Intelligent detection method and system for dlblc tumor cells in bone marrow smear

Abstract

The application relates to a DLBCL tumor cell intelligent detection method and system in bone marrow smears, which comprises the following steps: acquiring a full-field digital pathology image of a bone marrow smear; performing overlapping block processing to obtain a plurality of sub-image blocks; performing region of interest screening on the sub-image blocks to retain ROI image blocks containing effective cells; performing standardization processing on the ROI image blocks to obtain a target detection input image; detecting the target detection input image by using a deep learning target detection model to output a bounding box target containing DLBCL tumor cells and other cells; performing edge filtering processing on the bounding box target to eliminate the bounding box corresponding to the edge incomplete cells; counting the number of DLBCL tumor cells and other cells; calculating the proportion of DLBCL tumor cells in all nucleated cells to output a detection result. The application realizes automatic identification and quantitative analysis of DLBCL cells, has small calculation overhead, and has a high recognition rate for cross-block cells.

Description

Technical Field

[0001] This invention relates to the field of computer artificial intelligence technology, specifically to an intelligent detection method and system for DLBCL tumor cells in bone marrow smears. Background Technology

[0002] Existing deep learning-based digital pathological image cell detection workflows typically employ standard medical image analysis procedures, including image preprocessing, image segmentation, region filtering, target detection, and result statistics.

[0003] For example, for a whole-slide image (WSI) of bone marrow smears, the high-resolution WSI is first divided into multiple image patches of fixed size. Then, all image patches are directly input into a target detection network for cell detection. The target detection network performs bounding box localization and category prediction for cells in each image patch, outputting the detection results for DLBCL cells and non-DLBCL cells. Finally, based on the detection results of each image patch, the number of target cells in the whole-slide image is accumulated and the proportion of DLBCL cells is calculated. The target detection stage typically uses a standard target detection network structure, such as a single-stage detection algorithm based on a convolutional neural network, to extract features from the input image patches, fuse multi-scale features, and perform target classification regression, outputting the location and category information of the cell targets. However, this approach has the following problems:

[0004] (1) Direct detection of the entire block results in high computational overhead.

[0005] Due to the ultra-high resolution and large size of bone marrow smears (WSI), a single image typically needs to be segmented into numerous image patches. In existing standard workflows, all segmented image patches are directly input into the object detection network for inference computation. Many of these image patches may contain only background areas, staining artifacts, or areas without valid cells. Performing full object detection computation on such low-information image patches results in a large amount of invalid computation, significantly increasing GPU inference time and computational resource consumption.

[0006] (2) Low cross-block cell recognition rate

[0007] Because the slicing process may divide the same cell into different blocks, resulting in only a portion of the cells in each block. Under the existing standard procedure, the detection and classification accuracy of the target cells that have been divided into different blocks is low when input into the detection network, causing the overall results to deviate from the actual results.

[0008] (3) The manual statistical process is inefficient.

[0009] In existing standard procedures, post-processing of test results and calculation of DLBCL ratios may still rely on manual or semi-automatic statistical methods, making it difficult to achieve standardization and efficient calculation.

[0010] Therefore, it is necessary to develop a new intelligent detection method and system for DLBCL tumor cells in bone marrow smears. Summary of the Invention

[0011] The purpose of this invention is to provide an intelligent detection method and system for DLBCL tumor cells in bone marrow smears, which can realize automatic identification and quantitative analysis of DLBCL cells with low computational overhead and high recognition rate for cells across blocks.

[0012] In a first aspect, the present invention provides a method for intelligent detection of DLBCL tumor cells in bone marrow smears, comprising the following steps:

[0013] S1: Acquire full-field digital pathological images of bone marrow smears;

[0014] S2: The full-view digital pathological image is subjected to overlapping and segmentation processing to obtain multiple sub-image blocks;

[0015] S3: Perform region of interest (ROI) filtering on the sub-image blocks, retaining ROI image blocks containing valid cells and removing invalid image blocks;

[0016] S4: Standardize the ROI image block to obtain the target detection input image;

[0017] S5: Use a deep learning target detection model to detect the target input image and output bounding boxes containing DLBCL tumor cells and other cells;

[0018] S6: Perform edge filtering on the bounding box target to remove bounding boxes corresponding to cells with incomplete edges;

[0019] S7: Count the number of DLBCL tumor cells and other cells;

[0020] S8: Calculate the proportion of DLBCL tumor cells among all nucleated cells and output the detection results.

[0021] The above technical solution addresses the problems of existing methods for detecting DLBCL tumor cells in bone marrow smears, which rely on manual slide reading, are highly subjective, inefficient, and lack accuracy. It also resolves the issues of large amounts of unnecessary computation and low cross-block cell recognition rates in existing automated detection processes. This invention achieves automated and standardized detection of DLBCL tumor cells through a complete process including cell image acquisition, overlapping segmentation, ROI screening, standardization, target detection, edge filtering, counting, and proportion calculation. This eliminates the influence of subjective bias in manual slide reading, reduces unnecessary computation, and improves detection efficiency and accuracy. Simultaneously, it provides a quantitative output of the proportion of DLBCL cells, offering reliable quantitative evidence for clinical diagnosis, staging, treatment, and prognostic assessment, thus meeting the clinical need for efficient and standardized auxiliary diagnostic tools.

[0022] In one possible implementation, in S2, the full-view digital pathology image (WSI) is processed by overlapping blocks using a sliding window method. The window size, overlap ratio, and sliding step size are set to be adapted to the cell density characteristics of bone marrow smears. The full-view digital pathology image is traversed through the sliding window to obtain a set of several uniform sub-image blocks with overlapping areas between adjacent blocks.

[0023] In the above technical solutions, the existing technologies for full-view digital pathology image slicing are prone to cell fragmentation and missed detection, and there is no standardized slicing method. This invention addresses these issues by employing a sliding window method, setting a window size, overlap ratio, and sliding step size that are appropriate for the cell density of bone marrow smears. This ensures uniform slicing size, effectively prevents cell fragmentation in overlapping areas of adjacent slices, reduces the risk of missed cell detection across slices, and guarantees the integrity of subsequent cell testing.

[0024] In one possible implementation, the sliding window has a size of 1600×1600px, an overlap ratio of 0.2, and a sliding step size of 1280px.

[0025] In the above technical solution, the specific size of the sliding window (1600×1600px), the overlap ratio (0.2), and the sliding step size (1280px) are further clarified to make the slicing operation more feasible and standardized. The parameter settings are adapted to the characteristics of the ultra-high resolution of bone marrow smear WSI. This can avoid the low efficiency of subsequent processing due to excessively large slicing, and can minimize cell fragmentation by using a reasonable overlap ratio. At the same time, it controls the number of sub-image blocks, balances computational efficiency and detection integrity, and ensures the efficient progress of subsequent ROI screening and target detection.

[0026] In one possible implementation, in step S3, each sub-image block is first normalized and scaled to a preset binary classification model input size, and then the scaled sub-image blocks are input into a ResNet50-based binary classification model for ROI filtering. A classification threshold is defined. When the classification probability value of a sub-image block is greater than or equal to the classification threshold, it is determined to be an ROI image block containing effective cells and is retained to form an ROI image block set. When the classification probability value is less than the classification threshold, it is determined to be an invalid image block without effective cells and is discarded.

[0027] The above technical solution addresses the problems of large amounts of invalid computation, long GPU inference time, and resource waste caused by directly inputting the entire image block into the target detection network in existing technologies. This invention normalizes and scales sub-image blocks to an appropriate size, inputs them into a ResNet50-based binary classification model for ROI selection, and combines this with a classification threshold to quickly remove invalid image blocks such as background blocks and stained artifact blocks that do not contain valid cells, retaining only ROI image blocks containing valid cells. This significantly reduces the computational load of subsequent target detection, improves the inference efficiency of the entire detection process, and simultaneously purifies the data to be detected, reducing the interference of invalid data on the detection results and indirectly improving the detection accuracy.

[0028] One possible implementation is that the standardization process in S4 includes: uniformly scaling the ROI image patches to fit the input size of the deep learning object detection model to form an object detection input set.

[0029] In the above technical solution, to address the issues of size differences, uneven illumination, and insufficient contrast in different ROI image patches, the ROI image patches are uniformly scaled to fit the input size of the target detection model, forming a standardized target detection input set. This eliminates the impact of size differences on detection accuracy. At the same time, it provides input data of uniform specifications for subsequent deep learning target detection models, ensuring the consistency and stability of model inference, avoiding detection deviations caused by inconsistent input specifications, and improving the accuracy of cell recognition and classification.

[0030] In one possible implementation, the deep learning object detection model in S5 is a YOLOv8 object detection model, and each output bounding box includes location coordinates, width and height information, category label and confidence score, wherein the category label is DLBCL or other cell.

[0031] The above technical solution addresses the problems of insufficient specificity, low recognition accuracy, and incomplete output information in existing target detection models for DLBCL cell detection in bone marrow smears. This invention employs the YOLOv8 target detection model, which boasts advantages such as high specificity and high speed. It is well-suited to the characteristics of high heterogeneity and uneven distribution of DLBCL cells in bone marrow smears, enabling accurate identification of DLBCL tumor cells from other cells. The output bounding box includes location coordinates, width and height information, category label, and confidence level. This clearly identifies cell location and category, and the confidence level characterizes the reliability of recognition, providing accurate and complete basic data for subsequent edge filtering and counting, reducing false positives and false negatives.

[0032] One possible implementation is that the edge filtering process in S6 specifically involves: calculating the center coordinates of each bounding box and setting an image edge safety threshold; retaining only bounding boxes whose center coordinates meet the preset safety range and whose cell morphology is complete, and removing bounding boxes whose center coordinates exceed the safety range and whose cell edges are cut by the image, so as to avoid interference from incomplete cells on the statistical results.

[0033] The above technical solutions address the problems of incomplete edge cells, duplicate counting across blocks, and incomplete cells interfering with the accuracy of detection results caused by overlapping blocks. This invention calculates the center coordinates of the bounding box and sets an edge safety threshold, retaining only bounding boxes with complete cell morphology and discarding bounding boxes corresponding to incomplete cells with cut edges. This effectively reduces the interference of incomplete cells on classification and counting, avoids duplicate counting, and improves the accuracy and reliability of cell identification and counting.

[0034] In one possible implementation, in step S7, the number of cells with the category label DLBCL in the filtered bounding box is counted. and the number of cells categorized as "other cell". .

[0035] The above technical solution addresses the problems of low efficiency and susceptibility to counting errors in existing cell counting methods, which rely on manual or semi-automatic processes. This invention automates cell counting by statistically analyzing the number of two cell types (DLBCL and other cells) within the filtered bounding box. This avoids subjective errors and operational mistakes associated with manual counting, ensuring the accuracy and uniqueness of the counting results. It provides precise basic data for subsequent calculations of the DLBCL cell percentage, thus improving the reliability of the detection results.

[0036] In one possible implementation, in S8, the proportion of DLBCL cells to all nucleated cells is calculated:

[0037]

[0038] in: For the final output, This represents the number of cells with the category label DLBCL within the bounding box. This represents the number of cells labeled "other cell".

[0039] The above technical solution addresses the problems of existing technologies where the calculation of the proportion of DLBCL cells relies on manual methods, has low standardization, and is prone to calculation errors. This invention, through a clearly defined proportion calculation formula, automatically calculates the proportion of DLBCL tumor cells to all nucleated cells based on automated statistical analysis of the two cell types, eliminating errors from manual calculations and ensuring the accuracy and standardization of the proportion results. The final output proportion results can be directly used for the quantitative assessment of DLBCL bone marrow infiltration in clinical practice.

[0040] Secondly, the intelligent detection system for DLBCL tumor cells in bone marrow smears according to the present invention includes:

[0041] The image acquisition module is used to acquire and obtain full-field digital pathological images of bone marrow smears;

[0042] The image slicing module is communicatively connected to the image acquisition module and is used to perform overlapping slicing processing on the full-view digital pathology image to obtain multiple sub-image blocks.

[0043] The ROI filtering module is communicatively connected to the image slicing module and is used to filter the regions of interest of the sub-image blocks, retaining ROI image blocks containing valid cells and removing invalid image blocks.

[0044] The image standardization module is communicatively connected to the ROI filtering module and is used to standardize the ROI image blocks to obtain the target detection input image.

[0045] The target detection module is communicatively connected to the image normalization module. It has a pre-trained deep learning target detection model built in, which is used to detect the target input image and output bounding boxes containing DLBCL tumor cells and other cells.

[0046] The bounding box filtering module is communicatively connected to the target detection module and is used to perform edge filtering on the bounding box target, remove the bounding boxes corresponding to cells with incomplete edges, and output the valid bounding boxes.

[0047] The cell counting module is communicatively connected to the bounding box filtering module and is used to extract the category labels corresponding to the effective bounding boxes and count the number of DLBCL tumor cells and other cells besides DLBCL tumor cells, respectively.

[0048] The result calculation and output module is communicatively connected to the cell counting module and is used to calculate the proportion of DLBCL tumor cells to all nucleated cells and output the detection results. Attached Figure Description

[0049] Figure 1 This is a schematic diagram illustrating the principle of the intelligent detection method for DLBCL tumor cells in bone marrow smears in this application embodiment;

[0050] Figure 2 This is a flowchart of the intelligent detection method for DLBCL tumor cells in bone marrow smears in the embodiments of this application;

[0051] Figure 3 This is a network architecture diagram of a binary classification model based on ResNet50 in an embodiment of this application;

[0052] Figure 4 This is an architecture diagram of the YOLOv8 object detection model in the embodiments of this application;

[0053] Figure 5 This is a block diagram of the intelligent detection system for DLBCL tumor cells in bone marrow smears in this application embodiment. Detailed Implementation

[0054] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0055] like Figure 1 and Figure 2 As shown, a smart detection method for DLBCL tumor cells in bone marrow smears includes the following steps:

[0056] S1: Obtain full-field digital pathological images of bone marrow smears.

[0057] S2: Overlapping and slicing the full-view digital pathology image to obtain multiple sub-image blocks.

[0058] S3: Perform region of interest (ROI) filtering on sub-image blocks, retaining ROI image blocks containing valid cells and removing invalid image blocks.

[0059] S4: Standardize the ROI image patch to obtain the target detection input image.

[0060] S5: A deep learning object detection model is used to detect objects in the input image and output bounding boxes containing DLBCL tumor cells and other cells.

[0061] S6: Perform edge filtering on the bounding box targets and remove bounding boxes corresponding to cells with incomplete edges.

[0062] S7: Count the number of DLBCL tumor cells and other cells.

[0063] S8: Calculate the proportion of DLBCL tumor cells among all nucleated cells and output the detection results.

[0064] This application addresses the technical problems of existing methods for detecting DLBCL tumor cells in bone marrow smears, which rely on manual slide reading, are highly subjective, inefficient, and lack accuracy, as well as the large amount of unnecessary calculations and low cross-block cell recognition rate in existing automated detection processes. This application achieves automated and standardized detection of DLBCL tumor cells through a complete workflow including image acquisition, overlapping block segmentation, ROI screening, standardization processing, target detection, edge filtering, counting, and proportion calculation. It eliminates the subjective bias of manual slide reading, reduces unnecessary calculations, and improves detection efficiency and accuracy. Simultaneously, it provides quantitative output of the proportion of DLBCL cells, offering reliable quantitative evidence for clinical diagnosis, staging, treatment, and prognostic assessment, thus meeting the clinical need for efficient and standardized auxiliary diagnostic tools.

[0065] In one possible embodiment, in S2, the sliding window method is used to perform overlapping block processing on the whole field of view digital pathology image (WSI). The window size, overlap ratio and sliding step size are set to adapt to the cell density characteristics of bone marrow smears. The whole field of view digital pathology image is traversed through the sliding window to obtain a set of several uniform sub-image blocks with overlapping areas between adjacent blocks.

[0066] To address the issues of cell fragmentation and missed detection in existing full-view digital pathology image slicing techniques, and the lack of standardized slicing methods, this application employs a sliding window method. This method sets a window size, overlap ratio, and sliding step size that are appropriate for the cell density of bone marrow smears, ensuring uniform slicing specifications. The overlapping areas of adjacent slices effectively prevent cell fragmentation, reducing the risk of missed cross-slice cells and guaranteeing the integrity of subsequent cell testing.

[0067] In one possible embodiment, in S3, each sub-image block is first normalized and scaled to a preset binary classification model input size, and then the scaled sub-image blocks are input into a ResNet50-based binary classification model for ROI filtering; a classification threshold is defined, and when the classification probability value of a sub-image block is greater than or equal to the classification threshold, it is determined to be an ROI image block containing effective cells and is retained to form an ROI image block set; when the classification probability value is less than the classification threshold, it is determined to be an invalid image block without effective cells and is discarded.

[0068] This application addresses the problems of large amounts of unnecessary computation, long GPU inference time, and wasted resources caused by directly inputting the entire image block into the object detection network in existing technologies. This application normalizes and scales sub-image blocks to an appropriate size before inputting them into a ResNet50-based binary classification model for ROI selection. Combined with a classification threshold, it can quickly remove invalid image blocks such as background blocks and stained artifact blocks that do not contain valid cells, retaining only ROI image blocks containing valid cells. This significantly reduces the computational load of subsequent object detection, improves the inference efficiency of the entire detection process, and simultaneously purifies the data to be detected, reducing the interference of invalid data on the detection results and indirectly improving the detection accuracy.

[0069] In one possible embodiment, in S3, each sub-image block is first normalized and scaled to a preset binary classification model input size, and then the scaled sub-image blocks are input into a ResNet50-based binary classification model for ROI filtering; a classification threshold is defined, and when the classification probability value of a sub-image block is greater than or equal to the classification threshold, it is determined to be an ROI image block containing effective cells and is retained to form an ROI image block set; when the classification probability value is less than the classification threshold, it is determined to be an invalid image block without effective cells and is discarded.

[0070] In one possible embodiment, the normalization process in S4 includes: uniformly scaling the ROI image patches to fit the input size of the deep learning object detection model to form an object detection input set.

[0071] This application addresses issues such as size differences, uneven illumination, and insufficient contrast in different ROI image patches. By uniformly scaling ROI image patches to an input size suitable for the object detection model, a standardized object detection input set is formed, eliminating the impact of size differences on detection accuracy. At the same time, it provides input data of uniform specifications for subsequent deep learning object detection models, ensuring the consistency and stability of model inference, avoiding detection deviations caused by inconsistent input specifications, and improving the accuracy of cell recognition and classification.

[0072] In one possible embodiment, the deep learning object detection model in S5 is the YOLOv8 object detection model, and each output bounding box contains location coordinates, width and height information, class label and confidence score, with the class label being DLBCL or othercell.

[0073] This application addresses the shortcomings of existing target detection models in detecting DLBCL cells in bone marrow smears, including insufficient specificity, low recognition accuracy, and incomplete output information. This application employs the YOLOv8 target detection model, which boasts high accuracy and speed, and is well-suited to the heterogeneous and unevenly distributed characteristics of DLBCL cells in bone marrow smears. It can accurately identify DLBCL tumor cells from other cells. The output bounding box includes location coordinates, width and height information, category label, and confidence score, clearly defining cell location and category, and using the confidence score to characterize the reliability of recognition. This provides accurate and complete basic data for subsequent edge filtering and counting, reducing false positives and false negatives.

[0074] In one possible embodiment, the edge filtering process in S6 specifically involves: calculating the center coordinates of each bounding box and setting an image edge safety threshold; retaining only bounding boxes whose center coordinates meet the preset safety range and whose cell morphology is complete, and removing bounding boxes whose center coordinates exceed the safety range and whose cell edges are cut by the image, thereby avoiding interference from incomplete cells on the statistical results.

[0075] This application addresses the problems of incomplete edge cells, duplicate counting across blocks, and incomplete cells interfering with the accuracy of detection results caused by overlapping blocks. By calculating the center coordinates of the bounding box and setting an edge safety threshold, this application retains only bounding boxes with complete cell morphology and removes bounding boxes corresponding to incomplete cells whose edges have been cut. This effectively reduces the interference of incomplete cells on classification and counting, avoids duplicate counting, and improves the accuracy and reliability of cell identification and counting.

[0076] In one possible embodiment, in S7, the number of cells with the category label DLBCL in the filtered bounding box is counted. and the number of cells categorized as "other cell". .

[0077] This application addresses the problems of low efficiency and susceptibility to counting errors in existing cell counting technologies, which rely on manual or semi-automatic methods. By statistically analyzing the number of two cell types (DLBCL and other cells) within the filtered bounding box, this application automates cell counting, avoiding subjective errors and operational mistakes inherent in manual counting. This ensures the accuracy and uniqueness of the counting results, providing precise foundational data for subsequent calculations of DLBCL cell proportions and improving the reliability of the detection results.

[0078] In one possible embodiment, in S8, the proportion of DLBCL cells to all nucleated cells is calculated:

[0079]

[0080] in: For the final output, This represents the number of cells with the category label DLBCL within the bounding box. This represents the number of cells labeled "other cell".

[0081] This application addresses the problems of existing technologies where the calculation of the proportion of DLBCL cells relies on manual methods, has low standardization, and is prone to calculation errors. This application uses a clearly defined proportion calculation formula, based on automated statistical analysis of the number of two cell types, to automatically calculate the proportion of DLBCL tumor cells among all nucleated cells, eliminating errors from manual calculations and ensuring the accuracy and standardization of the proportion results. The final output proportion results can be directly used for the quantitative assessment of DLBCL bone marrow infiltration in clinical practice.

[0082] A preferred embodiment of this application:

[0083] S1: Using the Bionovation Image Cytometry (BionovationBiotech Inc.) cytometry device, all bone marrow smears were scanned at 0.1 μm / pixel (100x oil immersion) to obtain full-field digital pathological images (WSI) and generate complete digitized bone marrow smear images.

[0084] S2: The full-view digital pathology image is divided into image blocks according to a sliding window size of 1600×1600px, an overlap ratio of 0.2, and a sliding step size of stride=1600×(1−0.2)=1280px. The full-view digital pathology image is divided into multiple sub-image blocks, which are called the sub-image block set P: ,in: , This represents the nth sub-image block. Setting an overlap ratio of 0.2 can effectively reduce the risk of missed detection when cells are located at the block boundary.

[0085] S3: For each sub-image patch, first normalize and scale it to... The scaled sub-image patch is then input into a ResNet50-based binary classification model for fast region filtering. The ResNet50-based binary classification model outputs categories defined as Region of Interest (ROI) and Non-ROI (Non-ROI). The classification function is defined as follows: For example: when When, it is determined to be an invalid image block and is directly removed. At that time, ROI image patches that are identified as containing effective cells are retained and stored, forming a set of ROI image patches. ,in For the first This step involves pre-processing image patches representing regions of interest. This process removes a large number of background areas and invalid image patches, reducing the computational burden of subsequent high-complexity detection.

[0086] Classification function The classification function is determined by the relationship between the classification probability and the classification threshold: when the classification probability is greater than or equal to the classification threshold, the classification function is set to 1, and the image is identified as a ROI (Region of Interest); when the classification probability is less than the classification threshold, the classification function is set to 0, and the image is identified as a non-ROI (Region of Interest) image.

[0087] S4: Scale all ROI image blocks uniformly to [size]. To form the target detection input set ,in, This is the Zth region of interest image after scaling to 640×640px.

[0088] S5: Input each ROI image patch into the YOLOv8 object detection model. The YOLOv8 object detection model outputs multiple bounding box object sets. , This represents the j-th bounding box detected by the YOLOv8 object detection model. Each bounding box carries the object's coordinates in the image, its width and height, object category, and detection confidence information. Category labels are defined as: DLBCL and other cell, i.e.: .

[0089] S6: Considering that some cells may be located at the edge of the image patch after overlapping slicing, resulting in incomplete cell morphology and reduced classification accuracy, this application further sets boundary filtering rules to address this issue:

[0090] Calculate each bounding box center coordinates ( , ),in , ,in, The x-coordinate of the center coordinate. The vertical coordinate of the Chinese coordinate system. The coordinates of the top left corner of the bounding box. The width of the bounding box. The height of the bounding box; set the boundary threshold d=75, if any of the following conditions are met: Or Or Or If a cell is found to be located in an image edge region, then that cell is determined to be in such a region. Since cells in this region are often truncated, incomplete, or duplicated across blocks, the bounding box is discarded and not included in subsequent statistics. That is, only cells satisfying the following conditions are retained: The complete cell target, for example, retains Z1 bounding boxes.

[0091] S7: Count the number of DLBCL cells in the filtered test results. Other cell numbers .for example: Z2, It is Z3.

[0092] S8: Calculate the proportion of DLBCL cells among all nucleated cells: R = Z2 / Z1.

[0093] To further illustrate the overall architecture design, resource scheduling, and parallel processing scheme of the pathological WSI image block screening and target detection in this application, the specific implementation method of this application will be described in detail below from four dimensions: data structure design and memory management, computation step reconstruction and scheduling, hardware and software framework collaboration, and parallel distributed expansion.

[0094] A. Data Structure Design and Memory Management Strategies

[0095] The system maintains multiple data structures for process control and status recording: an Activation Cache dictionary with path names as keys and corresponding path data as values; a List to record the names of image chunks that pass ROI screening; and a Checkpoint Stack to record statistics for the current WSI file, including the total number of chunks, the number of ROI image chunks, etc. Number of categories The system includes information on the number of categories, etc.; it also maintains an ActivationCache dictionary, where the key is the WSI chip name and the value is the corresponding processed data.

[0096] In the processing flow, the WSI image is segmented into several sub-image blocks and saved to a specified local path; each sub-image block is input into a ResNet50-based binary classification model for ROI screening. If the screening fails, the memory space corresponding to the sub-image block is immediately released; if the screening passes, the name of the sub-image block is added to a list; after the ROI image blocks that have passed the screening complete YOLOv8 object detection, the memory space they occupy is released.

[0097] B. Restructuring and Scheduling of Computational Steps

[0098] The entire process from WSI segmentation to target detection in the standard workflow is broken down into two independent stages: the input of WSI segmentation results into the ROI screening step, and the input of the screened ROI image blocks into the target detection step.

[0099] This application adopts a design strategy of trading storage space for system traceability and verifiability. That is, by increasing the storage overhead of intermediate results, the processing data and status information of each stage are retained, thereby improving the system's operational stability and the ability to verify the results.

[0100] C. Synergy with computing hardware and frameworks

[0101] After completing the WSI image slicing operation and reading intermediate results on the CPU side, the sub-image blocks to be processed are transferred to the GPU memory for inference calculation using an asynchronous data transfer method.

[0102] By implementing a custom All_step method, core steps such as image segmentation, ResNet inference, and YOLO inference are integrated, and the checkpoint recording and recomputation logic of this application are embedded inside the method.

[0103] D. Parallel and Distributed Scaling

[0104] In the WSI image slicing process, a multi-threaded parallel processing method is used to perform image slicing operations, thereby improving slicing efficiency and system processing speed.

[0105] In this application, besides the 1600×1600 px sliding window and 0.2 overlap ratio in the embodiments, other sliding windows and overlap ratios can also be used for segmentation. The screening model, besides the ResNet50-based binary classification model, can also use other binary classification models for screening. The deep learning object detection model, besides the YOLOv8 object detection model, can use other detection models for detection and classification, but based on experimental data comparison, the YOLOv8 object detection model is better. The boundary threshold, besides 75, can also be dynamically set according to the input size.

[0106] like Figure 3 The diagram shown is a network architecture diagram of a binary classification model based on ResNet50. It undertakes the first-stage rapid ROI screening task in the two-stage detection of this application, and is specifically used to identify regions containing effective cells from bone marrow smear WSI blocks, thereby reducing the computational load of subsequent detection.

[0107] 1. Input layer: Receives sub-image patches scaled to 224×224px, matching the model's standard input size.

[0108] 2. Feature extraction backbone:

[0109] The first layer uses 7×7 convolution (conv) (k=7, s=2, p=3) + max pooling to quickly downsample and extract basic cell texture, staining, and contour features;

[0110] The feature extraction body is composed of four stacked residual modules, which abstract key features such as bone marrow cell morphology, nucleocytoplasmic ratio, and staining uniformity layer by layer, and can stably distinguish cell regions from blank background, impurities, and staining artifacts.

[0111] The residual module mainly uses 3×3 convolution (k=3, s=1, p=1) to extract local cell features finely without changing the feature map size.

[0112] Residual connections avoid gradient vanishing and improve robustness in extracting features of cells with low contrast and morphological heterogeneity.

[0113] 3. Explanation of key convolution parameters

[0114] k (kernel_size): Convolutional kernel size, which controls the receptive field range and feature extraction granularity;

[0115] s (stride): sliding step size, controlling the downsampling rate of the feature map;

[0116] p (padding): Edge padding, which keeps the feature map size stable and preserves cell edge information.

[0117] 4. Classification Output Layer:

[0118] Global average pooling (avgpool) and fully connected layers (FC) output binary classification probabilities, with categories being ROI (containing valid cells) and Non-ROI (invalid regions). Combined with a set threshold, rapid filtering is achieved, retaining only ROI blocks to enter the next stage of target detection, significantly reducing the GPU computational burden.

[0119] Figure 4 This is the architecture of the YOLOv8 target detection model for detecting DLBCL tumor cells in bone marrow smears. It undertakes the high-precision cell detection and classification task in the second phase of this application and is optimized for the pathological characteristics of dense, varied, and morphologically diverse cells in bone marrow smears.

[0120] 1. Input layer: Receives ROI image blocks that have been standardized to 640×640px, providing a uniform input for the model and ensuring consistent cell detection scale.

[0121] 2. Backbone Feature Extraction Network

[0122] The core units are convolutional layers (Conv), C2f modules, and SPPF modules. The initial convolution uses k=3, s=2, and p=1 to complete downsampling and basic feature extraction. The C2f module enhances gradient flow and feature expression, improving the feature extraction capability of small cells and atypical DLBCL cells. The SPPF structure integrates multi-scale receptive fields to adapt to the characteristics of cells of different sizes and disordered distribution in bone marrow smears.

[0123] 3. Standard Definition of Convolution Parameters

[0124] k (kernel_size) is the kernel size, controlling the receptive field of the feature map; s (stride) is the stride, controlling the scaling and downsampling of the feature map; p (padding) is the edge padding, maintaining the stability of the feature map size and preserving key cell edge features; h is the image height, and w is the image width, ensuring that the aspect ratio of the feature map is consistent with the original ROI image; c is the number of channels, corresponding to the feature extraction dimension. These three are combined to adapt the image processing workflow for DLBCL detection in bone marrow smears. c_in is the number of input channels, and c_out is the number of output channels.

[0125] 4. Neck Feature Fusion Structure

[0126] The FPN+PAN multi-scale fusion structure is adopted to integrate deep semantic features with shallow localization features, thereby enhancing the detection capabilities of dense cells and small target cells, and adapting to high cell density scenarios in bone marrow smears.

[0127] 5. Head detection head

[0128] An anchor-free detection mechanism is adopted to simplify the target matching process and improve inference speed and localization accuracy. It outputs three detection scales: P3 (stride=8), P4 (stride=16), and P5 (stride=32), corresponding to small, medium, and large cells, respectively, fully covering all cell morphologies in bone marrow smears. The output includes: cell location coordinates, bounding box width and height, category label (DLBCL / other cell), and detection confidence.

[0129] 6. The loss function and output are jointly optimized using CIoU bounding box loss, BCE classification loss, and DFL distribution focus loss to improve the accuracy of DLBCL cell classification and bounding box regression, providing reliable data support for subsequent edge filtering, cell counting, and proportion calculation.

[0130] 7. Scene adaptability balances detection speed and accuracy, enabling stable identification of DLBCL tumor cells in complex bone marrow smear backgrounds, effectively reducing missed and false detections, and meeting the needs of intelligent clinical pathology detection.

[0131] The following is a description of each abbreviation:

[0132] ROI: Region of Interest, specifically refers to the region containing viable cells in a bone marrow smear image block.

[0133] WSI: Whole Slide Image, which is a complete digital image obtained after scanning a bone marrow smear, and is the initial input for the entire testing process.

[0134] Conv: Convolutional Layer Figure 3 , Figure 4 The core feature extraction unit is used to extract features such as texture, contour, and staining of bone marrow cells.

[0135] Conv2d: A two-dimensional convolutional layer, it is the core module for feature extraction. It is used to perform convolution operations on the input DLBCL cell image to extract the cell's low-level texture, edges, and shallow features, providing initial feature support for subsequent MaxPool, BatchNorm2d processing, and detection head recognition.

[0136] BCE: Binary Cross Entropy Loss. Figure 4 It is used to optimize the binary classification task of DLBCL cells with other nucleated cells.

[0137] CIoU: Complete Intersection over Union Loss. Figure 4 This is used to optimize the accuracy of cell bounding box regression and solve the localization error in dense cell scenes.

[0138] DFL: Distribution Focal Loss Figure 4 This is used to optimize the accuracy of bounding box coordinate prediction and adapt to the strong morphological heterogeneity of DLBCL cells.

[0139] FPN: Feature Pyramid Network Figure 4 The core of the Neck structure is used to fuse multi-scale cellular features and improve small cell detection capabilities.

[0140] PAN: Path Aggregation Network Figure 4 It is a component of the Neck structure and works with FPN to achieve precise fusion of deep and shallow features.

[0141] DLBCL: Diffuse Large B-Cell Lymphoma, the core detection target of this application, namely the type of tumor cells to be identified in bone marrow smears.

[0142] C2f: Cross Stage Partial network with 2 convolutions and fusion, the core unit of the Backbone feature extraction network in the YOLOv8 object detection model. It is used to enhance gradient flow and feature expression. Through dynamic feature aggregation and multi-branch feature reuse, it improves the feature extraction capability of small-sized and irregularly shaped DLBCL cells, while taking into account lightweight and inference efficiency.

[0143] Concat: Concatenate, a technique used to fuse features from different layers, integrate cell-related feature information, and improve detection accuracy.

[0144] Split: This function splits the fused feature map as needed to adapt to different levels of detection requirements and assists in the extraction of cell features.

[0145] Bottleneck: The bottleneck module is used to compress and optimize the input features, filter key features, reduce redundant information, and retain core cell features (such as the morphology and staining features of DLBCL cells), improve the efficiency of model operation, avoid detection errors caused by feature redundancy, adapt to the scenario requirements of bone marrow smear DLBCL detection, and is a key module connecting different feature layers in the YOLOv8 object detection model architecture.

[0146] Upsample: Upsampling is used to enlarge low-resolution feature maps, restore cell detail features, and compensate for image information lost during downsampling. It ensures that the features of DLBCL cells (especially small and atypical cells) are not lost, and that cell morphology can be clearly identified during detection. It is a key step in the YOLOv8 target detection model architecture that connects feature extraction and detection output. It is adapted to the bone marrow smear DLBCL detection scenario to ensure that cell features are not missed and that the location is accurate.

[0147] Backbone: The backbone network is the core component of the YOLOv8 object detection model architecture. It is responsible for extracting cell features (such as cell morphology and staining information) from bone marrow smear images. It is the foundation for subsequent feature fusion, detection and recognition, and supports the normal operation of the entire detection process, ensuring accurate capture of the features of DLBCL cells.

[0148] Head: The detection head is the core of the model's output. It is responsible for converting the extracted cell features into specific detection results (such as cell location and category), directly outputting the detection data of DLBCL cells, and providing support for subsequent counting and analysis.

[0149] Detect: The detection module / detection head is used to receive features extracted by the backbone network, identify DLBCL cells, and output detection data such as cell location and category, providing support for subsequent counting and analysis. It is the core output module of the model.

[0150] Head: The detection head is the core of the model's output. It is responsible for converting the extracted cell features into specific detection results (such as cell location and category), directly outputting the detection data of DLBCL cells, and providing support for subsequent counting and analysis.

[0151] BatchNorm2d: A two-dimensional batch normalization layer used to normalize the DLBCL cell features extracted by convolution, stabilize the model training process, improve feature extraction efficiency, ensure uniform distribution of cell features, and provide stable feature input for the detection head to accurately identify cells.

[0152] Bbox Loss: This is the core loss function for model optimization. It measures the error between the predicted cell bounding boxes and the ground truth labeled boxes in DLBCL. It supervises the model to continuously optimize detection accuracy, ensuring that the output cell location data is accurate and providing a reliable basis for subsequent cell counting and localization analysis.

[0153] CLS Loss: Classification loss is the core loss function for model optimization. It measures the error between the predicted and actual categories of DLBCL cells, supervises the model to accurately identify DLBCL cells, distinguish between normal and diseased cells, and ensures the accuracy of the output cell category data, thus providing support for subsequent cell classification analysis.

[0154] To improve the accuracy, stability, and generalization ability of target recognition, the relevant parameters of the ResNet50 model and the YOLOv8 model were optimized.

[0155] In one possible implementation, the ResNet50 model is used for feature extraction and classification of target images or target regions. The input image size of the ResNet50 model is set to 224*224; the training batch size is set to 16; the number of training epochs is set to 50; the initial learning rate is set to 0.001; the optimizer is SGD; the loss function is BCE; and the number of output classes is set to 2. During training, data augmentation techniques such as random cropping, random flipping, and rotation are also employed to enhance the model's adaptability to images under different lighting, angle, scale variations, and complex background conditions.

[0156] The YOLOv8 model is used for target detection and localization in images. The model version used is yolov8m.pt; the input image size is set to 640*640; the training batch size is set to 16; the number of training epochs is set to 100; the initial learning rate is set to 0.001; the optimizer is SGD; and the number of detection classes is set to 3. During inference, the confidence threshold is set to 0.25, the IoU threshold for non-maximum suppression is set to 0.75, and the maximum number of detected targets is set to 100. These parameter settings effectively filter out low-confidence detection results, reduce duplicate detection boxes, and improve the accuracy of target localization and the reliability of detection results.

[0157] By configuring the parameters of the ResNet50 and YOLOv8 models as described above, the YOLOv8 model can quickly and accurately locate the target region in the image, reducing the interference of the background region on the recognition results. The ResNet50 model can further perform deep feature extraction and fine classification of the target region, thereby improving the accuracy, recall, and stability of the overall recognition process. These parameter configurations can improve the convergence effect during model training, enhance the model's adaptability to complex scenes, changes in lighting, target occlusion, changes in shooting angle, and changes in target scale, thus improving the recognition accuracy and generalization ability of the entire recognition system.

[0158] like Figure 5 As shown in this embodiment, an intelligent detection system for DLBCL tumor cells in bone marrow smears includes an image acquisition module, an image segmentation module, an ROI screening module, an image normalization module, a target detection module, a bounding box filtering module, a cell counting module, and a result calculation and output module. The image acquisition module is used to acquire and obtain full-field digital pathological images of bone marrow smears.

[0159] The image tiling module communicates with the image acquisition module to perform overlapping tiling on the full-view digital pathology image, resulting in multiple sub-image patches. The ROI filtering module communicates with the image tiling module to filter the sub-image patches for regions of interest, retaining ROI image patches containing valid cells and discarding invalid ones. The image normalization module communicates with the ROI filtering module to normalize the ROI image patches, obtaining the target detection input image. The target detection module communicates with the image normalization module, incorporating a pre-trained deep learning target detection model to detect targets in the input image, outputting bounding boxes containing DLBCL tumor cells and other cells. The bounding box filtering module communicates with the target detection module to perform edge filtering on the bounding box targets, discarding bounding boxes corresponding to cells with incomplete edges and outputting valid bounding boxes. The cell counting module communicates with the bounding box filtering module to extract the category labels corresponding to the valid bounding boxes, separately counting the number of DLBCL tumor cells and other cells. The result calculation output module communicates with the cell counting module to calculate the proportion of DLBCL tumor cells among all nucleated cells and output the detection results.

[0160] This application constructs a two-stage cascaded deep learning inference process for ultra-high resolution bone marrow smear (WSI). The innovation of this process lies in replacing large-scale, inefficient object detection computations with low-complexity classification computations, significantly reducing the number of object detection calls from a system execution perspective and improving overall inference efficiency.

[0161] To address the issue of ultra-high resolution in WSI images and the inability to directly perform overall inference, this application employs a fixed-size sliding window segmentation strategy and sets an overlap ratio to ensure that edge cells are covered as completely as possible. In terms of computer system implementation, a key technical aspect of this application is the persistent storage of segmentation results and results from subsequent stages, thereby ensuring traceability of intermediate results, ease of resuming calculations from breakpoints, and convenient clinical review.

[0162] This application introduces a post-processing filtering rule based on the center position of the bounding box after object detection. For each bounding box output by the YOLOv8 object detection model, only targets whose center point coordinates meet the requirements are retained to remove incomplete cells at the edges, reducing duplicate counting caused by overlapping slices and improving classification accuracy.

[0163] This application separately counts the number of DLBCL and other cells and automatically calculates R, thereby eliminating errors from manual statistics.

[0164] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for intelligent detection of DLBCL tumor cells in bone marrow smears, characterized in that, Includes the following steps: S1: Acquire full-field digital pathological images of bone marrow smears; S2: The full-view digital pathological image is subjected to overlapping and segmentation processing to obtain multiple sub-image blocks; S3: Perform region of interest (ROI) filtering on the sub-image blocks, retaining ROI image blocks containing valid cells and removing invalid image blocks; S4: Standardize the ROI image block to obtain the target detection input image; S5: Use a deep learning target detection model to detect the target input image and output bounding boxes containing DLBCL tumor cells and other cells; S6: Perform edge filtering on the bounding box target to remove bounding boxes corresponding to cells with incomplete edges; S7: Count the number of DLBCL tumor cells and other cells; S8: Calculate the proportion of DLBCL tumor cells among all nucleated cells and output the detection results.

2. The intelligent detection method for DLBCL tumor cells in bone marrow smears according to claim 1, characterized in that, In step S2, the sliding window method is used to perform overlapping block processing on the full-view digital pathology image. The window size, overlap ratio and sliding step size are set to match the cell density characteristics of bone marrow smears. The full-view digital pathology image is traversed through the sliding window to obtain a set of several uniform sub-image blocks with overlapping areas between adjacent blocks.

3. The intelligent detection method for DLBCL tumor cells in bone marrow smears according to claim 2, characterized in that, The sliding window has a size of 1600×1600px, an overlap ratio of 0.2, and a sliding step size of 1280px.

4. The intelligent detection method for DLBCL tumor cells in bone marrow smears according to claim 1, characterized in that, In step S3, each sub-image block is first normalized and scaled to the preset input size of the binary classification model, and then the scaled sub-image blocks are input into the ResNet50-based binary classification model for ROI filtering. A classification threshold is defined. When the classification probability value of a sub-image patch is greater than or equal to the classification threshold, it is determined to be a ROI image patch containing effective cells and is retained to form a set of ROI image patches. When the classification probability value is less than the classification threshold, it is determined to be an invalid image patch without effective cells and is removed.

5. The intelligent detection method for DLBCL tumor cells in bone marrow smears according to claim 1, characterized in that, The standardization process in step S4 includes: uniformly scaling the ROI image patches to fit the input size of the deep learning object detection model to form an object detection input set.

6. The intelligent detection method for DLBCL tumor cells in bone marrow smears according to claim 1, characterized in that, The deep learning object detection model in step S5 is the YOLOv8 object detection model. Each output bounding box contains location coordinates, width and height information, category label and confidence score. The category label is DLBCL or other cell.

7. The intelligent detection method for DLBCL tumor cells in bone marrow smears according to claim 6, characterized in that, The edge filtering process in step S6 specifically involves: calculating the center coordinates of each bounding box and setting an image edge safety threshold; retaining only bounding boxes whose center coordinates meet the preset safety range and whose cell morphology is complete, and removing bounding boxes whose center coordinates exceed the safety range and whose cell edges are cut by the image, so as to avoid interference from incomplete cells on the statistical results.

8. The intelligent detection method for DLBCL tumor cells in bone marrow smears according to claim 7, characterized in that, In step S7, the number of cells with the category label DLBCL in the filtered bounding box is counted. and the number of cells with the category label "othercell" .

9. The intelligent detection method for DLBCL tumor cells in bone marrow smears according to claim 8, characterized in that, In step S8, the proportion of DLBCL cells among all nucleated cells is calculated: in: For the final output, This represents the number of cells with the category label DLBCL within the bounding box. This represents the number of cells labeled "other cell".

10. A smart detection system for DLBCL tumor cells in bone marrow smears, characterized in that, include: The image acquisition module is used to acquire and obtain full-field digital pathological images of bone marrow smears; The image slicing module is communicatively connected to the image acquisition module and is used to perform overlapping slicing processing on the full-view digital pathology image to obtain multiple sub-image blocks. The ROI filtering module is communicatively connected to the image slicing module and is used to filter the regions of interest of the sub-image blocks, retaining ROI image blocks containing valid cells and removing invalid image blocks. An image standardization module, which is communicatively connected to the ROI filtering module, is used to standardize the ROI image blocks to obtain the target detection input image. The target detection module is communicatively connected to the image normalization module. It has a pre-trained deep learning target detection model built in, which is used to detect the target input image and output bounding boxes containing DLBCL tumor cells and other cells. The bounding box filtering module is communicatively connected to the target detection module and is used to perform edge filtering on the bounding box target, remove the bounding boxes corresponding to cells with incomplete edges, and output the valid bounding boxes. The cell counting module is communicatively connected to the bounding box filtering module and is used to extract the category labels corresponding to the effective bounding boxes and count the number of DLBCL tumor cells and other cells besides DLBCL tumor cells, respectively. The result calculation and output module is communicatively connected to the cell counting module and is used to calculate the proportion of DLBCL tumor cells to all nucleated cells and output the detection results.

Images