Lesion recognition method and system for non-staining biopsy cells

Abstract

The invention relates to the field of medical image processing, in particular to a lesion recognition method and system for non-staining biopsy cells, and the method comprises the steps: carrying out the imaging collection of the non-staining biopsy cells flowing at a high speed through the time sequence synchronous control of pulse laser and high-speed imaging, and obtaining an original cell image; performing image preprocessing on the original cell image to obtain a preprocessed cell image; performing multi-target cutting on the preprocessed cell image to obtain a standardized single cell image; extracting multi-scale features of the standardized single cell image based on a deep learning model to obtain a lesion recognition result corresponding to the non-staining biopsy cells; and based on a preset structured report generation algorithm and a statistical distribution modeling mechanism, performing structured integration on the lesion recognition result to obtain a structured cell diagnosis report. Through pulse laser imaging and deep learning feature extraction, automatic lesion recognition of non-staining biopsy cells is realized, a report is efficiently generated, and accurate diagnosis is assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and system for identifying lesions in non-stained biopsy cells. Background Art

[0002] In pathological diagnosis, the identification of lesions in biopsy cells is a key link in disease screening and diagnosis. This is especially true for unstained biopsy samples, which do not require chemical staining and can preserve the original morphology of cells, and have important application value in rapid diagnosis scenarios.

[0003] However, existing stain-free biopsy cell identification technology has significant limitations: on the one hand, the contrast between unstained cells and the background is low, and imaging under high-speed flow conditions is prone to produce ghosting, making it difficult to obtain clear images and extract standardized single cells; on the other hand, traditional identification relies on manual reading, which is highly subjective and inefficient. Existing AI-assisted methods also have difficulty in accurately quantifying the degree of cell pathology due to incomplete feature extraction, and the reliability and applicability of the identification results are limited.

[0004] Therefore, how to break through the bottleneck of poor quality of stain-free cell imaging and insufficient recognition accuracy, and build an efficient and accurate stain-free biopsy cell lesion identification solution to improve the objectivity, efficiency and clinical applicability of diagnosis has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] In order to solve the above-mentioned problems in the prior art, the present invention provides a method and system for identifying lesions in non-stained biopsy cells.

[0006] A first aspect of the present invention provides a method for identifying lesions in unstained biopsy cells, comprising: By synchronously controlling the timing of pulsed laser and high-speed imaging, the high-speed flowing unstained biopsy cells are imaged and collected to obtain the original cell images. performing image preprocessing on the original cell image to obtain a preprocessed cell image; Based on a preset target cutting strategy, multi-target cutting is performed on the pre-processed cell image to obtain a standardized single cell image; Extracting multi-scale features of the standardized single-cell image based on a deep learning model to obtain lesion recognition results corresponding to the unstained biopsy cells, wherein the lesion recognition results include benign and malignant judgment results of the unstained biopsy cells, lesion scores, interpretable heat maps of the recognition basis, and feature data supporting offline interactive interpretation; Based on a preset structured report generation algorithm and statistical distribution modeling mechanism, the lesion identification results are structured and integrated to obtain a structured cell diagnosis report containing heat map overlay images.

[0007] A second aspect of the present invention provides a lesion identification system for non-stained biopsy cells, comprising: Ultra-high-speed imaging module, used to image and capture high-speed flowing unstained biopsy cells through synchronous control of pulsed laser and high-speed imaging timing to obtain original cell images; An image preprocessing module, configured to perform image preprocessing on the original cell image to obtain a preprocessed cell image; An image segmentation module, configured to perform multi-target segmentation on the pre-processed cell image based on a preset target segmentation strategy to obtain a standardized single cell image; An intelligent recognition module is configured to extract multi-scale features of the standardized single-cell image based on a deep learning model to obtain lesion recognition results corresponding to the unstained biopsy cells, wherein the lesion recognition results include benign and malignant judgment results of the unstained biopsy cells, lesion scores, interpretable heat maps of the recognition basis, and feature data supporting offline interactive interpretation; The result output module is used to perform structured integration of the lesion identification results based on a preset structured report generation algorithm and statistical distribution modeling mechanism to obtain a structured cell diagnosis report containing a heat map overlay image.

[0008] The beneficial effects of the present invention are reflected in the following aspects: first, through the time-synchronous control of pulsed laser and high-speed imaging and image preprocessing, the technical bottleneck of low contrast of non-stained cell imaging is broken through, and clear imaging and standardized single-cell extraction of high-speed flow cells are achieved, which significantly improves the efficiency of sample processing and the stability of image quality. Secondly, based on the deep learning model, multi-scale features are extracted and lesion identification results are generated, which overcomes the defects of traditional pathological diagnosis that relies on staining and is highly subjective, and realizes accurate quantitative judgment of benign and malignant cells and the degree of lesions, thereby improving the objectivity and accuracy of the identification results. Finally, combined with the structured report generation mechanism, it provides the clinic with an integrated diagnostic output containing key information, while supporting subsequent interactive interpretation, which not only meets the core requirements of pathological diagnosis for efficiency and accuracy, but also enhances the traceability and clinical applicability of the results, greatly improving the diagnostic efficacy and medical collaboration efficiency in the non-stained biopsy scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a schematic flow chart of a method for identifying lesions in non-stained biopsy cells provided by the present invention; Figure 2 This is a schematic structural diagram of a lesion identification system for non-stained biopsy cells provided by the present invention. DETAILED DESCRIPTION

[0010] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0011] The existing technologies and their main problems can be summarized as follows: 1. High-throughput bottlenecks in image acquisition and processing: The image acquisition and processing mechanisms of existing cell imaging systems (such as TCT and imaging flow cytometers) are limited in efficiency. To obtain high-definition images, traditional systems often require extended exposure times or multiple scans, increasing the processing time of a single sample. Furthermore, data processing capabilities for large sample sizes have not been improved simultaneously, creating a throughput bottleneck in the "imaging-processing" chain. This makes it difficult to adapt to the needs of real-time, rapid, large-scale screening, and this efficiency shortfall is particularly prominent in grassroots or mobile testing scenarios.

[0012] 2. Insufficient recognition accuracy of low-contrast / unstained cells: Existing cell recognition technologies (such as traditional edge detection, threshold segmentation algorithms, and some deep learning models) have inherent flaws when processing transparent or unstained cell samples. Due to the low contrast between cells and background and blurred boundaries in these samples, traditional algorithms are susceptible to background noise interference, resulting in incomplete extraction of the target region. Even with morphological optimization, it is still difficult to accurately distinguish cells from the background, which can easily lead to missed detection of abnormal cells. Some deep learning models also suffer from low recognition accuracy due to the insufficient proportion of low-contrast samples in the training data.

[0013] 3. Complexity of sample preprocessing and deployment limitations: Existing mainstream technologies (such as TCT) rely on multi-step pretreatment mechanisms, including centrifugation, chemical staining, and slide preparation and fixation. These processes are not only cumbersome but also rely on specialized consumables, centrifugal equipment, and specialized operators. While some automated systems simplify operations, they still have strict requirements for operating environments such as temperature and cleanliness. This makes deployment difficult in resource-limited settings such as primary healthcare institutions and mobile testing vehicles, limiting the technology's universal applicability.

[0014] 4. Lack of explainability of decisions in intelligent recognition systems: Existing cell recognition technologies based on deep learning (such as CNNs) operate in a "black box" fashion, with their core feature extraction and classification processes unable to provide clear decision-making evidence (e.g., which cell morphological features support a "abnormal" diagnosis). This makes it difficult for medical professionals to verify the validity of the results, leading to reduced trust in the systems and a direct impact on the actual adoption of these technologies in clinical diagnosis.

[0015] Example 1: Reference Manual Figure 1 , which shows a flow chart of a method for identifying lesions in non-stained biopsy cells provided by the present invention.

[0016] An embodiment of the present invention provides a method for identifying lesions in non-stained biopsy cells, comprising: S1: Through the timing synchronization control of pulsed laser and high-speed imaging, the high-speed flowing unstained biopsy cells are imaged and collected to obtain the original cell image.

[0017] This step addresses the problems of blurred imaging, uneven illumination, motion smear, and insufficient frame rate in the existing cell image acquisition process. By combining the particle image formation mechanism under high-speed flow conditions, an image acquisition system is constructed that integrates optics, fluid control, imaging synchronization, and parameter control mechanisms.

[0018] Specifically, a pulsed laser is used as the illumination source. A spatial light modulation system shapes the laser beam into a near-circular spot with a Gaussian energy distribution. This concentrates the illumination intensity at the cell edge, enhancing the discernibility of cell boundaries. After beam expansion, collimation, and uniformity, the spot is projected perpendicularly onto the imaging acquisition window of the microfluidic channel.

[0019] To ensure uniform illumination within the imaging acquisition window, a multi-order aspheric lens group is introduced into the optical path design to control the radial diffusion of illumination energy and achieve high uniformity coverage.

[0020] Given that cells flow at a constant velocity in microfluidic channels, conventional continuous illumination without illumination modulation would produce image smearing. To address this, the present invention designed a pulse-width modulated laser drive mechanism that matches the flow rate. This allows the light source to flash only at the instant of image acquisition, thereby freezing the cell motion trajectory and improving transient image clarity.

[0021] In one possible embodiment, the process of S1 is to synchronously control the timing of pulsed laser and high-speed imaging to image and collect high-speed flowing unstained biopsy cells to obtain the original cell image, including: Equipped with a pulsed laser light source and high-speed imaging equipment, the unstained biopsy cells are introduced into the microfluidic channel and the flow velocity of the unstained biopsy cells is controlled to obtain a flow cell sample to be imaged; The delay time is calculated through the synchronous trigger mechanism to control the timing synchronization of laser pulse emission and image acquisition, and the flow cell sample entering the imaging acquisition window is photographed to obtain the initial cell image; When the signal-to-noise ratio of the initial cell image is lower than a preset threshold, the laser pulse energy or the imaging device gain parameter is dynamically adjusted to obtain an optimized cell image; Cells are continuously photographed based on optimized parameters to ultimately obtain high-contrast original cell images without smear.

[0022] Specifically, during the system configuration and sample preparation phase, a pulsed laser light source (such as a semiconductor laser) can be set as the illumination source, and a high-speed imaging device (such as a high-frame-rate CMOS camera) can be configured. The laser light source must support pulse modulation, and its output beam must be spatially shaped, expanded, and homogenized to form a light field with uniform energy distribution, which is projected vertically onto the imaging acquisition window of the microfluidic channel. An unstained biopsy cell sample (such as cervical exfoliated cells) is then introduced into the microfluidic channel, and the cell flow speed is adjusted to the target value (for example, above 5 m / s) through a precision fluid control system to form a high-speed, stable flow of cell samples to be imaged. During the timing synchronization and initial imaging phase, a synchronous triggering mechanism for the pulsed laser and the high-speed imaging device is established to capture the flow cell sample entering the imaging acquisition window and obtain the initial cell image.

[0023] In one possible embodiment, a delay time is calculated by a synchronous trigger mechanism to control the timing synchronization of laser pulse emission and image acquisition, and a flow cell sample entering the imaging acquisition window is photographed to obtain an initial cell image, specifically including: The synchronization trigger mechanism calculates the delay time according to the following formula: in, t d is the delay time required for unstained biopsy cells to enter the sensing area and the imaging acquisition window. L is the distance from the sensing area to the imaging acquisition window, v f The real-time flow rate of unstained biopsy cells; Control the single-frame exposure time of high-speed imaging equipment to meet the following anti-smear conditions: in, t e is the single-frame exposure time of the high-speed imaging device, is the pixel pitch of the image sensor; At the moment when the unstained biopsy cells enter the sensing area, a delay time is added, and this moment is used as the trigger point to synchronously trigger the laser pulse emission and the high-speed imaging device exposure to capture the unstained biopsy cells and obtain a high-contrast initial cell image without smear.

[0024] For example, in the anti-smear condition above, the left side of the inequality represents the actual displacement of cells in the field of view during a single frame exposure, while the right side represents the maximum displacement allowed in the image, which is one pixel wide. This constraint defines the minimum frame rate or maximum exposure time required by the imaging system at the target flow rate, ensuring that cell boundaries are not overlapped or blurred in the image.

[0025] To achieve precise imaging timing control, the system introduces an FPGA-based synchronous trigger mechanism, establishing a time logic coupling between the cell passing through the sensing area and the camera acquisition window. The cell passing signal is obtained through the front optical sensor or fluid disturbance detection device. The FPGA generates precise delay timing based on the signal and drives the laser pulse and image acquisition simultaneously. Its core scheduling mechanism is based on the above time control relationship , through precise control , in order to lock the imaging timing of a single cell, and further enable image acquisition to start just when the cell completely enters the imaging acquisition window, thereby improving the cell centering rate and imaging efficiency in the image.

[0026] Considering the large differences in image contrast and brightness under different sample types and background conditions, an image quality feedback control mechanism is also introduced in the high-speed acquisition process. Specifically, by calculating the grayscale mean in the image frame and variance , real-time evaluation of frame-level signal-to-noise ratio (SNR): This metric is used as a quantitative standard for image clarity and segmentability, combined with a pre-set minimum acceptable threshold Once detected , automatically adjusts the laser pulse energy or camera analog gain factor to ensure that the acquired image remains within the algorithm's processing range. This mechanism effectively overcomes brightness fluctuations caused by varying cell concentrations, refractive indices, or background light interference, improving the consistency of imaging quality across the entire device.

[0027] This step also addresses optical path drift and thermal stability during continuous acquisition. By introducing a thermal compensation structure, an optical axis alignment locker, and an imaging surface reflection calibration system, fine-tuning and closed-loop correction of the laser illumination path and imaging surface focal length are achieved. This ensures a stable positional relationship between the laser illumination and the imaging acquisition window even after long-term operation, thus preventing image drift and loss of focus caused by optical axis offset.

[0028] After image acquisition is completed, the image data will be transmitted in real time in a streaming form for subsequent image preprocessing and image segmentation steps, providing the raw data basis with stable edge features and high contrast for subsequent cell identification.

[0029] The S1 utilizes laser pulse modulation and timing synchronization with a high-speed camera, combined with flow velocity-frame rate coupling constraints and real-time fluid detection feedback mechanisms to strictly control single-frame exposure time and cell movement displacement within the pixel level, achieving high-definition, smear-free imaging of unstained cells in high-speed flow. This solves the image blur, motion smear, and insufficient frame rate issues often encountered with traditional imaging technologies in high-speed flow samples, particularly improving imaging distortion of unstained, transparent cells. This provides high-fidelity data for subsequent identification and meets the needs of large-scale, rapid screening.

[0030] S2: performing image preprocessing on the original cell image to obtain a preprocessed cell image.

[0031] In one possible implementation, performing image preprocessing on the original cell image to obtain a preprocessed cell image specifically includes: Performing Gaussian filtering on the original cell image to suppress image noise in the original cell image and obtain a smooth image; Performing local contrast equalization on the smoothed image, and enhancing the distinction between cells and background in the smoothed image by grayscale stretching to obtain an enhanced image; Perform phase consistency analysis and gradient calculation on the enhanced image to extract the cell edge features corresponding to the low-contrast area in the enhanced image to form an edge image; Binarization and morphological operations are performed on the edge image to remove noise and connection broken boundaries in the edge image and obtain the connected domain boundary image; The connected domain boundary image is used as the preprocessed cell image.

[0032] It is worth noting that the task of image preprocessing is to perform noise suppression, contrast enhancement, and edge enhancement on high-speed cell images without relying on staining, thereby extracting image regions with clear boundaries and ensuring the accuracy and standardization of subsequent recognition model input. This step is suitable for cell samples with characteristics such as high transparency, large morphological differences, uneven cell density, and complex backgrounds. At the same time, it must meet the requirements of high-throughput processing and ensure that image structural information is not lost or deformed while maintaining processing speed.

[0033] Specifically, in order to address the high-frequency electronic noise and low-light disturbance that may exist in the imaging process, a joint image preprocessing method based on Gaussian filtering and local contrast equalization is first adopted. , its Gaussian filtering operation in the spatial domain can be expressed as: in, represents the smoothed image after smoothing, is a two-dimensional Gaussian kernel function, defined as: in, is the standard deviation of the Gaussian kernel, which controls the filtering strength.

[0034] This filtering process can effectively remove random noise, but it will weaken the edge information. Therefore, a local contrast enhancement operator needs to be introduced after filtering.

[0035] For each pixel , calculate its The maximum grayscale value in the neighborhood , minimum grayscale value , and then perform local enhancement through linear stretching transformation: Get the enhanced image after enhancement , its edge gradient is more obvious, providing a basis for subsequent edge extraction.

[0036] In terms of edge extraction, considering the problems of weak edge response and poor robustness of traditional edge detection methods such as Sobel and Canny in transparent cells or low-contrast images, the present invention introduces a composite edge extraction algorithm based on image gradient field convolution and phase consistency analysis.

[0037] First, calculate the gradient image through the first-order derivative , and extract the main edge direction through the directional high-pass filter. On this basis, the phase consistency function is introduced, and its expression is: in, For the The amplitude under the frequency channel, is the phase, is the average phase, is the noise threshold, For small amounts, prevent the denominator from being zero.

[0038] The phase consistency function can reflect the consistency of the frequency domain structure changes in the image and can detect boundary information in areas with weak intensity changes. It is suitable for edge extraction of images with insufficient staining or transparent cells.

[0039] After obtaining the edge image, the system binarizes it and removes isolated noise points and connects broken boundaries through morphological operations (opening and closing operations), thereby obtaining candidate cell areas with strong connectivity and clear boundaries.

[0040] S2 extracts stable structural information from weak grayscale regions through phase congruency analysis and combines it with the gradient operator results. Combined with a multi-target selection mechanism using non-maximum suppression, it achieves precise extraction of low-contrast, transparent cell edges. This solves the problem of traditional edge detection's poor recognition of unstained and adherent cells, effectively stripping away overlapping regions and providing clear edge features for subsequent segmentation. This improves the robustness of image preprocessing and ensures the quality and stability of the input data.

[0041] S3: Based on the preset target cutting strategy, multi-target cutting is performed on the preprocessed cell image to obtain a standardized single-cell image.

[0042] In one possible embodiment, based on a preset target segmentation strategy, multi-target segmentation is performed on the preprocessed cell image to obtain a standardized single cell image, specifically including: Fitting the minimum bounding rectangle to each connected area in the preprocessed cell image to obtain the first candidate box set; Based on a preset contrast scoring function, calculating the contrast score of the image area corresponding to each candidate frame in the first candidate frame set to obtain a contrast score set; Based on the contrast score set, the candidate boxes in the first candidate box set are sorted in descending order to obtain a second candidate box set; Traverse each candidate box in the second candidate box set, calculate the intersection-and-union ratio of each candidate box with the remaining candidate boxes, and obtain an intersection-and-union ratio set; For each candidate frame in the second candidate frame set, if its intersection-over-union ratio with other candidate frames is greater than a preset threshold, only the candidate frames with higher contrast scores are retained, and the remaining candidate frames are suppressed to obtain the target candidate frame set; The single-cell area corresponding to the target candidate frame set is size-normalized to obtain a standardized single-cell image.

[0043] It's worth noting that, given that cervical cells often adhere, overlap, and vary significantly in size, traditional connected domain-based object extraction methods are prone to missegmentation or misidentification of adhered regions. To address this, the present invention designs a target segmentation strategy based on a combination of minimum enclosing rectangles and non-maximum suppression.

[0044] For each connected area, first fit its minimum enclosing rectangle to obtain a set of candidate boxes For all According to the contrast scoring function within the image region Sort in descending order, then traverse each box and calculate the intersection over union (IoU) with the rest of the boxes, defined as follows: like If the score is greater than a preset threshold, the candidate boxes with higher scores are retained, suppressing redundant annotations in overlapping areas. This method ensures that only the most representative target boundaries are retained in densely populated or overlapping areas, improving the discriminability and uniqueness of the cuts.

[0045] The final output single-cell image blocks are uniformly adjusted to a fixed-size standard input format, that is, size standardization is performed to obtain standardized single-cell images, and they enter the next link together with the original cell image position information for subsequent multi-classification diagnosis and lesion feature determination.

[0046] The above-mentioned image preprocessing and target cutting process is characterized by automation, adaptability and parallelism. It can maintain stable image quality control capabilities and target cutting consistency under high-throughput input streams, laying a stable foundation for back-end intelligent analysis and clinical interpretation.

[0047] Based on a preset segmentation strategy, S3 performs multi-target segmentation on preprocessed cell images through bounding rectangle fitting and intersection-over-union constraints, effectively separating adherent cells and extracting standardized single-cell images. This solves the problems of failed cell target extraction or overlapping recognition in traditional segmentation algorithms, ensuring that the output single-cell images have uniform size and unique boundaries, providing standardized input for subsequent intelligent recognition and improving the accuracy and stability of subsequent recognition.

[0048] S4: Based on the deep learning model, multi-scale features of standardized single-cell images are extracted to obtain lesion recognition results corresponding to unstained biopsy cells. The lesion recognition results include the benign and malignant judgment results of unstained biopsy cells, lesion scores, interpretable heat maps of the recognition basis, and feature data that support offline interactive interpretation.

[0049] It is worth noting that this step is the core step of the present invention for realizing cell classification, lesion determination and subtype labeling. It undertakes the task of extracting semantic features from image structure information and making medical significance judgments. It is the key link for the present invention to realize early diagnosis and automatic decision-making assistance.

[0050] Specifically, given that cervical exfoliated cells exhibit weak detail, complex texture distribution, and strong individual variability when unstained or weakly stained, traditional shallow classifiers (such as support vector machines or K-nearest neighbor algorithms) perform poorly in extracting high-dimensional features and identifying heterogeneous samples. Therefore, this step constructs a deep neural architecture based on the fusion of a multi-branch convolutional network and an attention mechanism. This architecture can simultaneously extract key information such as cell boundary texture, nuclear-cytoplasmic ratio, and nuclear atypia at different scales. It also utilizes an attention weighting mechanism to dynamically adjust the focus on local regions, thereby enhancing sensitivity to subtle features of early lesions.

[0051] In one possible implementation, multi-scale features of standardized single-cell images are extracted based on a deep learning model to obtain lesion recognition results corresponding to unstained biopsy cells, specifically including: The standardized single-cell image is input into a multi-branch convolutional network, and the cell features in the standardized single-cell image are extracted in parallel through convolution kernels of different scales to obtain a multi-branch feature map; The channel weights of the multi-branch feature maps are calculated through the attention mechanism and weighted fusion is performed to obtain the fused feature map; Through the classification branch, the fusion feature map is used to determine the benign and malignant cells and lesion subtypes, and the target category is output; Through the regression branch, the fusion feature map is scored for the degree of cell abnormality to obtain a score value; the classification branch and the regression branch are run in parallel; Based on the target category and fused feature map, an interpretable heat map is generated in real time through the gradient-weighted class activation mapping algorithm; Data integration of target categories, scoring values, and interpretability heatmaps was performed to obtain lesion recognition results corresponding to unstained biopsy cells.

[0052] Specifically, the input of the neural network is a set of standardized cell images ,in h is the image height, w is the image width, is the number of channels.

[0053] The neural network structure consists of three parts: The first part is the feature extraction backbone network, which uses multi-scale convolution kernels to build multiple parallel convolution paths and define the The output of the convolution is: in, For the The convolution kernel of the path, represents the convolution operation, is the bias term, is the rectified linear activation function.

[0054] The feature maps of all paths are combined and then weighted fused through the channel attention module. This step is done by Calculate its global average pooling value , and then introduce the learnable weight vector , and get the weighted output: in, is the final fusion feature map, For the The importance weights of each channel can be learned through a fully connected layer and Softmax normalization. This mechanism can significantly enhance the response to abnormal areas or weak structures, effectively improving the discrimination ability of the present invention in low-contrast images.

[0055] After the feature fusion is completed, it is sent to the parallel classification branch and regression branch.

[0056] The classification branch is used to determine whether cells are benign or malignant and output lesion subtype labels. It uses the cross-entropy loss function for supervised learning and is defined as: in, is the one-hot encoding of the actual category, is the predicted category probability, is the total number of categories, k It is the category number.

[0057] The regression branch is used to determine the degree of cell abnormality or morphological score, using mean square error loss: in, is the actual score of the sample, is the model prediction value, is the number of samples.

[0058] In one possible implementation, based on the target category and the fused feature map, an interpretable heat map is generated in real time using a gradient-weighted class activation mapping algorithm, specifically including: Get the target category score corresponding to the target category; Calculate the gradient matrix of the target category score to the fusion feature map; Calculate the global mean of the gradient matrix corresponding to each channel in the fusion feature map to obtain the channel weight data; The channel weight data is weighted and summed with the fusion feature map to generate an interpretable heat map.

[0059] In order to solve the problem of insufficient credibility caused by the "black box" nature of AI diagnosis in clinical applications, the present invention integrates an explainability enhancement mechanism at the inference end.

[0060] Specifically, a visualization method based on gradient-weighted class activation mapping is introduced to calculate the contribution of each spatial position to the final classification result by backpropagating the gradient of the model output relative to the intermediate convolutional features.

[0061] For the feature maps, whose weights are defined by the following expression: in, represents the target category score, For the Channels at position ( i , j ), For the The weight of each channel, is the normalization factor.

[0062] The final interpretability heatmap is obtained by the following weighted combination: This heat map is superimposed on the original cell image to visualize the basis for classification decisions, allowing users to intuitively understand the logical correspondence between the model's focus areas and the diagnostic results, thereby enhancing clinical adoption. The user in this disclosure can be a doctor or related personnel, and this is not limited to this.

[0063] S4 uses a deep model that integrates multi-path convolution and channel attention mechanisms to extract cell features at different scales in parallel. Combined with class activation mapping, it generates interpretable heatmaps, enabling accurate judgment and visualization of benign and malignant cell types and lesion severity. This addresses the challenges of traditional CNN models in extracting subtle lesion features and lacking interpretability in their decision-making processes. It enhances the model's ability to identify and generalize early-stage lesions, while also increasing the clinical adoption of lesion identification results through transparency.

[0064] S5: Based on the preset structured report generation algorithm and statistical distribution modeling mechanism, the lesion identification results are structured and integrated to obtain a structured cell diagnosis report with heat map overlay images.

[0065] This step is the information carrier and feedback hub for end users in the technical solution of the present invention. It is responsible for presenting the lesion identification results in a clinically acceptable, explainable and interactive manner, and supports docking with the hospital information system (HIS) or electronic medical record system (EMR) to achieve closed-loop management and traceability analysis of diagnostic data.

[0066] Specifically, three types of data are first received: the first is the benign and malignant classification results of cell images; the second is the probability prediction value and corresponding score of each lesion type; and the third is the interpretable feature heat map generated based on gradient backpropagation.

[0067] In order to ensure consistency and structured information communication, the system constructs a multi-channel result fusion model and uses a structured report generation algorithm to standardize and integrate heterogeneous output data.

[0068] Assume that the total number of cell samples is , No. The classification results of samples are , the lesion score is , the feature attention area heat map is , then the output record of each cell can be represented as a triple: all are collected and form the sample aggregate output matrix , and then index mapping is performed based on the patient information of the cell source, sample timestamp and test batch number to establish the data structure entries at the diagnosis level.

[0069] In order to enhance the hierarchical interpretability of the diagnostic results, the system introduces a statistical distribution modeling mechanism to aggregate the scores. Standardized analysis is performed under the assumption of normal distribution. The sample mean score is defined as , the variance is , then each rating can be mapped to a standard score: according to Cells are classified into three categories based on their position on a standard normal distribution: normal, mildly suspicious, and highly abnormal. These are then graphically represented using different color codes. This mechanism allows doctors to intuitively identify potentially high-risk samples based on the score distribution, without having to understand the internal mechanisms of the deep network.

[0070] In terms of heat map presentation, the present invention presents each cell image with its corresponding focus heat map. Perform channel superposition to generate a color artifact image , where the red channel value of each pixel is defined by the following formula: in, is the red channel of the original image, is the heatmap overlay weight coefficient. This image is used to assist doctors or other users in identifying key areas of interest for the model and verifying lesion identification results. Specifically, for example, a user can click on a specific area to access the corresponding raw cell image, characteristic distribution curve, and interpretation logic, thereby establishing a closed loop of human-computer interactive interpretation.

[0071] To enhance compatibility with clinical systems, the result output module corresponding to step S5 also has a built-in standardized interface that supports the HL7 and FHIR protocols. This unified data structure connects to the hospital's internal diagnostic record management system, enabling archiving, retrieval, and sharing of interpretation results. Each report comes with a data summary, AI score distribution charts, key image annotation pages, and interpretable documentation, making the diagnostic report both technically sound and medically readable.

[0072] In addition, in order to evaluate the credibility of the model output, the result output module introduces an output confidence evaluation mechanism, which realizes the result quality control by calculating the prediction consistency index of the model on multiple interpretation paths. , define the consistency measurement function as: in, Representation Label The number of occurrences, is the total number of paths. This consistency value is attached to the result as a credibility label for doctors to refer to during the interpretation process.

[0073] S5 utilizes a structured report generation algorithm and statistical distribution modeling mechanism to integrate lesion identification results with heatmap overlays, creating a three-in-one output structure of images, scores, and heatmaps, supporting user interactive verification and feedback. This addresses the issues of traditional systems with single output formats and a lack of transparent explanations, achieving a precise correspondence between diagnostic results, the original cell image regions, and the model response basis, improving the visibility and operability of the results. It also supports human-computer interactive interpretation, enhancing clinical practicality.

[0074] In one possible embodiment, before performing structured integration of lesion identification results based on a preset structured report generation algorithm and statistical distribution modeling mechanism to obtain a structured cell diagnosis report containing a heat map overlay image, the method further includes: In response to the user's questioning of the lesion identification results, a preset offline interpretation mechanism is triggered; Through an offline interpretation mechanism, key features that affect lesion recognition results are extracted based on standardized single-cell images, the corresponding fusion feature maps of standardized single-cell images, and the model decision logic of the deep learning model; Generate visual explanation content for questioned instructions based on key features; The visual explanation content is displayed in association with the lesion identification results, so that users can interactively verify the visual explanation content and the identification results; Receive user feedback on the visual interpretation content, associate the feedback with the lesion identification results, and form a closed-loop interaction record.

[0075] Specifically, when a user questions the lesion identification results, a preset offline interpretation mechanism is triggered. In this invention, users can click on a specific area in the structured cytology diagnostic report to access the corresponding raw cell image, characteristic distribution curve, and interpretation logic. The interpretation logic can include the cell morphology parameters that the model focuses on (such as nuclear-cytoplasmic ratio and edge smoothness), as well as the degree to which each parameter influences the lesion identification results.

[0076] In one possible implementation, the offline interpretation mechanism includes a LIME offline interpretation mechanism and a SHAP offline interpretation mechanism. In response to a user's questioning instruction regarding the lesion identification result, the preset offline interpretation mechanism is triggered, including: In response to a user's questioning instruction on the lesion identification result, at least one of the LIME offline interpretation mechanism and the SHAP offline interpretation mechanism is triggered.

[0077] Specifically, if the LIME offline interpretation mechanism is triggered: based on the standardized single-cell image, a sample set is generated by local area perturbation, and combined with the sensitivity rules to image features in the model decision logic, a local linear substitution model is trained to simulate the decision trend of the original model; the superpixel area (i.e., key features) that has a significant impact on the lesion recognition results is screened out through model weight sorting, and it is superimposed with the original cell image to generate a visual explanation map, which serves as a visual explanation content to intuitively display the relationship between the local features of the image and the recognition results.

[0078] If the SHAP offline interpretation mechanism is triggered, the cell morphological parameters that the model decision logic focuses on (such as nuclear-cytoplasmic ratio, edge smoothness, etc.) are extracted as key features from the fusion feature map corresponding to the standardized single-cell image, and the Shapley value of each parameter is calculated based on the feature interaction rules to quantify its marginal contribution to the recognition result; a waterfall chart is generated by sorting the contribution, and the influence relationship between the features and the recognition results is displayed by distinguishing between positive and negative contributions, which serves as a visual explanation content to clearly present the decision basis at the feature level.

[0079] It is worth noting that both the questioning instruction and the feedback instruction are issued through a preset interactive method. The feedback instruction can be issued through any one or more of the following interactive methods, which are not limited in the present invention: 1. Result confirmation / correction: Click "Approve" or "Disapprove" to correct the score and mark the reason for the disapproval.

[0080] 2. Annotation: Mark missed features on the image and add text annotations (such as "Nuclear morphology here is normal").

[0081] 3. Quantitative rating: Provide feedback on the effectiveness of explanations through star ratings or scales (e.g., “Explanation clarity: high / medium / low”).

[0082] The query instruction may be issued through any one or more of the following interactive methods, which are not limited in the present invention: 1. Graphic click: Click on key areas such as heat maps, scores, or preset buttons (such as "Questionable results") to trigger targeted questions.

[0083] 2. Option selection: Use the drop-down menu or checkbox to select preset query types such as "Misclassification" and "Insufficient explanation".

[0084] 3. Text input: Enter specific questions in the text box (e.g., “Why is this cell judged to be highly abnormal?”).

[0085] 4. Voice assistance: Voice commands (such as "question the third cell score") are adapted to multiple operation scenarios.

[0086] This step uses the LIME or SHAP offline interpretation mechanism, based on standardized single-cell images, fused feature maps, and model decision logic, to extract key features that influence recognition results (such as nuclear morphology and edge features) and generate explanatory information, providing targeted explanations for questioned instructions. This resolves the "black box" problem of intelligent recognition results, allowing users to understand the recognition basis through interactive verification, enhancing the credibility and traceability of the results and addressing the lack of interpretability of traditional AI model decisions.

[0087] Example 2 Reference Manual Figure 2 , shows a structural schematic diagram of a lesion identification system for non-stained biopsy cells provided by the present invention.

[0088] An embodiment of the present invention provides a lesion identification system 20 for non-stained biopsy cells, comprising: The ultra-high-speed imaging module 201 is used to image and capture high-speed flowing unstained biopsy cells through synchronous control of pulsed laser and high-speed imaging timing to obtain original cell images; An image preprocessing module 202 is configured to perform image preprocessing on the original cell image to obtain a preprocessed cell image; An image segmentation module 203 is configured to perform multi-target segmentation on the pre-processed cell image based on a preset target segmentation strategy to obtain a standardized single cell image; an intelligent recognition module 204 for extracting multi-scale features of the standardized single-cell image based on a deep learning model to obtain lesion recognition results corresponding to the unstained biopsy cells, wherein the lesion recognition results include benign and malignant judgment results of the unstained biopsy cells, lesion scores, an interpretable heat map of the recognition basis, and feature data supporting offline interactive interpretation; The result output module 205 is used to perform structured integration on the lesion recognition results based on a preset structured report generation algorithm and statistical distribution modeling mechanism to obtain a structured cell diagnosis report containing a heat map overlay image.

[0089] In a possible implementation, the ultra-high-speed imaging module 201 is specifically configured to: A pulsed laser light source and a high-speed imaging device are configured to introduce unstained biopsy cells into a microfluidic channel and control the flow velocity of the unstained biopsy cells to obtain a flow cell sample to be imaged; Calculating the delay time through a synchronous trigger mechanism to control the timing synchronization of laser pulse emission and image acquisition, photographing the flow cell sample entering the imaging acquisition window, and obtaining an initial cell image; When the signal-to-noise ratio of the initial cell image is lower than a preset threshold, the laser pulse energy or the imaging device gain parameter is dynamically adjusted to obtain an optimized cell image; Cells are continuously photographed based on optimized parameters to ultimately obtain high-contrast original cell images without smear.

[0090] In a possible implementation, the ultra-high-speed imaging module 201 is specifically configured to: The synchronization trigger mechanism calculates the delay time according to the following formula: in, t d is the delay time required for unstained biopsy cells to enter the sensing area and the imaging acquisition window. L is the distance from the sensing area to the imaging acquisition window, v f The real-time flow rate of unstained biopsy cells; The single-frame exposure time of the high-speed imaging device is controlled to meet the following anti-smear conditions: in, t e is the single-frame exposure time of the high-speed imaging device, is the pixel pitch of the image sensor; At the moment when the unstained biopsy cells enter the sensing area, the delay time is added, and this moment is used as the trigger point to synchronously trigger the laser pulse emission and the high-speed imaging device exposure to photograph the unstained biopsy cells and obtain a high-contrast initial cell image without smear.

[0091] In a possible implementation, the image preprocessing module 202 is specifically configured to: performing Gaussian filtering on the original cell image to suppress image noise in the original cell image to obtain a smooth image; performing local contrast equalization processing on the smoothed image, and enhancing the distinction between cells and background in the smoothed image by grayscale value stretching to obtain an enhanced image; performing phase consistency analysis and gradient calculation on the enhanced image, extracting cell edge features corresponding to low-contrast areas in the enhanced image, and forming an edge image; performing binarization and morphological operations on the edge image to remove noise points and connection broken boundaries in the edge image, thereby obtaining a connected domain boundary image; The connected domain boundary image is used as the preprocessed cell image.

[0092] In a possible implementation, the image segmentation module 203 is specifically configured to: Fitting a minimum bounding rectangle to each connected area in the preprocessed cell image to obtain a first candidate box set; Calculating, based on a preset contrast scoring function, a contrast score of the image area corresponding to each candidate frame in the first candidate frame set to obtain a contrast score set; Based on the contrast score set, sorting the candidate boxes in the first candidate box set in descending order to obtain a second candidate box set; Traversing each candidate box in the second candidate box set, calculating the intersection-and-union ratio of each candidate box with the remaining candidate boxes, and obtaining an intersection-and-union ratio set; For each candidate frame in the second candidate frame set, if its intersection-over-union ratio with other candidate frames is greater than a preset threshold, only the candidate frames with higher contrast scores are retained, and the remaining candidate frames are suppressed to obtain the target candidate frame set; The single-cell region corresponding to the target candidate frame set is size-normalized to obtain a standardized single-cell image.

[0093] In a possible implementation, the intelligent identification module 204 is specifically configured to: Inputting the standardized single cell image into a multi-branch convolutional network, and extracting cell features in the standardized single cell image in parallel through convolution kernels of different scales to obtain a multi-branch feature map; Calculating the channel weights of the multi-branch feature maps through an attention mechanism and performing weighted fusion to obtain a fused feature map; Through the classification branch, the fusion feature map is used to determine the benign and malignant cells and the lesion subtype, and the target category is output; The fusion feature map is scored for cell abnormality through a regression branch to obtain a score value; the classification branch and the regression branch are run in parallel; Based on the target category and the fused feature map, generating an interpretable heat map in real time through a gradient-weighted class activation mapping algorithm; Data integration is performed on the target category, the score value, and the interpretable heat map to obtain a lesion recognition result corresponding to the unstained biopsy cells.

[0094] In a possible implementation, the intelligent identification module 204 is specifically configured to: Obtaining a target category score corresponding to the target category; Calculating the gradient matrix of the target category score to the fused feature map; Calculate the global mean corresponding to the gradient matrix of each channel in the fused feature map to obtain channel weight data; The channel weight data is weighted and summed with the fusion feature map to generate an interpretable heat map.

[0095] In a possible embodiment, the lesion identification system 20 of the non-stained biopsy cells further includes an offline interpretation module 206, and the offline interpretation module 206 is specifically configured to: Before the structured integration of the lesion identification results is performed based on the preset structured report generation algorithm and statistical distribution modeling mechanism to obtain a structured cell diagnosis report containing a heat map overlay image, in response to a user's questioning instruction regarding the lesion identification results, a preset offline interpretation mechanism is triggered; Extracting key features that affect the lesion recognition result through the offline interpretation mechanism based on the standardized single-cell image, the fusion feature map corresponding to the standardized single-cell image, and the model decision logic of the deep learning model; Based on the key features, generating visual explanation content for the questioned instruction; The visual explanation content is associated with the lesion identification result and displayed so that the user can interactively verify the visual explanation content and the identification result; Receive user feedback instructions on the visual interpretation content, associate the feedback instructions with the lesion identification results, and form a closed-loop interaction record.

[0096] In a possible implementation, the offline interpretation mechanism includes a LIME offline interpretation mechanism and a SHAP offline interpretation mechanism, and the offline interpretation module 206 is specifically configured to: In response to a user's questioning instruction on the lesion identification result, at least one of a LIME offline interpretation mechanism and a SHAP offline interpretation mechanism is triggered.

[0097] The embodiment of the present invention provides a lesion identification system 20 for non-stained biopsy cells, which can implement the steps and effects of the lesion identification method for non-stained biopsy cells in Example 1. To avoid repetition, the present invention will not elaborate on them.

[0098] The beneficial effects of the present invention are embodied in: First, through the synchronous control of pulsed laser and high-speed imaging and image preprocessing, the technical bottleneck of low contrast in non-stained cell imaging was broken through, and clear imaging of high-speed flow cells and standardized single-cell extraction were achieved, which significantly improved the efficiency of sample processing and the stability of image quality. Secondly, based on the deep learning model, multi-scale features were extracted and lesion recognition results were generated, which overcame the defects of traditional pathological diagnosis that relied on staining and was highly subjective, and achieved accurate quantitative judgment of the benign and malignant nature of cells and the degree of lesions, improving the objectivity and accuracy of the recognition results. Finally, combined with the structured report generation mechanism, it provides the clinic with an integrated diagnostic output containing key information, while supporting subsequent interactive interpretation, which not only meets the core requirements of pathological diagnosis for efficiency and accuracy, but also enhances the traceability and clinical applicability of the results, greatly improving the diagnostic efficacy and medical collaboration efficiency in the non-stained biopsy scenario.

[0099] In the description of the embodiments of the present invention, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "center", "top", "bottom", "top", "bottom", "inside", "outside", "inner side", "outer side" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. Among them, "inside" refers to an internal or enclosed area or space. "Periphery" refers to the area surrounding a specific component or specific area.

[0100] In the description of the embodiments of the present invention, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly specifying the number of the technical features indicated. Therefore, a feature specified as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0101] In the description of the embodiments of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "installed," "connected," "connected," and "assembled" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to direct connections, indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention in specific circumstances.

[0102] In the description of the embodiments of the present invention, specific features, structures, materials or characteristics may be combined in an appropriate manner in any one or more embodiments or examples.

[0103] In describing the embodiments of the present invention, it should be understood that "-" and "~" represent a range between two values, and the range includes the endpoints. For example, "AB" represents a range greater than or equal to A and less than or equal to B. "A~B" represents a range greater than or equal to A and less than or equal to B.

[0104] In describing the embodiments of the present invention, the term "and / or" is used herein to describe a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. Furthermore, the character " / " is generally used herein to indicate that the associated objects are in an "or" relationship.

[0105] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying lesions in non-stained biopsy cells, characterized in that: include: By synchronously controlling the timing of pulsed laser and high-speed imaging, the high-speed flowing unstained biopsy cells are imaged and collected to obtain the original cell images. performing image preprocessing on the original cell image to obtain a preprocessed cell image; Based on a preset target cutting strategy, multi-target cutting is performed on the pre-processed cell image to obtain a standardized single cell image; Extracting multi-scale features of the standardized single-cell image based on a deep learning model to obtain lesion recognition results corresponding to the unstained biopsy cells, wherein the lesion recognition results include benign and malignant judgment results of the unstained biopsy cells, lesion scores, interpretable heat maps of the recognition basis, and feature data supporting offline interactive interpretation; Based on a preset structured report generation algorithm and statistical distribution modeling mechanism, the lesion identification results are structured and integrated to obtain a structured cell diagnosis report containing a heat map overlay image.

2. The method according to claim 1, characterized in that The method uses pulsed laser and high-speed imaging to synchronously control the timing of the pulsed laser and the high-speed imaging to collect images of the high-speed flowing unstained biopsy cells to obtain the original cell images, including: A pulsed laser light source and a high-speed imaging device are configured to introduce unstained biopsy cells into a microfluidic channel and control the flow velocity of the unstained biopsy cells to obtain a flow cell sample to be imaged; Calculating the delay time through a synchronous trigger mechanism to control the timing synchronization of laser pulse emission and image acquisition, photographing the flow cell sample entering the imaging acquisition window, and obtaining an initial cell image; When the signal-to-noise ratio of the initial cell image is lower than a preset threshold, the laser pulse energy or the imaging device gain parameter is dynamically adjusted to obtain an optimized cell image; Cells are continuously photographed based on optimized parameters to ultimately obtain high-contrast original cell images without smear.

3. The method according to claim 2, characterized in that The delay time is calculated by the synchronous trigger mechanism to control the timing synchronization of laser pulse emission and image acquisition, and the flow cell sample entering the imaging acquisition window is photographed to obtain an initial cell image, including: The synchronization trigger mechanism calculates the delay time according to the following formula: in, t d is the delay time required for unstained biopsy cells to enter the sensing area and the imaging acquisition window. L is the distance from the sensing area to the imaging acquisition window, v f The real-time flow rate of unstained biopsy cells; The single-frame exposure time of the high-speed imaging device is controlled to meet the following anti-smear conditions: in, t e is the single-frame exposure time of the high-speed imaging device, is the pixel pitch of the image sensor; At the moment when the unstained biopsy cells enter the sensing area, the delay time is added, and this moment is used as the trigger point to synchronously trigger the laser pulse emission and the high-speed imaging device exposure to photograph the unstained biopsy cells and obtain a high-contrast initial cell image without smear.

4. The method according to claim 1, wherein The performing image preprocessing on the original cell image to obtain a preprocessed cell image includes: performing Gaussian filtering on the original cell image to suppress image noise in the original cell image to obtain a smooth image; performing local contrast equalization processing on the smoothed image, and enhancing the distinction between cells and background in the smoothed image by grayscale value stretching to obtain an enhanced image; performing phase consistency analysis and gradient calculation on the enhanced image, extracting cell edge features corresponding to low-contrast areas in the enhanced image, and forming an edge image; performing binarization and morphological operations on the edge image to remove noise points and connection broken boundaries in the edge image, thereby obtaining a connected domain boundary image; The connected domain boundary image is used as the preprocessed cell image.

5. The method according to claim 1, wherein The method of performing multi-target cutting on the pre-processed cell image based on a preset target cutting strategy to obtain a standardized single cell image includes: Fitting a minimum bounding rectangle to each connected area in the preprocessed cell image to obtain a first candidate box set; Calculating, based on a preset contrast scoring function, a contrast score of the image area corresponding to each candidate frame in the first candidate frame set to obtain a contrast score set; Based on the contrast score set, sorting the candidate boxes in the first candidate box set in descending order to obtain a second candidate box set; Traversing each candidate box in the second candidate box set, calculating the intersection-and-union ratio of each candidate box with the remaining candidate boxes, and obtaining an intersection-and-union ratio set; For each candidate frame in the second candidate frame set, if its intersection-over-union ratio with other candidate frames is greater than a preset threshold, only the candidate frames with higher contrast scores are retained, and the remaining candidate frames are suppressed to obtain the target candidate frame set; The single-cell region corresponding to the target candidate frame set is size-normalized to obtain a standardized single-cell image.

6. The method according to claim 1, characterized in that The step of extracting multi-scale features of the standardized single-cell image based on a deep learning model to obtain a lesion recognition result corresponding to the unstained biopsy cell includes: Inputting the standardized single cell image into a multi-branch convolutional network, and extracting cell features in the standardized single cell image in parallel through convolution kernels of different scales to obtain a multi-branch feature map; Calculating the channel weights of the multi-branch feature maps through an attention mechanism and performing weighted fusion to obtain a fused feature map; Through the classification branch, the fusion feature map is used to determine the benign and malignant cells and the lesion subtype, and the target category is output; The fusion feature map is scored for cell abnormality through a regression branch to obtain a score value; the classification branch and the regression branch are run in parallel; Based on the target category and the fused feature map, generating an interpretable heat map in real time through a gradient-weighted class activation mapping algorithm; Data integration is performed on the target category, the score value, and the interpretable heat map to obtain a lesion recognition result corresponding to the unstained biopsy cells.

7. The method according to claim 6, characterized in that The method of generating an interpretable heat map in real time based on the target category and the fused feature map by a gradient weighted class activation mapping algorithm includes: Obtaining a target category score corresponding to the target category; Calculating the gradient matrix of the target category score to the fused feature map; Calculate the global mean corresponding to the gradient matrix of each channel in the fused feature map to obtain channel weight data; The channel weight data is weighted and summed with the fusion feature map to generate an interpretable heat map.

8. The method according to claim 1, characterized in that Before performing structured integration on the lesion identification results based on the preset structured report generation algorithm and statistical distribution modeling mechanism to obtain a structured cell diagnosis report containing a heat map overlay image, the method further includes: In response to a user's questioning instruction regarding the lesion identification result, triggering a preset offline interpretation mechanism; Extracting key features that affect the lesion recognition result through the offline interpretation mechanism based on the standardized single-cell image, the fusion feature map corresponding to the standardized single-cell image, and the model decision logic of the deep learning model; Based on the key features, generating visual explanation content for the questioned instruction; The visual explanation content is associated with the lesion identification result and displayed so that the user can interactively verify the visual explanation content and the identification result; Receive user feedback instructions on the visual interpretation content, associate the feedback instructions with the lesion identification results, and form a closed-loop interaction record.

9. The method according to claim 8, characterized in that The offline interpretation mechanism includes a LIME offline interpretation mechanism and a SHAP offline interpretation mechanism. In response to the user's questioning instruction on the lesion identification result, the preset offline interpretation mechanism is triggered, including: In response to a user's questioning instruction on the lesion identification result, at least one of a LIME offline interpretation mechanism and a SHAP offline interpretation mechanism is triggered.

10. A lesion identification system for non-stained biopsy cells, characterized in that: include: Ultra-high-speed imaging module, used to image and capture high-speed flowing unstained biopsy cells through synchronous control of pulsed laser and high-speed imaging timing to obtain original cell images; An image preprocessing module, configured to perform image preprocessing on the original cell image to obtain a preprocessed cell image; An image segmentation module is used to perform multi-target segmentation on the pre-processed cell image based on a preset target segmentation strategy to obtain a standardized single cell image; An intelligent recognition module is configured to extract multi-scale features of the standardized single-cell image based on a deep learning model to obtain lesion recognition results corresponding to the unstained biopsy cells, wherein the lesion recognition results include benign and malignant judgment results of the unstained biopsy cells, lesion scores, interpretable heat maps of the recognition basis, and feature data supporting offline interactive interpretation; The result output module is used to perform structured integration of the lesion identification results based on a preset structured report generation algorithm and statistical distribution modeling mechanism to obtain a structured cell diagnosis report containing a heat map overlay image.

Images