A method and system for identifying lesions in unstained biopsy cells

Abstract

This invention relates to the field of medical image processing, and more particularly to a method and system for identifying lesions in unstained biopsy cells. The method includes: acquiring images of rapidly flowing unstained biopsy cells by time-synchronized control of pulsed laser and high-speed imaging to obtain raw cell images; preprocessing the raw cell images to obtain preprocessed cell images; performing multi-target segmentation on the preprocessed cell images to obtain standardized single-cell images; extracting multi-scale features from the standardized single-cell images based on a deep learning model to obtain lesion identification results corresponding to the unstained biopsy cells; and integrating the lesion identification results in a structured manner based on a preset structured report generation algorithm and statistical distribution modeling mechanism to obtain a structured cell diagnostic report. This invention achieves automated lesion identification in unstained biopsy cells through pulsed laser imaging and deep learning feature extraction, efficiently generating reports and assisting in accurate diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for identifying lesions in unstained biopsy cells. Background Technology

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

[0003] However, existing unstained biopsy cell identification technologies have significant limitations: on the one hand, unstained cells have low contrast with the background, and imaging under high-speed flow conditions is prone to ghosting, making it difficult to obtain clear images and extract standardized single cells; on the other hand, traditional identification relies on manual slide reading, which is highly subjective and inefficient, and existing AI-assisted methods are also difficult to accurately quantify the degree of cell lesions due to incomplete feature extraction, thus limiting the reliability and applicability of the identification results.

[0004] Therefore, how to overcome the bottlenecks of poor imaging quality and insufficient recognition accuracy of unstained cells, and to build an efficient and accurate scheme for identifying unstained biopsy cell lesions in order to improve the objectivity, efficiency and clinical applicability of diagnosis, has become an urgent technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention provides a method and system for identifying lesions using unstained biopsy cells.

[0006] A first aspect of the present invention provides a method for identifying lesions in unstained biopsy cells, comprising:

[0007] By using time-synchronized control of pulsed laser and high-speed imaging, images of unstained biopsy cells flowing at high speed are acquired to obtain raw cell images.

[0008] The original cell image is preprocessed to obtain a preprocessed cell image;

[0009] Based on a preset target cutting strategy, the preprocessed cell image is subjected to multi-target cutting to obtain a standardized single-cell image.

[0010] Multi-scale features of the standardized single-cell image are extracted based on a deep learning model to obtain the lesion identification results corresponding to the unstained biopsy cells. The lesion identification results include the benign or malignant judgment results of the unstained biopsy cells, lesion scores, interpretability heatmaps of the identification basis, and feature data that supports offline interactive interpretation.

[0011] 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 overlaid heatmap images.

[0012] A second aspect of the present invention provides a lesion identification system for unstained biopsy cells, comprising:

[0013] The ultra-high-speed imaging module is used to acquire images of unstained biopsy cells flowing at high speed by controlling the timing of pulsed laser and high-speed imaging, thereby obtaining raw cell images.

[0014] The image preprocessing module is used to perform image preprocessing on the original cell image to obtain a preprocessed cell image;

[0015] The image segmentation module is used to perform multi-target segmentation on the preprocessed cell image based on a preset target segmentation strategy to obtain a standardized single-cell image.

[0016] The intelligent recognition module is used to extract multi-scale features of the standardized single-cell image based on a deep learning model to obtain the lesion recognition result corresponding to the unstained biopsy cell. The lesion recognition result includes the benign or malignant judgment result of the unstained biopsy cell, the lesion score, the interpretability heatmap of the recognition basis, and feature data that supports offline interactive interpretation.

[0017] The results output module is used to integrate the lesion identification results in a structured manner based on a preset structured report generation algorithm and statistical distribution modeling mechanism, so as to obtain a structured cell diagnosis report containing heatmap overlay images.

[0018] The beneficial effects of this invention are as follows: First, by using time-synchronous control of pulsed laser and high-speed imaging, along with image preprocessing, the technical bottleneck of low contrast in unstained cell imaging is overcome, achieving clear imaging of high-speed flowing cells and standardized single-cell extraction, significantly improving sample processing efficiency and image quality stability. Second, by extracting multi-scale features and generating lesion identification results based on a deep learning model, the invention overcomes the shortcomings of traditional pathological diagnosis, which relies on staining and is highly subjective, achieving precise quantitative judgment of cell benignity / malignancy and lesion severity, thus improving the objectivity and accuracy of the identification results. Finally, combined with a structured report generation mechanism, the invention provides clinicians with integrated diagnostic output containing key information, while supporting subsequent interactive interpretation. This 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 diagnostic efficacy and medical collaboration efficiency in unstained biopsy scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for identifying lesions in unstained biopsy cells provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of a lesion identification system for unstained biopsy cells provided by the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The existing technologies and their main problems can be summarized in the following aspects:

[0023] 1. High-throughput bottleneck in image acquisition and processing:

[0024] Existing cell imaging systems (such as TCT and imaging flow cytometers) have limitations in image acquisition and processing mechanisms. To obtain high-resolution images, traditional systems often require extended exposure times or multiple scans, leading to increased processing time per sample. At the same time, data processing capabilities for large sample scenarios have not been improved accordingly, creating a throughput bottleneck in the "imaging-processing" chain. This makes it difficult to meet the needs of real-time, rapid, large-scale screening, especially in grassroots or mobile testing scenarios where efficiency shortcomings are particularly prominent.

[0025] 2. Insufficient recognition accuracy for low-contrast / unstained cells:

[0026] Existing cell identification technologies (such as traditional edge detection, threshold segmentation algorithms, and some deep learning models) have inherent limitations when processing transparent or unstained cell samples. These samples have low contrast between cells and background, and blurred boundaries. Traditional algorithms are easily affected by background noise, leading to incomplete extraction of the target region. Even with morphological optimization, it is still difficult to accurately distinguish cells from the background, easily resulting in missed detection of abnormal cells. Some deep learning models also suffer from low recognition accuracy due to insufficient proportion of low-contrast samples in their training data.

[0027] 3. Complexity and deployment limitations of sample preprocessing:

[0028] Current mainstream technologies (such as TCT) rely on multi-step pretreatment mechanisms, requiring processes such as centrifugation, chemical staining, and slide fixation. This is not only cumbersome to operate but also dependent on specialized consumables, centrifugation equipment, and professional operators. While some automated systems simplify operations, they still have high requirements for operating environments such as temperature and cleanliness, making deployment difficult in resource-constrained primary healthcare institutions and mobile testing vehicles, thus limiting the technology's versatility.

[0029] 4. Lack of interpretability in the decisions made by intelligent recognition systems:

[0030] Existing cell identification technologies based on deep learning (such as CNNs) operate on a "black box" mechanism in their core feature extraction and classification processes, failing to provide clear decision-making criteria (such as which cell morphological features support an "abnormal" judgment). Medical personnel struggle to verify the validity of the results, leading to decreased trust in the system and directly impacting the actual adoption rate of such technologies in clinical diagnosis.

[0031] Example 1:

[0032] Reference manual attached Figure 1 The diagram shows a flowchart of a method for identifying lesions in unstained biopsy cells provided by the present invention.

[0033] This invention provides a method for identifying lesions in unstained biopsy cells, comprising:

[0034] S1: By controlling the timing of pulsed laser and high-speed imaging, images of unstained biopsy cells flowing at high speed are acquired to obtain raw cell images.

[0035] This step addresses existing problems in cell image acquisition, such as blurred images, uneven illumination, motion blur, and insufficient frame rate. It combines the particle image formation mechanism under high-speed flow conditions to construct an image acquisition system that integrates optics, flow control, imaging synchronization, and parameter adjustment mechanisms.

[0036] Specifically, pulsed modulated laser is used as the illumination source. The laser beam is shaped into a near-circular spot with a Gaussian energy distribution through a spatial light modulation system, making the illumination intensity more concentrated in the cell edge region and enhancing the resolvability of the cell boundary. After beam expansion, collimation, and homogenization, the spot is vertically projected onto the imaging acquisition window of the microfluidic channel.

[0037] To ensure uniform illumination within the imaging acquisition window, a multi-stage aspherical lens group is introduced into the optical path design to control the radial diffusion of illumination energy and achieve high uniformity coverage.

[0038] Considering that cells flow at a constant velocity in microfluidic channels, image motion blur would occur under conventional continuous illumination without illumination modulation. Therefore, this invention designs a pulse-width modulation laser driving mechanism that matches the flow rate, causing the light source to flash only at the moment of image acquisition, thereby freezing the cell motion trajectory and improving transient image clarity.

[0039] In one possible implementation, S1 involves timing synchronization control of pulsed laser and high-speed imaging to image and acquire raw cell images of rapidly flowing, unstained biopsy cells, including:

[0040] By configuring a pulsed laser source and a high-speed imaging device, unstained biopsy cells are introduced into a microfluidic channel and the flow rate of the unstained biopsy cells is controlled to obtain a flow cell sample to be imaged.

[0041] The delay time is calculated by a synchronous triggering mechanism to control the timing synchronization between laser pulse emission and image acquisition, and the flowing cell sample entering the imaging acquisition window is photographed to obtain the initial cell image;

[0042] If the signal-to-noise ratio of the initial cell image is lower than a preset threshold, the laser pulse energy or the gain parameter of the imaging device is dynamically adjusted to obtain an optimized cell image.

[0043] By continuously photographing cells using optimized parameters, high-contrast raw cell images without motion blur are finally obtained.

[0044] Specifically, in the system configuration and sample preparation stage, a pulsed laser 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 source needs to support pulse modulation, and its output beam, after spatial shaping, beam expansion, and homogenization, forms a light field with uniform energy distribution, which is then vertically projected onto the imaging acquisition window of the microfluidic channel. Unstained biopsy cell samples (such as cervical exfoliated cells) are then introduced into the microfluidic channel, and the cell flow velocity is adjusted to a target value (e.g., above 5 m / s) through a precision fluid control system, forming a high-speed, stable flow of cells to be imaged. In the timing synchronization and initial imaging stage, a synchronous triggering mechanism between the pulsed laser and the high-speed imaging device is established to capture images of the flowing cell samples entering the imaging acquisition window, obtaining initial cell images.

[0045] In one possible implementation, a delay time is calculated through a synchronous triggering mechanism to control the timing synchronization of laser pulse emission and image acquisition, thereby capturing images of the flowing cell sample entering the imaging acquisition window to obtain an initial cell image. Specifically, this includes:

[0046] The synchronous triggering mechanism calculates the delay time based on the following formula:

[0047]

[0048] in, t d The time delay required for unstained biopsy cells to travel from the sensing area to the imaging acquisition window. L The distance from the sensing area to the imaging acquisition window. v f Real-time flow rate for unstained biopsy cells;

[0049] The single-frame exposure time of high-speed imaging equipment must meet the following anti-ghosting conditions:

[0050]

[0051] in, t e This refers to the single-frame exposure time of a high-speed imaging device. The pixel pitch of the image sensor;

[0052] At the moment when unstained biopsy cells enter the sensing area, a delay time is added, and this moment is used as the trigger point to simultaneously trigger the laser pulse emission and the high-speed imaging device exposure, so as to capture the unstained biopsy cells and obtain a high-contrast initial cell image without ghosting.

[0053] For example, in the above anti-ghosting condition, the left side of the inequality represents the actual displacement of a cell in the field of view during a single frame exposure time, while the right side represents the maximum allowable displacement range in the image, i.e., one pixel width. This constraint is used to define the minimum frame rate or maximum permissible exposure time required by the imaging system at the target flow rate, thereby ensuring that cell boundaries do not overlap or become blurred in the image recording.

[0054] To achieve precise imaging timing control, the system introduces an FPGA-based synchronous triggering mechanism, establishing a temporal logical coupling between the cell passage sensing area and the camera acquisition window. Cell passage signals are acquired via a front-mounted optical sensor or fluid disturbance detection device. The FPGA generates precise delay timing based on this signal and simultaneously drives the laser pulse and image acquisition. Its core scheduling mechanism is based on the aforementioned timing control relationship. Through precise control This allows for locking onto the timing of imaging individual cells, further ensuring that image acquisition begins precisely when the cell is fully within the imaging acquisition window, thereby improving cell centering rate and imaging efficiency in the image.

[0055] Considering the significant differences in image contrast and brightness under different sample types and background conditions, an image quality feedback control mechanism is introduced into the high-speed acquisition process. Specifically, this is achieved by calculating the average grayscale value in the image frame. and variance Real-time evaluation of frame-level signal-to-noise ratio (SNR):

[0056]

[0057] This metric serves as a quantitative standard for image sharpness and separability, combined with a pre-defined minimum acceptable threshold. Once detected The system automatically adjusts the laser pulse energy or the camera's analog gain coefficient 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 overall imaging quality.

[0058] Meanwhile, the design of this step also pays special attention to the issues of optical path drift and thermal stability during continuous acquisition. By introducing a thermal compensation structure, an optical axis collimator lock, and an imaging surface reflection calibration system, it is possible to achieve fine-tuning and closed-loop correction of the laser illumination path and the focal length of the imaging surface, ensuring that the positional relationship between the laser illumination and the imaging acquisition window remains stable after long-term operation, thereby preventing image shift and defocusing caused by optical axis offset.

[0059] After image acquisition is completed, the image data will be transmitted in real time in streaming form to facilitate subsequent image preprocessing and image segmentation steps, providing a raw data foundation with stable edge features and high contrast for subsequent cell identification.

[0060] S1 achieves high-definition imaging of unstained cells without motion blur by using laser pulse modulation and time-synchronized control of a high-speed camera, combined with flow velocity-frame rate coupling constraints and a real-time fluid detection feedback mechanism. This is achieved by strictly controlling the exposure time of a single frame and the displacement of cell movement within the pixel level, enabling high-resolution imaging of unstained cells in high-speed flowing samples without motion blur. It solves the problems of image blurring, motion blur, and insufficient frame rate in traditional imaging techniques for high-speed flowing samples, especially improving the imaging distortion of unstained transparent cells, providing high-fidelity data for subsequent identification and meeting the needs of large-scale rapid screening.

[0061] S2: Perform image preprocessing on the original cell image to obtain the preprocessed cell image.

[0062] In one possible implementation, image preprocessing is performed on the original cell image to obtain a preprocessed cell image, specifically including:

[0063] Gaussian filtering is applied to the original cell image to suppress image noise and obtain a smooth image.

[0064] Local contrast equalization is performed on the smoothed image, and the distinction between cells and background in the smoothed image is enhanced by stretching gray values ​​to obtain an enhanced image;

[0065] Phase consistency analysis and gradient calculation are performed on the enhanced image to extract cell edge features corresponding to low-contrast regions in the enhanced image, forming an edge image;

[0066] Binarization and morphological operations are performed on the edge image to remove noise and broken connection boundaries, resulting in a connected domain boundary image.

[0067] The connected component boundary image is used as the preprocessed cell image.

[0068] It is worth noting that the task of image preprocessing is to perform noise suppression, contrast enhancement, and edge enhancement on the rapidly acquired cell images without relying on staining, thereby extracting image regions with clear boundaries to ensure the accuracy and standardization of the 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, while also meeting the requirements of high-throughput processing, ensuring that image structural information is not lost or distorted while maintaining processing speed.

[0069] Specifically, to address the potential high-frequency electronic noise and low-light disturbances during the imaging process, a joint image preprocessing method based on Gaussian filtering and local contrast equalization is first employed. For any image frame... Its Gaussian filtering operation in the spatial domain can be expressed as:

[0070]

[0071] in, This represents the smoothed image. The two-dimensional Gaussian kernel function is defined as follows:

[0072]

[0073] in, The standard deviation of the Gaussian kernel controls the filtering strength.

[0074] This filtering process can effectively remove random noise, but it weakens edge information. Therefore, a local contrast enhancement operator needs to be introduced after filtering.

[0075] For each pixel Calculate its Maximum gray value in the neighborhood Minimum grayscale value Then, local enhancement is achieved through linear stretching transformation:

[0076]

[0077] The enhanced image is obtained. Its edge gradient is more obvious, which provides a basis for subsequent edge extraction.

[0078] In terms of edge extraction, considering the problem that traditional edge detection methods such as Sobel and Canny have weak edge response and poor robustness in transparent cells or low-contrast images, this invention introduces a composite edge extraction algorithm based on image gradient field convolution and phase consistency analysis.

[0079] First, the gradient image is calculated using the first derivative. The main edge direction is extracted using a directional high-pass filter. Based on this, a phase coherence function is introduced, whose expression is:

[0080]

[0081] in, For the first Amplitude of each frequency channel For phase, For average phase, Noise threshold To keep the quantity small, to prevent the denominator from being zero.

[0082] Phase consistency functions can reflect the consistency of frequency domain structural changes in an image and can detect boundary information in regions with weak intensity changes. They are suitable for edge extraction in images of insufficiently stained or transparent cells.

[0083] After obtaining the edge image, the system binarizes it and removes isolated noise points and broken boundaries through morphological operations (opening and closing operations) to obtain candidate cell regions with strong connectivity and clear boundaries.

[0084] S2 extracts stable structural information from weak grayscale regions through phase consistency analysis and fuses it with gradient operator results. Combined with a multi-target selection mechanism of non-maximum suppression, it achieves accurate extraction of low-contrast, transparent cell edges. This solves the problem of poor recognition of unstained and adherent cells in traditional edge detection, effectively strips overlapping regions, provides clear edge features for subsequent segmentation, improves the robustness of image preprocessing, and ensures the quality and stability of input data.

[0085] S3: Based on a preset target cutting strategy, perform multi-target cutting on the preprocessed cell image to obtain a standardized single-cell image.

[0086] In one possible implementation, based on a preset target segmentation strategy, the preprocessed cell image is subjected to multi-target segmentation to obtain a standardized single-cell image, specifically including:

[0087] The minimum bounding rectangle is fitted to each connected region in the preprocessed cell image to obtain the first candidate box set;

[0088] Based on a preset contrast scoring function, the contrast score of the image region corresponding to each candidate box in the first candidate box set is calculated to obtain a contrast score set.

[0089] Based on the contrast score set, the candidate boxes in the first candidate box set are sorted in descending order to obtain the second candidate box set;

[0090] Iterate through each candidate box in the second candidate box set, calculate the intersection-union ratio (IUR) of each candidate box with the other candidate boxes, and obtain the IUR set.

[0091] For each candidate box in the second candidate box set, if the intersection-union ratio with other candidate boxes is greater than a preset threshold, only the candidate box with the higher contrast score is retained, and the other candidate boxes are suppressed to obtain the target candidate box set.

[0092] The size of the single-cell regions corresponding to the target candidate box set is standardized to obtain a standardized single-cell image.

[0093] It is worth noting that, considering the often-adhesive, overlapping, and significant size variations of cervical cells, traditional target extraction methods based on connected components are prone to missegmentation or misidentification of adhered regions. Therefore, this invention designs a target segmentation strategy based on a combination of minimum bounding rectangle and non-maximum suppression.

[0094] For each connected region, first fit its minimum bounding rectangle to obtain a set of candidate boxes. For all According to the contrast scoring function within the image region Sort the boxes in descending order, then iterate through each box and calculate the Intersection over Union (IoU) with the remaining boxes, as defined below:

[0095]

[0096] like If the score exceeds a preset threshold, the candidate bounding box with the higher score is 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 discriminativeness and uniqueness of the cut.

[0097] The final output single-cell image blocks are uniformly adjusted to a fixed-size standard input format, i.e., size standardization processing is performed to obtain standardized single-cell images, which are then entered into the next stage along with the original cell image location information for subsequent multi-classification diagnosis and lesion feature determination.

[0098] The aforementioned image preprocessing and target segmentation processes are characterized by automation, adaptability, and parallelism. They can maintain stable image quality control and target segmentation consistency under high-throughput input streams, laying a stable foundation for intelligent analysis and clinical interpretation in the backend.

[0099] Based on a preset segmentation strategy, S3 performs multi-target segmentation on preprocessed cell images through bounding rectangle fitting and intersection-union ratio (IU) constraints, achieving effective separation of adherent cells and standardized single-cell image extraction. This solves the problems of cell target extraction failure or overlapping recognition in traditional segmentation algorithms, ensuring that the output single-cell images have uniform size and unique boundaries. This provides standardized input for subsequent intelligent recognition, improving the accuracy and stability of subsequent recognition.

[0100] S4: Based on a deep learning model, multi-scale features of standardized single-cell images are extracted to obtain lesion identification results corresponding to unstained biopsy cells. The lesion identification results include the benign or malignant judgment results of unstained biopsy cells, lesion scores, interpretable heatmaps of the identification basis, and feature data that supports offline interactive interpretation.

[0101] 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 structural information and making medical significance judgments. It is a key link in the present invention to realize early diagnosis and automatic assisted decision-making.

[0102] Specifically, given that cervical exfoliated cells exhibit weak detail, complex texture distribution, and strong individual variability under unstained or weakly stained conditions, traditional shallow classifiers (such as support vector machines or K-nearest neighbors algorithms) perform poorly in high-dimensional feature extraction and heterogeneous sample identification. Therefore, this step constructs a deep neural structure based on the fusion of multi-branch convolutional networks and attention mechanisms. This structure can simultaneously extract key information such as cell boundary texture, nucleocytoplasmic ratio, and nuclear atypia at different scales. Simultaneously, it utilizes an attention weighting mechanism to dynamically adjust the focus on local regions, thereby improving the sensitivity to weak features of early lesions.

[0103] In one possible implementation, multi-scale features of standardized single-cell images are extracted based on a deep learning model to obtain lesion identification results corresponding to unstained biopsy cells, specifically including:

[0104] Standardized single-cell images are input into a multi-branch convolutional network, and cell features in the standardized single-cell images are extracted in parallel through convolutional kernels of different scales to obtain multi-branch feature maps;

[0105] The channel weights of the multi-branch feature maps are calculated using an attention mechanism and weighted fusion is performed to obtain the fused feature map.

[0106] By using classification branches, the fused feature map is used to determine the benign or malignant nature of cells and the subtype of lesions, and the target category is output.

[0107] The regression branch scores the degree of cell abnormality in the fused feature map to obtain a score value; the classification branch and the regression branch run in parallel.

[0108] Based on the target category and the fused feature map, an interpretable heatmap is generated in real time using a gradient-weighted class activation mapping algorithm.

[0109] By integrating the target category, score, and interpretability heatmap, the lesion identification results corresponding to unstained biopsy cells are obtained.

[0110] Specifically, the input to the neural network is a set of standardized cell images. ,in h Image height, w Image width, This represents the number of channels.

[0111] The neural network structure consists of three parts:

[0112] The first part is the feature extraction backbone network, which uses multi-scale convolutional kernels to construct multiple parallel convolutional paths, defining the... The output of the path convolution is:

[0113]

[0114] in, For the first Convolutional kernels for each path, This represents the convolution operation. For bias terms, To modify the linear activation function.

[0115] The feature maps of all paths are merged and then weighted and fused using a channel attention module. This step involves processing the feature map of each channel... Calculate its global average pooling value Then, learnable weight vectors are introduced. The weighted output is obtained as follows:

[0116]

[0117] in, For the final fused feature map, For the first The importance weights of each channel can be learned through fully connected layers and Softmax normalization. This mechanism can significantly enhance the response capability to anomalous regions or features with weak structures, effectively improving the discrimination capability of this invention in low-contrast images.

[0118] After feature fusion is complete, it is fed into parallel classification and regression branches.

[0119] The classification branch is used to determine the benignity or malignancy of cells and output lesion subtype labels. Supervised learning is performed using the cross-entropy loss function, defined as:

[0120]

[0121] in, For the one-hot encoding of the actual category, To predict class probabilities, The total number of categories, k This is the category number.

[0122] The regression branch is used to determine the degree of cell abnormality or morphological score, employing mean squared error loss.

[0123]

[0124] in, For the actual scores of the sample, These are the model's predicted values. This represents the number of samples.

[0125] In one possible implementation, an interpretable heatmap is generated in real time based on the target category and the fused feature map using a gradient-weighted class activation mapping algorithm, specifically including:

[0126] Obtain the target category score corresponding to the target category;

[0127] Calculate the gradient matrix of the target category score with respect to the fused feature map;

[0128] Calculate the global mean of the gradient matrix corresponding to each channel in the fused feature map to obtain the channel weight data;

[0129] The channel weight data and the fused feature map are weighted and summed to generate an interpretable heatmap.

[0130] To address the lack of credibility caused by the "black box" nature of AI diagnosis in clinical applications, this invention integrates an interpretability enhancement mechanism at the inference end.

[0131] Specifically, a visualization method based on gradient-weighted class activation mapping is introduced. By backpropagating the gradient of the model output relative to the intermediate convolutional features, the contribution of each spatial location to the final classification result is calculated.

[0132] For the Each feature map has a weight defined by the following expression:

[0133]

[0134] in, Indicates the target category score. For the first Each channel is located at ( i , j The activation value of ) For the first The weight of each channel, This is the normalization factor.

[0135] The final interpretability heatmap is obtained by the following weighted combination:

[0136]

[0137] This heatmap will be overlaid on the original cell image to visualize the classification decision criteria, allowing users to intuitively understand the logical correspondence between the model's focus area and the diagnostic results, thus enhancing clinical adoption. The user in this disclosure can be a doctor or related personnel; this invention does not limit this possibility.

[0138] S4, through a deep model that integrates multi-path convolution and channel attention mechanisms, extracts cellular features at different scales in parallel and generates interpretable heatmaps using class activation mapping, achieving accurate judgment and visualization of cell benign / malignant and lesion severity. It addresses the shortcomings of traditional CNN models in extracting weak lesion features and lacking interpretability in the decision-making process, enhancing the model's ability to identify early lesions and its generalization capabilities. Furthermore, by making the results transparent, it improves the clinical acceptance of lesion identification results.

[0139] 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 containing overlaid heatmap images.

[0140] This step is the information carrier and feedback hub for end users in the technical solution of this invention. It is responsible for presenting the lesion identification results in a clinically acceptable, interpretable and interactive manner, and supports the connection with the hospital information system (HIS) or electronic medical record system (EMR) to realize closed-loop management and traceability analysis of diagnostic data.

[0141] Specifically, the system first receives three types of data: the first is the benign or 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 heatmap generated based on gradient backpropagation.

[0142] To ensure consistent and structured information delivery, the system constructs a multi-channel result fusion model and uses a structured report generation algorithm to standardize and integrate heterogeneous output data.

[0143] Let the total number of cell samples be , No. The classification results for each sample are The lesion score was The feature-focused region heatmap is Then the output record of each cell can be represented as a triple:

[0144]

[0145] all The samples are collected and used to construct the total output matrix of the sample set. Then, based on the patient information of the cell source, the sample timestamp, and the test batch number, an index mapping is performed to establish diagnostic-level data structure entries.

[0146] To enhance the interpretability of diagnostic results, the system introduces a statistical distribution modeling mechanism, which integrates the score set... Standardization analysis is performed under the assumption of normal distribution. The sample mean score is defined as... The variance is Each rating can then be mapped to a standard score:

[0147]

[0148] according to Based on their location within a standard normal distribution, cells can be categorized into three areas: normal, slightly suspicious, and highly abnormal, and represented graphically using different color codes. This mechanism allows doctors to intuitively identify potentially high-risk samples through the scoring distribution without needing to understand the internal workings of deep networks.

[0149] In terms of heatmap presentation, this invention associates each cell image with its corresponding heatmap of interest. Channel overlay is performed to generate a color artifact image. The red channel value of each pixel is defined by the following formula:

[0150]

[0151] in, The red channel of the original image. This is a heatmap with weighted coefficients. This image is used to assist doctors or other users in identifying key areas of interest for the model and to manually verify lesion identification results. Specifically, for example, users can click on a specific area to access its corresponding original cell image, feature distribution curve, and interpretation logic, thereby constructing a human-computer interactive interpretation loop.

[0152] To enhance compatibility with clinical systems, the result output module for step S5 also incorporates a standardized interface supporting HL7 and FHIR protocols. Through a unified data structure, it connects to the hospital's internal diagnostic record management system, enabling the archiving, retrieval, and sharing of interpretation results. Each report includes a data summary, an AI score distribution chart, a highlighted image page, and interpretable documentation, ensuring that the diagnostic report is not only technically sound but also medically readable.

[0153] Furthermore, to evaluate the reliability of the model's output, the results output module introduces an output confidence evaluation mechanism, which achieves result quality control by calculating the prediction consistency index of the model across multiple interpretation paths. This mechanism is used to obtain predicted labels for the same cell image in different feature subspaces. Define the consistency metric function as follows:

[0154]

[0155] in, Indicates label Number of times it appears This represents the total number of paths. This consistency value is appended to the results as a confidence label for doctors to reference during the interpretation process.

[0156] S5 integrates lesion identification results with overlaid heatmap images through a structured report generation algorithm and statistical distribution modeling mechanism, constructing a three-in-one output structure of image, score, and heatmap, supporting user interaction verification and feedback. It solves the problems of traditional systems having a single output format and lacking transparent explanation, achieving precise correspondence between diagnostic results and the original cell image region and model response basis, improving the visibility and operability of the results, while also supporting human-computer interactive interpretation, enhancing clinical applicability.

[0157] In one possible implementation, before structurally integrating 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 overlaid heatmap images, the method further includes:

[0158] In response to a user's questioning of the lesion identification results, a preset offline interpretation mechanism is triggered;

[0159] Through an offline interpretation mechanism, based on standardized single-cell images, the fusion feature maps corresponding to standardized single-cell images, and the model decision logic of deep learning models, key features affecting lesion recognition results are extracted.

[0160] Based on key features, generate visual explanations of the challenge commands;

[0161] The visual explanations are displayed in conjunction with the lesion identification results, so that users can interact and verify the visual explanations and identification results.

[0162] It receives user feedback instructions on the visual explanation content, associates the feedback instructions with the lesion recognition results, and forms a closed-loop interaction record.

[0163] Specifically, when a user challenges the lesion identification result, a preset offline interpretation mechanism is triggered. In this invention, users can click on a specific area in the structured cell diagnostic report to retrieve the corresponding original cell image, feature distribution curve, and interpretation logic. The interpretation logic can include cell morphology parameters that the model focuses on (such as nucleocytoplasmic ratio, edge smoothness, etc.), and the degree of influence of each parameter on the lesion identification result.

[0164] 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 challenge to the lesion identification results, a preset offline interpretation mechanism is triggered, including:

[0165] In response to a user's questioning command regarding the lesion identification results, at least one of the LIME offline interpretation mechanism and the SHAP offline interpretation mechanism is triggered.

[0166] Specifically, if the LIME offline interpretation mechanism is triggered: Based on standardized single-cell images, a sample set is generated by perturbating local regions. Combined with the sensitivity rules of 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 regions (i.e. key features) that significantly affect the lesion recognition results are selected by ranking the model weights. These regions are then superimposed with the original cell image to generate a visual interpretation map, which serves as the visual interpretation content to intuitively show the correlation between local image features and recognition results.

[0167] If the SHAP offline interpretation mechanism is triggered, cell morphology parameters (such as nucleocytoplasmic ratio, edge smoothness, etc.) that the model decision logic is concerned with are extracted from the fusion feature map corresponding to the standardized single-cell image as key features. 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 plot is generated by sorting by contribution, and the influence relationship between features and recognition results is shown by distinguishing between positive and negative contributions, which serves as a visual interpretation to clearly present the decision basis at the feature level.

[0168] It is worth noting that both the questioning command and the feedback command are issued through a preset interaction method. The feedback command can be issued through any one or more of the following interaction methods, and this invention does not limit the specific methods used:

[0169] 1. Result Confirmation / Correction: Click "Approval" or "Disapproval" to correct the score and indicate the reason for the error for the disapproved item.

[0170] 2. Annotation: Mark missed features on the image and add text annotations (such as "Nuclear morphology is normal here").

[0171] 3. Quantitative scoring: Feedback on the effectiveness of the explanation is provided through star ratings or scales (such as "Explanation clarity: high / medium / low").

[0172] The challenge command can be issued through any one or more of the following interactive methods, and this invention does not limit the scope of the challenge:

[0173] 1. Graphical Click: Click on key areas such as heatmaps and ratings, or preset buttons (such as "Result in Doubt") to trigger targeted questioning.

[0174] 2. Option selection: Select the preset question type such as "Classification error" or "Insufficient explanation" through the drop-down menu or checkbox.

[0175] 3. Text Input: Enter your specific question in the text box (e.g., "Why was this cell judged to be highly abnormal?").

[0176] 4. Voice assistance: Voice commands (such as "question the score of the 3rd cell") are adapted to multiple operation scenarios.

[0177] This step utilizes LIME or SHAP offline interpretation mechanisms to extract key features (such as cell nucleus morphology and edge features) that influence the recognition results, based on standardized single-cell images, fused feature maps, and model decision logic, and generates explanatory information, enabling targeted explanations of questioning commands. This solves 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 compensating for the lack of interpretability in traditional AI model decisions.

[0178] Example 2

[0179] Reference manual attached Figure 2 The diagram shows a schematic of the structure of a lesion identification system for unstained biopsy cells provided by the present invention.

[0180] An embodiment of the present invention provides a lesion identification system 20 for unstained biopsy cells, comprising:

[0181] The ultra-high-speed imaging module 201 is used to acquire images of unstained biopsy cells flowing at high speed by timing synchronization control of pulsed laser and high-speed imaging, and obtain raw cell images.

[0182] Image preprocessing module 202 is used to perform image preprocessing on the original cell image to obtain a preprocessed cell image;

[0183] Image segmentation module 203 is used to perform multi-target segmentation on the preprocessed cell image based on a preset target segmentation strategy to obtain a standardized single-cell image;

[0184] The intelligent recognition module 204 is used to extract multi-scale features of the standardized single-cell image based on a deep learning model to obtain the lesion recognition result corresponding to the unstained biopsy cell. The lesion recognition result includes the benign or malignant judgment result of the unstained biopsy cell, the lesion score, the interpretability heatmap of the recognition basis, and feature data that supports offline interactive interpretation.

[0185] The result output module 205 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 overlaid heatmap images.

[0186] In one possible implementation, the ultra-high-speed imaging module 201 is specifically used for:

[0187] A pulsed laser source and a high-speed imaging device are configured to introduce unstained biopsy cells into a microfluidic channel and control the flow rate of the unstained biopsy cells to obtain a flow cell sample to be imaged.

[0188] The delay time is calculated by a synchronous triggering mechanism to control the timing synchronization of laser pulse emission and image acquisition, and the flowing cell sample entering the imaging acquisition window is photographed to obtain an initial cell image;

[0189] If the signal-to-noise ratio of the initial cell image is lower than a preset threshold, the laser pulse energy or the gain parameter of the imaging device is dynamically adjusted to obtain an optimized cell image.

[0190] By continuously photographing cells using optimized parameters, high-contrast raw cell images without motion blur are finally obtained.

[0191] In one possible implementation, the ultra-high-speed imaging module 201 is specifically used for:

[0192] The synchronization triggering mechanism calculates the delay time according to the following formula:

[0193]

[0194] in, t d The time delay required for unstained biopsy cells to travel from the sensing area to the imaging acquisition window. L The distance from the sensing area to the imaging acquisition window. v f Real-time flow rate for unstained biopsy cells;

[0195] The single-frame exposure time of the high-speed imaging device is controlled to meet the following anti-ghosting conditions:

[0196]

[0197] in, t e This refers to the single-frame exposure time of a high-speed imaging device. The pixel pitch of the image sensor;

[0198] At the moment the unstained biopsy cell enters 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, so as to capture the unstained biopsy cell and obtain a high-contrast initial cell image without ghosting.

[0199] In one possible implementation, the image preprocessing module 202 is specifically used for:

[0200] Gaussian filtering is applied to the original cell image to suppress image noise and obtain a smooth image.

[0201] The smoothed image is subjected to local contrast equalization processing, and the distinction between cells and background in the smoothed image is enhanced by stretching gray values ​​to obtain an enhanced image;

[0202] Phase consistency analysis and gradient calculation are performed on the enhanced image to extract cell edge features corresponding to low-contrast regions in the enhanced image, forming an edge image;

[0203] Binarize and perform morphological operations on the edge image to remove noise and broken connection boundaries, thereby obtaining a connected component boundary image.

[0204] The connected domain boundary image is used as the preprocessed cell image.

[0205] In one possible implementation, the image cropping module 203 is specifically used for:

[0206] For each connected region in the preprocessed cell image, fit the minimum bounding rectangle to obtain the first candidate box set;

[0207] Based on a preset contrast scoring function, the contrast score of the image region corresponding to each candidate box in the first candidate box set is calculated to obtain a contrast score set.

[0208] Based on the contrast score set, the candidate boxes in the first candidate box set are sorted in descending order to obtain the second candidate box set;

[0209] Iterate through each candidate box in the second candidate box set, calculate the intersection-union ratio (IUU) of each candidate box with the other candidate boxes, and obtain the IUU set;

[0210] For each candidate box in the second candidate box set, if the intersection-union ratio with other candidate boxes is greater than a preset threshold, only the candidate box with the higher contrast score is retained, and the other candidate boxes are suppressed to obtain the target candidate box set.

[0211] The size of the single-cell regions corresponding to the target candidate box set is standardized to obtain a standardized single-cell image.

[0212] In one possible implementation, the intelligent recognition module 204 is specifically used for:

[0213] The standardized single-cell image is input into a multi-branch convolutional network, and cell features in the standardized single-cell image are extracted in parallel by convolutional kernels of different scales to obtain a multi-branch feature map.

[0214] The channel weights of the multi-branch feature map are calculated using an attention mechanism and weighted fusion is performed to obtain a fused feature map.

[0215] By using classification branches, the fused feature map is used to determine the benign or malignant nature of cells and the lesion subtype, and the target category is output.

[0216] The fused feature map is scored for the degree of cell abnormality through a regression branch to obtain a score value; the classification branch and the regression branch are implemented in parallel.

[0217] Based on the target category and the fused feature map, an interpretable heatmap is generated in real time using a gradient-weighted class activation mapping algorithm.

[0218] By integrating the target category, the score value, and the interpretability heatmap, the lesion identification result corresponding to the unstained biopsy cells is obtained.

[0219] In one possible implementation, the intelligent recognition module 204 is specifically used for:

[0220] Obtain the target category score corresponding to the target category;

[0221] Calculate the gradient matrix of the target category score with respect to the fused feature map;

[0222] Calculate the global mean of the gradient matrix corresponding to each channel in the fused feature map to obtain the channel weight data;

[0223] The channel weight data and the fused feature map are weighted and summed to generate an interpretable heatmap.

[0224] In one possible implementation, the lesion identification system 20 for unstained biopsy cells further includes an offline interpretation module 206, which is specifically used for:

[0225] Before the structured integration of 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 overlaid heatmap images, a preset offline interpretation mechanism is triggered in response to the user's questioning command regarding the lesion identification results.

[0226] 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, key features affecting the lesion identification result are extracted.

[0227] Based on the aforementioned key features, a visual explanation of the challenge instruction is generated;

[0228] The visual explanation content is displayed in association with the lesion identification result, so that the user can interactively verify the visual explanation content and the identification result;

[0229] The system receives user feedback instructions on the visual explanation content and associates these instructions with the lesion identification results to form a closed-loop interaction record.

[0230] In one 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 used for:

[0231] In response to a user's challenge to the lesion identification result, at least one of the LIME offline interpretation mechanism and the SHAP offline interpretation mechanism is triggered.

[0232] The lesion identification system 20 for unstained biopsy cells provided in this embodiment of the invention can realize the steps and effects of the lesion identification method for unstained biopsy cells in Embodiment 1. To avoid repetition, the invention will not repeat the steps.

[0233] The beneficial effects of this invention are reflected in:

[0234] First, by achieving temporal synchronization control of pulsed laser and high-speed imaging, along with image preprocessing, the technical bottleneck of low contrast in unstained cell imaging was overcome. This enabled clear imaging of rapidly flowing cells and standardized single-cell extraction, significantly improving sample processing efficiency and image quality stability. Second, based on a deep learning model, multi-scale features were extracted and lesion identification results were generated. This overcame the shortcomings of traditional pathological diagnosis, which relies on staining and is highly subjective. It enabled precise quantitative judgment of cell benignity / malignancy and lesion severity, improving the objectivity and accuracy of the identification results. Finally, combined with a structured report generation mechanism, an integrated diagnostic output containing key information was provided to clinicians, while supporting subsequent interactive interpretation. This not only met the core requirements of pathological diagnosis for efficiency and accuracy but also enhanced the traceability and clinical applicability of the results, greatly improving diagnostic efficacy and medical collaboration efficiency in unstained biopsy scenarios.

[0235] In the description of the embodiments of the present invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "center," "top," "bottom," "top," "bottom," "inner," "outer," "inner side," and "outer side," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. "Inner side" refers to the interior or enclosed area or space. "Outer perimeter" refers to the area surrounding a specific component or specific area.

[0236] In the description of embodiments of the present invention, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of that feature. In the description of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0237] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "joining," and "assembly" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0238] In the description of embodiments of the present invention, specific features, structures, materials or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0239] In the description of the embodiments of the present invention, it should be understood that "-" and "~" represent a range of two numerical values, and this 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.

[0240] In the description of embodiments of the present invention, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0241] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying lesions in unstained biopsy cells, characterized in that, include: By using time-synchronized control of pulsed laser and high-speed imaging, images of unstained biopsy cells flowing at high speed are acquired to obtain raw cell images. The original cell image is preprocessed to obtain a preprocessed cell image; Based on a preset target cutting strategy, the preprocessed cell image is subjected to multi-target cutting to obtain a standardized single-cell image. Multi-scale features of the standardized single-cell image are extracted based on a deep learning model to obtain the lesion identification results corresponding to the unstained biopsy cells. The lesion identification results include the benign or malignant judgment results of the unstained biopsy cells, lesion scores, interpretability heatmaps of the identification basis, and feature data that supports 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 overlaid heatmap images. The method involves timing-synchronized control of pulsed laser and high-speed imaging to image and acquire raw cell images of high-speed flowing unstained biopsy cells, including: A pulsed laser source and a high-speed imaging device are configured to introduce unstained biopsy cells into a microfluidic channel and control the flow rate of the unstained biopsy cells to obtain a flow cell sample to be imaged. The delay time is calculated by a synchronous triggering mechanism to control the timing synchronization of laser pulse emission and image acquisition, and the flowing cell sample entering the imaging acquisition window is photographed to obtain an initial cell image; If the signal-to-noise ratio of the initial cell image is lower than a preset threshold, the laser pulse energy or the gain parameter of the imaging device is dynamically adjusted to obtain the optimized parameters. By continuously photographing cells based on optimized parameters, high-contrast raw cell images without motion blur are obtained. The process involves calculating the delay time through a synchronous triggering mechanism to control the timing synchronization between laser pulse emission and image acquisition, and then capturing images of the flowing cell sample entering the imaging acquisition window to obtain an initial cell image. This includes: The delay time is calculated using the synchronous triggering mechanism according to the following formula: in, t d The time delay required for unstained biopsy cells to travel from the sensing area to the imaging acquisition window. L The distance from the sensing area to the imaging acquisition window. v f Real-time flow rate for unstained biopsy cells; The single-frame exposure time of the high-speed imaging device is controlled to meet the following anti-ghosting conditions: in, t e This refers to the single-frame exposure time of a high-speed imaging device. 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, so as to capture the unstained biopsy cells and obtain a high-contrast initial cell image without ghosting. The process of extracting multi-scale features from the standardized single-cell image based on a deep learning model to obtain the lesion identification result corresponding to the unstained biopsy cells includes: The standardized single-cell image is input into a multi-branch convolutional network, and cell features in the standardized single-cell image are extracted in parallel by convolutional kernels of different scales to obtain a multi-branch feature map. The channel weights of the multi-branch feature map are calculated using an attention mechanism and weighted fusion is performed to obtain a fused feature map. By using classification branches, the fused feature map is used to determine the benign or malignant nature of cells and the lesion subtype, and the target category is output. The fused feature map is scored for the degree of cell abnormality through a regression branch to obtain a score value; the classification branch and the regression branch are implemented in parallel. Based on the target category and the fused feature map, an interpretable heatmap is generated in real time using a gradient-weighted class activation mapping algorithm. By integrating the target category, the score value, and the interpretability heatmap, the lesion identification result corresponding to the unstained biopsy cells is obtained.

2. The method according to claim 1, characterized in that, The step of preprocessing the original cell image to obtain a preprocessed cell image includes: Gaussian filtering is applied to the original cell image to suppress image noise and obtain a smooth image. The smoothed image is subjected to local contrast equalization processing, and the distinction between cells and background in the smoothed image is enhanced by stretching gray values ​​to obtain an enhanced image; Phase consistency analysis and gradient calculation are performed on the enhanced image to extract cell edge features corresponding to low-contrast regions in the enhanced image, forming an edge image; Binarize and perform morphological operations on the edge image to remove noise and broken connection boundaries, thereby obtaining a connected component boundary image. The connected domain boundary image is used as the preprocessed cell image.

3. The method according to claim 1, characterized in that, The preprocessed cell image is subjected to multi-target segmentation based on a preset target segmentation strategy to obtain a standardized single-cell image, including: For each connected region in the preprocessed cell image, fit the minimum bounding rectangle to obtain the first candidate box set; Based on a preset contrast scoring function, the contrast score of the image region corresponding to each candidate box in the first candidate box set is calculated 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 the second candidate box set; Iterate through each candidate box in the second candidate box set, calculate the intersection-union ratio (IUU) of each candidate box with the other candidate boxes, and obtain the IUU set; For each candidate box in the second candidate box set, if the intersection-union ratio with other candidate boxes is greater than a preset threshold, only the candidate box with the higher contrast score is retained, and the other candidate boxes are suppressed to obtain the target candidate box set. The size of the single-cell regions corresponding to the target candidate box set is standardized to obtain a standardized single-cell image.

4. The method according to claim 1, characterized in that, The step of generating an interpretable heatmap in real time based on the target category and the fused feature map using a gradient-weighted class activation mapping algorithm includes: Obtain the target category score corresponding to the target category; Calculate the gradient matrix of the target category score with respect to the fused feature map; Calculate the global mean of the gradient matrix corresponding to each channel in the fused feature map to obtain the channel weight data; The channel weight data and the fused feature map are weighted and summed to generate an interpretable heatmap.

5. The method according to claim 1, characterized in that, Before the structured integration of 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 overlaid heatmap images, the following steps are also included: In response to a user's questioning command regarding the lesion identification results, a preset offline explanation mechanism is triggered; 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, key features affecting the lesion identification result are extracted. Based on the aforementioned key features, a visual explanation of the challenge instruction is generated; The visual explanation content is displayed in association with the lesion identification result, so that the user can interactively verify the visual explanation content and the identification result; The system receives user feedback instructions on the visual explanation content and associates these instructions with the lesion identification results to form a closed-loop interaction record.

6. The method according to claim 5, characterized in that, The offline interpretation mechanism includes the LIME offline interpretation mechanism and the SHAP offline interpretation mechanism. The step of triggering a preset offline interpretation mechanism in response to a user's challenge to the lesion identification result includes: In response to a user's challenge to the lesion identification result, at least one of the LIME offline interpretation mechanism and the SHAP offline interpretation mechanism is triggered.

7. A lesion identification system for unstained biopsy cells, characterized in that, include: The ultra-high-speed imaging module is used to acquire images of unstained biopsy cells flowing at high speed by controlling the timing of pulsed laser and high-speed imaging, thereby obtaining raw cell images. The image preprocessing module is used to perform image preprocessing on the original cell image to obtain a preprocessed cell image; The image segmentation module is used to perform multi-target segmentation on the preprocessed cell image based on a preset target segmentation strategy to obtain a standardized single-cell image. The intelligent recognition module is used to extract multi-scale features of the standardized single-cell image based on a deep learning model to obtain the lesion recognition result corresponding to the unstained biopsy cell. The lesion recognition result includes the benign or malignant judgment result of the unstained biopsy cell, the lesion score, the interpretability heatmap of the recognition basis, and feature data that supports offline interactive interpretation. The results 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 overlaid heatmap images. The ultra-high-speed imaging module is specifically used for: A pulsed laser source and a high-speed imaging device are configured to introduce unstained biopsy cells into a microfluidic channel and control the flow rate of the unstained biopsy cells to obtain a flow cell sample to be imaged. The delay time is calculated by a synchronous triggering mechanism to control the timing synchronization of laser pulse emission and image acquisition, and the flowing cell sample entering the imaging acquisition window is photographed to obtain an initial cell image; If the signal-to-noise ratio of the initial cell image is lower than a preset threshold, the laser pulse energy or the gain parameter of the imaging device is dynamically adjusted to obtain the optimized parameters. By continuously photographing cells based on optimized parameters, high-contrast raw cell images without motion blur are obtained. The ultra-high-speed imaging module is specifically used for: The delay time is calculated using the synchronous triggering mechanism according to the following formula: in, t d The time delay required for unstained biopsy cells to travel from the sensing area to the imaging acquisition window. L The distance from the sensing area to the imaging acquisition window. v f Real-time flow rate for unstained biopsy cells; The single-frame exposure time of the high-speed imaging device is controlled to meet the following anti-ghosting conditions: in, t e This refers to the single-frame exposure time of a high-speed imaging device. 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, so as to capture the unstained biopsy cells and obtain a high-contrast initial cell image without ghosting. The intelligent recognition module is specifically used for: The standardized single-cell image is input into a multi-branch convolutional network, and cell features in the standardized single-cell image are extracted in parallel by convolutional kernels of different scales to obtain a multi-branch feature map. The channel weights of the multi-branch feature map are calculated using an attention mechanism and weighted fusion is performed to obtain a fused feature map. By using classification branches, the fused feature map is used to determine the benign or malignant nature of cells and the lesion subtype, and the target category is output. The fused feature map is scored for the degree of cell abnormality through a regression branch to obtain a score value; the classification branch and the regression branch are implemented in parallel. Based on the target category and the fused feature map, an interpretable heatmap is generated in real time using a gradient-weighted class activation mapping algorithm. By integrating the target category, the score value, and the interpretability heatmap, the lesion identification result corresponding to the unstained biopsy cells is obtained.

Images