyramid-structured files, typically stored in SVS, NDPI, or TIFF formats. Under this mechanism, the system sets the extraction window size to [size missing]. (e.g., 1024×1024 pixels), and raster scanning is performed at the highest resolution level (Level 0). Considering the "edge effect" that often occurs when convolutional neural networks process image edges, resulting in feature loss or decreased prediction accuracy, this embodiment specifically introduces overlapping regions between windows, with an overlap rate of... Setting it to [0.15, 0.25], this redundant sampling design, while increasing computational overhead slightly, effectively avoids noticeable mesh artifacts in the final stitched image.

[0045] After obtaining the image patch tensor, the system inputs it into a pre-built deep convolutional neural network model. To ensure that those skilled in the art can reproduce the performance of this model, this embodiment explicitly adopts an asymmetric "encoder-decoder" architecture. Specifically, the encoder path uses a ResNet-50 pre-trained on ImageNet as the backbone network, removing its original fully connected layers and retaining only the convolutional blocks from Conv1 to Conv5 to extract multi-scale abstract semantic features; the decoder path contains four upsampling modules, each of which expands the feature map size by a factor of 2 through bilinear interpolation and fuses the feature maps of the corresponding encoder layers through lateral connections. This architecture design is crucial for pathological image analysis because it can restore spatial resolution while preserving microscopic texture details that are highly valuable for diagnosis. The model's output layer contains a 1×1 convolutional layer and a sigmoid activation function, compressing the number of channels to 1 and outputting a single-channel probability map with values ​​ranging from (0,1).

[0046] It is worth noting that, given the often extreme imbalance between lesion areas (positive samples) and normal tissue background (negative samples) in pathological images, the training process of this model employs a specific hybrid loss function. Its definition is as follows:

[0047] ;

[0048] In the formula: The binary cross-entropy loss focuses on optimizing pixel-level classification accuracy. The Dice coefficient loss is defined as follows: (in For the prediction graph, (For real labels), focusing on optimizing the overall overlap of lesion areas; and The weighting coefficient is set to 0.5 in this embodiment to achieve a balance between pixel accuracy and region integrity.

[0049] After model inference is completed, the system needs to reassemble the discrete probability patches into a global probability distribution map. To eliminate stitching artifacts, the system does not use simple hard cutting in overlapping regions, but instead applies a linear weighted fusion strategy: for any pixel within the overlapping region... The final probability value is the distance-weighted average of the predicted values ​​of two adjacent windows.

[0050] Subsequently, to construct a fault-tolerant index structure, the system performs binarization and morphological augmentation operations on the generated probability distribution map. The system sets a probability threshold. (Typically a value of 0.5) Generate the initial binary mask However, considering the extremely high prognostic value of "infiltrative margins" in pathological diagnosis, and to avoid the risk of local missed detections that may exist in AI models, this embodiment enforces a morphological dilation operation based on physical scales. This operation is not based on a fixed number of pixels, but on real physical distances, and its calculation logic is defined as follows:

[0051] ;

[0052] ;

[0053] In the formula: Defined as the physical expansion radius, set at 500μm, this value is based on the minimum field of view required by pathologists to identify microinvasive lesions under low magnification. Defined as the scan resolution (μm / pixel) of the WSI slice, this parameter is preferentially read from the metadata in the file header; Defined as a default resolution constant, with a value of 0.25μm / pixel (corresponding to the scanning accuracy of a standard 40x objective lens), this fallback mechanism is used to handle abnormal situations when metadata is missing or corrupted, avoiding program crashes caused by division by zero; Indicates radius as The disk-shaped structural element was chosen instead of a rectangle to ensure isotropic expansion and avoid introducing unnatural geometric artifacts. This represents the morphological dilation operator.

[0054] Generated through the above steps The mask not only covers high-confidence cancerous areas but also includes a safety buffer of a certain width. This process ensures that the subsequently generated quadtree index can fully cover the microenvironment surrounding the lesion, allowing the client to obtain key contextual information about the lesion's edge even when requesting data at a low scaling rate, thus supporting doctors in making accurate TNM staging decisions.

[0055] See attached document Figure 1After the image analysis module generates high-dimensional pixel-level mask data, the index building module 120 is then started. Its core task is to convert these discrete, unstructured binary image data into a compact, streamable sparse quadtree structure.

[0056] To address the challenge of rapid retrieval of gigapixel-level pathological images under limited network bandwidth, this embodiment does not directly transmit the original pixel data. Instead, it constructs a logical navigation graph. The index construction module first establishes a logical quadtree grid that maps to the pyramid hierarchy of the original fully digital slice (WSI). The system is based on the physical resolution of the WSI (… The maximum depth of the quadtree is calculated using the given slice size and preset slice dimensions (usually 256×256 pixels). In this structure, each leaf node corresponds precisely to a specific slice at the bottom of the pyramid in spatial geometry, thus establishing a bijective relationship between the logical index and physical storage.

[0057] Subsequently, the system traverses all leaf nodes of the quadtree and calculates a "composite state bitmask" for each node based on multi-source heterogeneous data. To decouple the index structure from the feature data, this embodiment employs a multi-bit binary encoding mechanism, enabling a single node to simultaneously carry multi-dimensional attributes such as visual saliency, clinical annotation, and even gene expression. Let any node... The composite state bitmask is Its computational logic is defined as follows:

[0058] ;

[0059] In the formula: This indicates the total number of supported feature dimensions. In this embodiment, it is initially set to 32 bits (i.e., stored using the uint32 type) and can be extended to 64 bits. Indicates a bit index; Indicates the first An indicator function for a dimensional feature, which takes the value 0 or 1.

[0060] Specifically, for the most critical visual and clinical features in this embodiment, the indicator function is defined as follows:

[0061] For Bit 0 (visual saliency bit), the system calculates the spatial region corresponding to this node. Binarized saliency mask generated in the preceding steps The percentage of the intersection area. Indicator function. The judgment logic is as follows:

[0062] ;

[0063] in, The noise filtering threshold is set to 0.01 (i.e., 1%) in this embodiment. This threshold is set to filter out tiny disconnected regions caused by dust on the slide or noise from the scanner sensor, preventing excessive splitting of the quadtree due to invalid high-frequency details, thereby ensuring the sparsity of the index.

[0064] For Bit 1 (clinical annotation bit), the system will use the set of spatial coordinates from the received clinical text data. (Such as positive spots of fluorescence in situ hybridization (FISH) probes, surgical boundary markers) are mapped to a quadtree coordinate system. Indicator function The judgment logic is as follows:

[0065] ;

[0066] This logic means that as long as the node region contains at least one valid clinical annotation point, the bit is marked as 1, ensuring that key information annotated by doctors is not missed due to scaling. It's worth noting that, to ensure backward compatibility of the system architecture, this bitmask design reserves high-order space from Bit 2 to Bit N. For example, when the system accesses spatial transcriptome data from next-generation sequencing (NGS), Bit 2 can be directly assigned to represent "highly expressed gene mutation regions" without needing to reconstruct the underlying index data structure or recompile the client parser.

[0067] After assigning attribute values ​​to the leaf nodes, the system executes a bottom-up recursive pruning and merging algorithm. This step utilizes the spatial local correlation of the image to remove redundant information. The algorithm starts from the maximum depth... Begin by traversing upwards, level by level, until you reach the root node. For any parent node... and its four child nodes The system performs a homogeneity check:

[0068] ;

[0069] In the formula: This represents the operation logic on the parent node; Indicates the first The composite state bitmask value of each child node.

[0070] If the bitmask values ​​of the four child nodes are completely identical (e.g., all 0, indicating meaningless background; or all 3, indicating areas containing lesions and labeled), the system determines that the region has spatial homogeneity and performs a Merge operation: freeing the four child nodes from memory and directly assigning the unified bitmask value to the parent node. Simultaneously, the parent node is marked as "leaf state". Conversely, if there are differences between child nodes, the parent node retains the "branch state" and maintains the index pointers to the child nodes. This process effectively aggregates large areas of blank background or continuous lesion regions into a single high-level node, greatly reducing the amount of data transmission.

[0071] Finally, the system serializes the pruned quadtree structure into a compact binary stream. To improve network transmission efficiency, this embodiment employs a breadth-first search (BFS) strategy to write the node queue into a buffer, and the data structure of each node is compressed into a format of "1 byte depth + 1 byte type + 4 bytes bitmask". This binary index stream serves as a navigation map for subsequent client resource requests, ensuring that the client only needs to download a very small index file (typically less than 500KB) once to fully understand the macroscopic distribution and microscopic details of gigapixel-level images.

[0072] See attached document Figure 1 Once the index building module completes the logical quadtree planning, the resource generation module 130 is activated. Its core function is to transform massive amounts of raw pixel data into physical file units suitable for network distribution, specifically covering raster image slices, lightweight vector slices, and sparse labeled data slices.

[0073] As a crucial bridge connecting server-side storage and client-side rendering, the resource generation module 130 first performs a standardized image pyramid slicing task. In this embodiment, the system follows common slicing standards such as OpenSeadragon or Google Maps to establish a three-dimensional coordinate system (z,x,y), where z represents the scaling level (z=0 corresponds to the highest resolution), and x and y represent the grid column and row numbers. For each level, the system segments the original WSI image into tiles of fixed resolution (typically 256×256 pixels). Considering the conflicting demands for "viewing speed" and "diagnostic accuracy" in pathology slide reading scenarios, this embodiment employs an adaptive image compression strategy: for low magnification levels (e.g., z≤10), the system prioritizes generating WebP format files with high compression ratios to minimize initial screen loading delay; while for high magnification diagnostic levels (e.g., z>10), the system automatically switches to generating JPEG format files with low compression ratios (setting a quality factor Q≥90). This differentiated processing ensures extremely low bandwidth usage during macroscopic browsing, while preventing the chromatin texture of cell nuclei from being smoothed out by the compression algorithm during microscopic observation.

[0074] To provide a preview of the tissue outline before the high-resolution image download is complete, this embodiment generates a vector tile file (.vec). Unlike traditional methods that only store the image, this step calculates the "tissue-weighted average color" of each tile and encapsulates it into very small binary metadata.

[0075] When calculating the average color, simply taking the arithmetic mean of all pixels would severely dilute the tissue color due to the large bright white background of the slide, causing the generated thumbnail to appear washed out and thus lose its navigational purpose. Therefore, the system executes a pixel filtering algorithm based on perceived brightness. For any pixel within the tile... The system first calculates its perceived brightness according to the ITU-R BT.601 standard. :

[0076] ;

[0077] Subsequently, the system statistically analyzes the set of effectively organized pixels. The selection criteria are .in, In this embodiment, the background removal threshold is set to 240. This threshold is selected based on the optical characteristics of H&E stained sections: the brightness of the blank glass area is usually between 245 and 255, while even the lightest cytoplasmic stained area usually has a brightness below 235.

[0078] Based on the filtered pixel set, the system calculates the RGBA values ​​of the vector slices. To prevent division by zero errors in blank areas, a minimum pixel constraint is introduced into the calculation logic:

[0079] ;

[0080] In the formula: For set The total number of pixels in; This represents the summation of the channel values ​​corresponding to all pixels within the set; This is an opacity constant with a value of 255 (corresponding to complete opacity). The minimum effective pixel count is set to 5% of the total number of pixels in the tile (e.g., approximately 3200 pixels in a 256×256 tile). This parameter serves as a protection logic for numerical stability, preventing the calculation of meaningless color values ​​in areas containing only a very small amount of noise (such as dust) or completely blank areas. If the effective pixels are insufficient, the system directly outputs a fully transparent value (0,0,0,0), indicating that the area does not need to be rendered. The final generated .vec file occupies only 4 bytes, greatly reducing network transmission overhead.

[0081] When generating spatial label tile files (.json), the resource generation module 130 does not blindly generate corresponding metadata files for all tiles, but instead executes a "sparse generation strategy" to optimize storage space and I / O performance. This step is highly dependent on the composite state bitmask in the quadtree index generated by the preceding module.

[0082] The system traverses the quadtree index, locking only nodes whose Bit 1 (clinical annotation bit) is 1. For each matching node, the system extracts its covered global physical coordinate range and retrieves a set of feature point coordinates falling within that range from the clinical text database. Subsequently, in order for the client to render directly at the current layer, the system must map the global coordinates to relative pixel coordinates within the current slice. This mapping process must take into account the scaling effect caused by the pyramid hierarchy, and its calculation formula is as follows:

[0083] ;

[0084] ;

[0085] In the formula: The WSI global absolute pixel coordinates of the feature point at Level 0; For the current scaling level index ( ); , The column and row indices of the current slice within the grid; , The physical dimensions of the slice; This indicates the floor function.

[0086] After the calculation is complete, the system will contain The GeoJSON object is serialized and written to the storage system. For nodes with bit 1 set to 0, the server skips the file generation step. This "on-demand instantiation" mechanism ensures that the file system is not filled with tens of thousands of empty JSON files (avoiding the innode exhaustion problem), and at the same time allows the client to directly determine "no data" by using the preloaded index when requesting a non-existent tag file without initiating an invalid HTTP request, reducing the server's concurrency pressure.

[0087] See attached document Figure 1 The data transmission module 210, as the logical control unit of the client, runs in the user's browser or native client application. Its core function is to parse the lightweight index generated by the server and formulate the optimal resource loading strategy based on the user's real-time operation intention or remote collaboration command.

[0088] During the initialization phase of the consultation session, the data transmission module does not rush to request large amounts of image data. Instead, it prioritizes acquiring the binary index stream generated by the preceding module. The client uses a pre-built parser to deserialize this binary stream and reconstructs the image in local memory. The quadtree interface QueryTree(z,x,y) provides time-complexity query capabilities. In this embodiment, this interface not only stores the spatial topology of the image, but more importantly, it carries the composite state bitmask of each node. This allows the client to know "what content" and "whether it is worth requesting" of a region before initiating any network request, thereby reducing bandwidth waste caused by blind polling.

[0089] To support remote pathology consultations and teaching scenarios, this embodiment establishes a multi-source driven viewport state update mechanism. The system's input source is defined as local human-computer interaction events. (Including mouse drag, scroll wheel zoom, touch gestures) and remote collaboration commands (Master status frame received via WebSocket full-duplex channel).

[0090] Considering the unavoidable latency and jitter in network transmission, directly applying remote coordinates would cause discontinuous jumps in the viewport, resulting in visual discontinuity. Therefore, the system maintains a viewport state vector in real time. It is then smoothed using an "exponentially weighted moving average" algorithm. Its update logic is defined as follows:

[0091] ;

[0092] ;

[0093] In the formula: express The target viewport state at any given time, including the center point coordinates. and scaling levels ; As a control mode factor, it is divided into two states: "independent viewing" and "following". The sequence number is a monotonically increasing sequence number for remote command frames; the system only accepts sequences with a sequence number greater than the previous processed sequence number. The instructions are to prevent viewport rollback due to network out-of-order issues; This refers to the rendering viewport state from the previous moment; This is a smoothing coefficient, with a value range of (0, 1). In this embodiment... Set as a dynamic value: when (i.e., when a large jump occurs) A value of 1.0 is used to achieve instantaneous response; when there is a small movement, A value of 0.3 is used to filter out hand tremors or network noise.

[0094] Once the viewport status Upon a change, the module calculates the set of tiles currently covered by the viewport and executes a "multi-level request queue distribution strategy" for each tile. This step is the core of the bandwidth adaptive scheduling implementation in this embodiment. The system reads the composite status bitmask corresponding to the tile. It routes requests to different priority queues based on their characteristic bits.

[0095] Specific routing decision logic The definition is as follows:

[0096] ;

[0097] In the formula: This is a visual saliency bitmask constant (corresponding to Bit 0); This is a clinical annotation bitmask constant (corresponding to Bit 1); This is a high-priority queue used for loading JPEG / WebP image slices (.img). This queue employs a LIFO (Last-In, First-Out) strategy to ensure that the image currently in the user's field of view is loaded first. This is an independent annotation queue used to load GeoJSON vector annotation files (.json). This queue executes in parallel with image loading, without blocking each other, ensuring the immediate presentation of clinical information. This is a low-priority queue used to load vector slices (.vec) of the background region. This is a proxy routing flag used to force priority elevation.

[0098] For those distributed to The request (i.e., corresponding to) (For the non-significant region where Bit 0 is 0), the system does not immediately initiate an HTTP request, but instead performs de-jitter processing with time window constraints. The system sets a time threshold. The time window is 50ms. If the viewport moves out of the viewport within this window, the scheduled task is canceled. This design is based on the physical characteristics of pathology slide reading: when doctors quickly drag to find lesions, the viewport sweeps across a large amount of blank background. Loading these transient areas at this time is not only meaningless, but also consumes the browser's limited concurrent connections. Only when the viewport stays in a blank area for more than 50ms does the system determine that the user intends to view the overview color of that area, and then initiate a very lightweight .vec file request.

[0099] To address potential false negatives in AI preprocessing, or in scenarios where doctors genuinely need to examine "normal tissue" for comparative analysis, this embodiment introduces proxy routing logic. When the interaction module detects that the user's mouse hovers over an area marked as "background" (Bit 0=0) for more than a preset duration (e.g., 500ms), or the user manually increases the zoom level to the cellular level (e.g., z>15), the system will forcibly reset the settings. .

[0100] Based on the aforementioned formula, Setting this flag will trigger an "on-demand upgrade" mechanism: requests to this region will bypass the low-priority queue and be directly promoted to higher priority. The image is queued, and the `&render=dynamic` parameter is appended to the request URL. At this point, the client ignores the background markers in the index and forces a request for a high-resolution image tile of that area from the server. If the server detects that the static tile file for the corresponding path does not exist (due to storage optimization during resource generation), it routes to the real-time rendering interface, dynamically cropping and compressing the image based on the original large image data and returning the image stream. This combined hardware and software strategy ensures both extreme performance under normal conditions and data accessibility and rigorous diagnostics under extreme circumstances.

[0101] See attached document Figure 1 Once the data transmission module delivers resource streams of different priorities to the client, the image rendering module 220 takes over the processing logic and is configured to achieve smooth rendering of gigapixel-level images under limited browser computing power.

[0102] To decouple the rendering logic of data with different frequencies and avoid screen flickering caused by full redraws, the system first constructs a full viewport canvas container. In this embodiment, the container adopts a three-layer heterogeneous overlay architecture: the bottom layer (Layer 1) serves as a static raster layer, specifically used for rendering high-resolution JPEG / WebP slices, carrying high-fidelity pathological morphological information; the middle layer (Layer 2) serves as a dynamic vector layer, used for rendering lightweight color placeholders; and the top layer (Layer 3) serves as an interactive overlay layer, using HTML5 Canvas or SVG technology to draw clinical annotation and measurement tools. This layered design ensures that high-frequency mouse interactions (such as drawing measurement lines) only require refreshing the top-level canvas, without triggering the heavily overhead redraw of the bottom layer, thus improving the rendering frame rate.

[0103] In scenarios with fluctuating network conditions or rapid dragging, the underlying high-definition tiles often experience physical transmission delays. To eliminate the visual "gray-white grid" waiting effect, the system performs fast pre-rendering based on vector data in the intermediate layer. The rendering engine parses the .vec file obtained in the previous steps, reads the encapsulated RGBA color values, and draws rectangular color blocks at the corresponding coordinates.

[0104] At this point, if discrete color blocks are directly presented, the image will exhibit obvious discrete block artifacts, resulting in visual discontinuity. Therefore, this embodiment is configured to execute an "adaptive dynamic Gaussian blur" algorithm. This algorithm does not use a fixed blur kernel, but instead calculates the blur radius in real time based on the current viewport scaling state and device characteristics. Its calculation model is constructed based on the "hierarchical inverse law":

[0105] ;

[0106] In the formula: The pixel radius of the Gaussian blur kernel; This represents the absolute scaling level of the current viewport, typically ranging from 0 to... ; The device pixel ratio is used to eliminate rendering differences between Retina screens and regular displays, ensuring that the physical size of the blur effect is consistent. To prevent the denominator from approaching zero, the value is set to 0.1 for the numerically stable term. This is a visual smoothing constant, set to 20 in this embodiment. The physical meaning of this parameter is: when the user is at a low macroscopic magnification (… When the magnification is relatively small, the system applies a larger fuzz radius to fuse discrete tissue color patches into a continuous heatmap, simulating the macroscopic texture trend of the tissue; while when the user zooms in to a microscopic magnification (…), the system… When it is relatively large, Rapid decay reduces blurring interference with the potential cell textures that are about to be loaded, achieving a natural visual transition from macroscopic "seeing the trend" to microscopic "seeing the details".

[0107] Meanwhile, the underlying rendering pipeline loads high-priority .img image slices in parallel. Once an image slice at a certain location has been downloaded and decoded via asynchronous I / O, the system does not immediately perform a hard replacement, but instead executes a smooth inter-layer transition. Using alpha blending technology, the system adjusts the transparency of the underlying image. The transition is linear from 0 to 1.

[0108] To avoid GPU performance degradation caused by dual rendering, the system introduces an "inter-layer occlusion culling" mechanism. Once a specific region of the underlying layer is occluded... of When the value reaches 1.0 (completely opaque), the system immediately sends a command to clear the corresponding area of ​​the intermediate layer. The vector color blocks and blurring effect. This dynamic memory management strategy ensures that once the image is fully sharp, the GPU no longer wastes computing power rendering the occluded blurry placeholders.

[0109] When rendering the clinical annotation information (such as text labels and regions of interest (ROI) boundaries) contained in the top-level .json file, in order to solve the text readability problem caused by the complex background of the pathological image (dark cell nuclei intertwined with light glass background), this embodiment executes a "smart avoidance" rendering algorithm based on region statistics.

[0110] Drawing arbitrary text labels Previously, the rendering engine first calculated the center region of the tag's bounding box. (For example, a 10×10 pixel range), and calls the ctx.getImageData interface to sample the background pixel matrix of the bottom or intermediate layers. Subsequently, to avoid misjudgment caused by single-point sampling hitting noise points (such as impurities or isolated cell nuclei), the system calculates the average perceived brightness of the pixels in this area. :

[0111] ;

[0112] Based on the calculated average brightness value, the system dynamically determines the rendering strategy for foreground text. :

[0113] ;

[0114] In the formula: Fill the text with color; The brightness contrast threshold is set to 128 in this embodiment, which corresponds to the median value of the 8-bit grayscale space; The decision hysteresis interval is set to 15. This interval is introduced to handle areas of luminance blur. When the background brightness... When the text falls within the range [113, 143], it indicates that the contrast between the background and text colors is insufficient to support effective human visual recognition. At this point, the system triggers... Logically, draw a layer of transparency below the text. The semi-transparent gray background ensures that critical clinical information remains clearly readable against any complex tissue background, reducing the risk of misdiagnosis due to visual confusion.

[0115] See attached document Figure 1 To reduce the risk of missed diagnoses due to doctor fatigue, negligence, or network latency, the behavior audit module 230 is configured as a background daemon that runs silently. It constructs an objective diagnostic quality assessment system by quantifying the distribution of doctors' visual attention in real time.

[0116] The system starts a high-frequency sampling timer (in this embodiment, the sampling frequency is set). (Hz), real-time acquisition of the current viewport's state vector. This vector contains the viewport center coordinates. Current digital scaling level Mouse or touch point movement speed and image loading status indicator .

[0117] As a key improvement in this embodiment This is not a simple network request status flag, but a confirmation signal that pixel rendering is complete at the physical level. Specifically, this flag is deeply bound to the rasterization rendering pipeline at the underlying layer (Layer 1). The system only activates this flag when a high-fidelity slice within the viewport triggers the OnLoad event of a DOM element and is fully drawn onto the Canvas. Set to True. This logical judgment aims to distinguish between "the doctor is observing a clear tissue morphology" and "the doctor is waiting for a blurry placeholder to become clear," thereby eliminating invalid gaze duration caused by network congestion or decoding delays and ensuring the authenticity of audit data.

[0118] Based on the collected state vectors, the system performs real-time determination of "valid image viewing". To filter out unconscious browsing during rapid dragging, the system defines a valid image viewing determination function. Only at a certain moment The viewport area is only included in the effective diagnostic range when it simultaneously meets four criteria: sufficient magnification, smooth movement, sufficiently long dwell time, and clear image. The logical criteria are as follows:

[0119] ;

[0120] In the formula: The minimum diagnostic magnification threshold is set to 20x in this embodiment (corresponding to digital level z=12). This threshold is based on common sense in pathology: observations below 20x are usually only used to determine tissue structure (such as glandular arrangement) but cannot clearly distinguish nuclear atypia (such as mitotic figures), and therefore do not have diagnostic significance. This represents the instantaneous movement speed of the viewport; The visual stability threshold is set to 50 pixels per second. This parameter is based on the vestibular oculomotor reflex characteristics of the human eye: when the speed of the moving image exceeds this value, the retinal slippage speed is too large, causing the dynamic blur generated by the human eye to exceed the resolving power of the visual cortex, making it impossible to form a clear image; This refers to the cumulative fixation time in the same area. The cognitive processing time threshold is set to 500ms. This value corresponds to the minimum psychophysical time required for humans to complete feature recognition from visual signal input, ensuring that doctors have sufficient time to understand the image content.

[0121] To achieve low-latency coverage calculation, the system maintains a virtual raster mask corresponding to the full slide size, with a resolution mapping ratio of 1:256. All satisfying... The viewport area is projected onto this mask, and bitwise operations are used to mark the corresponding grid cells as 1, forming a "viewed area set". .

[0122] Subsequently, the module reads the metadata generated during the index building phase and extracts all node regions where Bit 0 is 1 (i.e., marked by the AI ​​algorithm as having suspected lesions or high value), forming a "high-risk region set". To quantify the completeness of this diagnosis, the system calculated the quality control coverage rate. Its computational model uses a piecewise function definition:

[0123] ;

[0124] In the formula: This represents the geometric area of ​​a set of regions to be calculated. In engineering, this is achieved by counting the number of set pixels in the raster mask. The intersection operation represents the set operation; The minimum threshold; when When (i.e., for negative slides that have been pre-screened by AI and determined to be completely normal), the formula enters the second branch, which then calculates the comprehensive review index: The total effective reading time for doctors; The baseline slide reading time (e.g., 120 seconds) is based on the amount of tissue in the slide. The area of ​​the region randomly sampled by the doctor for examination; This represents the total tissue area on the slice; The weighting coefficient is 0.5 in this embodiment. This branch logic ensures that even negative slides must undergo careful review for a sufficient period of time and area, reducing the risk of missed diagnoses due to AI-induced false negatives.

[0125] The system will calculate in real time The value is compared with the preset quality control threshold. (For example, 95%) is compared. As long as... The system will forcibly lock the "Submit for Review" button on the diagnostic report interface, making it grayed out and uninterrupted.

[0126] Simultaneously, the behavior auditing module sends a command to the top layer (Layer 3) of the image rendering module, triggering the drawing of the "dynamic navigation markers". The system first calculates the difference region. And calculate using the K-Means clustering algorithm. Geometric center of the largest connected region Then, render the pointer at the edge of the screen. The pulsed, flashing arrows. This interaction mechanism directly guides the doctor's viewpoint to unexamined high-risk areas until the doctor has completed effective image reading of these areas (i.e., satisfies the aforementioned effective image reading determination function). (defined four-dimensional conditions), such that Once the criteria are met, the submitting party's access is unlocked. This mandatory closed-loop quality control process, from a technical perspective, constructs a quality control mechanism of "AI early warning - behavior audit - mandatory review".

[0127] To verify the technical effectiveness of the system of the present invention in a real medical scenario, this embodiment constructs a cross-regional remote consultation environment for gastric cancer pathology for detailed explanation.

[0128] In this specific embodiment, the server first receives a fully digital slice of a gastroscopy biopsy, with a data volume of approximately 2.4GB. The background computing node then starts, calling a deep convolutional neural network model integrated with a ResNet-50 backbone to extract features from the image data. This model is specifically configured to identify the pathological features of signet ring cell carcinoma. It is worth noting that after identifying a suspected lesion area, the system does not merely mark the geometric boundaries, but performs a morphological dilation operation based on a physical radius of 500 micrometers, thereby generating a binary mask covering the core lesion and its surrounding microenvironment. Subsequently, the index building module constructs a sparse quadtree index based on this mask, mapping the high-dimensional image data into a binary index stream, laying the foundation for efficient subsequent transmission.

[0129] When a doctor initiates a consultation on a client configured as a general-purpose laptop, the data transmission module preferentially loads the aforementioned binary index stream. (See attached...) Figure 3 The resource scheduling curve shown in the diagram illustrates that during the initial browsing phase (0-12 seconds), doctors perform extensive and rapid dragging and zooming operations. At this time, the dashed line representing existing technology shows that frequent requests for high-resolution slices cause significant fluctuations in bandwidth usage, and the end-to-end rendering latency remains consistently above 600ms, easily leading to dropped frames or delayed response. In contrast, the system in this embodiment (shown by the solid line in the diagram) employs a multi-level request queue distribution strategy. For a large number of non-lesion background areas, the system downgrades high-resolution slice requests to low-priority vector slice requests. Therefore, in... Figure 3Within the 5-12 second range, the system's bandwidth usage remained consistently low and smooth, with rendering latency consistently below 100ms. This processing method visually preserved the macroscopic texture and tone of the tissue, avoiding the "grayish mesh" phenomenon. Only when the doctor's viewport coordinates stopped and zoomed in to the high-risk area marked by the index around the 15th second did the system trigger the "on-demand upgrade" mechanism, immediately loading the underlying high-resolution image slices. At this point, only a brief and controllable small peak appeared on the bandwidth curve, achieving precise response to key diagnostic areas.

[0130] Furthermore, to ensure diagnostic security, the behavior auditing module runs continuously as a background process. (See attached...) Figure 4 The correlation curve between audit coverage and missed diagnosis risk shown illustrates how the system collects viewport state vectors in real time and calculates the doctor's "coverage" of high-risk areas. As the viewing time progresses, the coverage curve rises in a step-like manner, while the "probability of missed diagnosis risk," derived from the inverse mapping of coverage, shows a monotonically decreasing trend. Figure 4 At time t1 (approximately second 30), the doctor attempts to click the submit button. At this point, the system detects that the coverage curve is approximately 75%, which has not yet reached the preset 95% safety threshold, and the corresponding risk of missed diagnoses remains high. Therefore, the system executes a logic interception and renders a dynamic navigation icon at the top layer of the view. Until time t2 (approximately second 48), the doctor, guided by the navigation, completes the review of the remaining high-risk areas. The coverage curve officially surpasses the 95% safety threshold, the probability of missed diagnoses approaches zero, and the system determines that the safety conditions are met, thus unlocking the submission permission, i.e., "Review Completed / Submission Allowed".

[0131] Based on the above mechanism, this embodiment organized a comparative test involving 100 gastrointestinal pathological slides. Experimental results showed that, in a weak network environment with a simulated bandwidth limit of 4Mbps, the average first-screen loading time of the experimental group was 0.4 seconds, a reduction of approximately 85% compared to the control group's 2.8 seconds. More importantly, thanks to… Figure 4 The mandatory review mechanism shown in the diagram successfully blocked all submission attempts that did not cover high-risk areas in the experimental group, achieving a 0% false negative rate for small lesions. In contrast, the control group experienced an 8% false negative rate for small lesions under the same conditions. In summary, this embodiment, by introducing sparse index flow and behavior auditing closed loop, effectively constructs a quality control mechanism for pathological diagnosis while reducing network resource overhead.

[0132] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based automated system for case consultation and pathological slide diagnosis, characterized in that, This includes server-side equipment and client-side equipment; The server-side device includes: an image analysis module for extracting features from the fully digital slice image to generate a binary mask covering the lesion region; an index construction module for constructing a quadtree index based on the binary mask and generating a binary index stream, wherein the index construction module divides the fully digital slice image into a quadtree grid and assigns a composite state bitmask to the nodes of the quadtree grid; the composite state bitmask includes at least: a visual saliency bit indicating that the node region intersects with the binary mask, and a clinical annotation bit indicating that the node region contains clinical text data coordinates; the index construction module performs a bottom-up pruning operation on the quadtree grid, and when the composite state bitmask values ​​of the child nodes to which the parent node belongs are consistent, the child nodes are merged and the parent node is retained, and finally the pruned structure is serialized into the binary index stream; and a resource generation module for generating image slice files and background-removed vector slice files. The client device includes: a data transmission module, used to receive the binary index stream and establish a local query tree, query the local query tree according to viewport coordinates, and divert resource requests to different priority queues according to the visual saliency bits of the corresponding node composite state bitmask; an image rendering module, used to construct a full viewport canvas, and overlay and draw the image slice file and the vector slice file on the full viewport canvas; and a behavior auditing module, used to collect viewport trajectory data, identify node areas with valid visual saliency bits as high-risk areas, calculate the viewing coverage rate of high-risk areas, and control the submission permission of diagnostic reports according to the viewing coverage rate. The behavior auditing module calculates the viewing coverage rate by determining whether the viewport state vector at the current moment meets the valid viewing conditions. The valid viewing conditions include: the viewport zoom level is greater than the minimum diagnostic magnification threshold, the viewport movement speed is less than the visual stability threshold, the viewport dwell time in the same area is greater than the cognitive processing time threshold, and the image slice file within the viewport range has completed physical rendering loading.

2. The AI-based case consultation and pathological slide automatic diagnosis system according to claim 1, characterized in that, The image analysis module calls a deep convolutional neural network model to infer the probability distribution map of lesions from the fully digital slice image. The image analysis module further performs a morphological dilation operation based on physical radius on the lesion probability distribution map, converting the lesion probability distribution map into the binary mask, so that the binary mask covers the core lesion and the surrounding microenvironment.

3. The AI-based case consultation and pathological slide automatic diagnosis system according to claim 1, characterized in that, The resource generation module generates the vector slice file in the following ways: The fully digital slice image is divided into patches, and the perceived brightness value of the pixels within each patch is calculated. The set of valid organization pixels whose perceived brightness value is lower than the background culling threshold is statistically analyzed. Calculate the weighted average of the pixel colors in the effective organized pixel set, and use the weighted average as the color data of the vector slice file; if the number of pixels in the effective organized pixel set is lower than a preset value, then mark the vector slice file as fully transparent.

4. The AI-based case consultation and pathological slide automatic diagnosis system according to claim 1, characterized in that, The data transmission module is configured with multi-level request queues; The data transmission module queries the local query tree based on the viewport coordinates and reads the composite status bitmask of the corresponding node; If the visual saliency bit of the composite state bitmask is valid, the request for the region of the corresponding node is allocated to a high-priority queue to load the image slice file; If the visual saliency bit of the composite state bitmask is invalid, the request for the region of the corresponding node is allocated to a low-priority queue to load the vector slice file, and time window de-jittering processing is performed on the requests in the low-priority queue.

5. The AI-based case consultation and pathological slide automatic diagnosis system according to claim 4, characterized in that, The data transmission module also performs proxy routing control; When the viewport stays in a region where the visual saliency bit of the composite state bitmask is invalid for a longer than a preset threshold, or the zoom level exceeds a cell-level threshold, the data transmission module forcibly increases the request priority of the region and requests the image slice file or a real-time rendered image stream from the server device to replace the vector slice file.

6. The AI-based case consultation and pathological slide automatic diagnosis system according to claim 1, characterized in that, The image rendering module adopts a three-layer heterogeneous overlay architecture; The bottom layer is a static raster layer, used to draw the image slice file; The intermediate layer is a dynamic vector layer used to draw the vector slice file; The image rendering module applies adaptive dynamic Gaussian blur processing to the intermediate layer, and the blur radius of the Gaussian blur processing is inversely proportional to the scaling level of the viewport; Once the image slice file of a specific region in the underlying layer has been rendered, the image rendering module clears the vector slice file and blur effect of the corresponding region in the intermediate layer.

7. The AI-based case consultation and pathological slide automatic diagnosis system according to claim 1, characterized in that, The resource generation module is also used to generate spatially labeled slice files containing clinical feature coordinates; The top layer of the image rendering module is an interactive overlay layer, used to draw text labels in the spatial annotation slice file; Before drawing the text label, the image rendering module samples the average brightness of the background pixels within the bounding box area of ​​the text label; If the contrast between the average brightness of the background pixels and the text color is lower than a preset threshold, the text color will be automatically switched or a semi-transparent background will be drawn below the text label.

8. The AI-based case consultation and pathological slide automatic diagnosis system according to claim 1, characterized in that, The behavior auditing module compares the image coverage rate with a preset security threshold in real time. If the image coverage rate is lower than the security threshold, the behavior audit module locks the diagnostic report submission permission, calculates the geometric center of the uncovered high-risk area, and renders a dynamic navigation marker pointing to the geometric center on the top layer of the full viewport canvas. The behavior audit module unlocks the permission to submit diagnostic reports only when the image coverage reaches the security threshold.

Images