Detection methods and systems, detection devices, equipment and storage media

Abstract

A detection method, system, device, equipment, and storage medium are disclosed. In the detection method, an optical image of each tissue slice is segmented to obtain multiple segmented images, and a block label is obtained for each segmented image in the tissue slice. Based on the positional correspondence between the segmented images and measurement points, and the block labels of the segmented images, a classification label corresponding to each measurement point is obtained. Compared to the optical image of each tissue slice, the segmented images and measurement points are smaller and more refined, improving the precision of the detection model in classification judgment, thus improving the accuracy and precision of detection. Furthermore, by obtaining multiple measurement points and their signal light characteristic curves on the detection sample, the efficiency of obtaining feature data is improved. The detection model can then perform predictive processing based on the raw data to obtain the corresponding classification label without needing to fit and analyze the signal light characteristic curves, thereby improving detection efficiency.

Description

Technical Field

[0001] The present invention relates to the field of optical detection, and more particularly to a detection method and system, a detection device, an equipment, and a storage medium. Background Technology

[0002] The choice of surgical extent has always been a focus of surgical treatment. Excessive resection can result in large wounds and damage to vital organs, while incomplete resection can lead to high recurrence rates. A major challenge is how to remove as much of the lesion as possible while preserving as much normal tissue as possible.

[0003] Currently, surgical resection of multifocal tumors, such as melanoma, basal cell carcinoma of the eyelid, fibrosarcoma protuberans, and external Paget's disease of the breast, often employs either locally extended resection or Mohs microsurgery. Locally extended resection typically involves an extra 1-5 cm around the tumor, but the appropriate extent of this extension remains controversial. Mohs microsurgery requires layer-by-layer histopathological examination of the surgical margins, which is time-consuming, involves numerous pathological examination points (hundreds or even thousands), resulting in a massive workload for pathologists. Furthermore, the quality and accuracy of intraoperative frozen section examination are not ideal, and positive margins cannot be completely avoided. Therefore, excellent margin assessment techniques are crucial to achieving both appropriate resection extent and preservation of as much normal tissue as possible.

[0004] However, the efficiency, precision, and accuracy of current testing still need to be improved. Summary of the Invention

[0005] The problem solved by the embodiments of the present invention is to provide a detection method and system, detection device, equipment and storage medium to improve the accuracy, precision and efficiency of detection.

[0006] To address the aforementioned problems, this invention provides a detection method, comprising: obtaining feature data of a test sample, the feature data including multiple measurement points located on the test sample and a feature curve of signal light generated by each measurement point; obtaining optical imaging data of tissue slices of the test sample at intervals of several locations; performing image segmentation on the optical image of each tissue slice based on the optical imaging data to obtain multiple segmented images; obtaining a block label for each segmented image in the tissue slice; and obtaining a classification label corresponding to each measurement point based on the positional correspondence between the segmented images and the measurement points, and the block label of the segmented images, wherein the measurement point and the corresponding classification label are FDU-PA230002-CN.

[0007] The signal light feature curves are used to form the first training dataset; an initial detection model is constructed; the initial detection model is trained using the first training dataset to obtain a trained detection model; the trained detection model is used to predict the measurement points and the corresponding signal light feature curves to obtain the corresponding classification labels.

[0008] Accordingly, this embodiment of the invention also provides a detection device, which is used to obtain feature data of the detection sample in the detection method described in this embodiment of the invention; the detection device includes: a light source for providing detection light; an optical fiber detection end connected to the light source, the optical fiber detection end being used as a detection end for detecting measurement points on the detection sample; wherein, the detection light is irradiated onto the measurement point through the optical fiber detection end, causing the detection sample at the measurement point location to generate signal light; and a detection module connected to the optical fiber detection end, the detection module being used to detect the signal light generated by the detection sample at the measurement point location and obtain the signal light feature curve of each measurement point.

[0009] Accordingly, this embodiment of the invention also provides a detection system, comprising: a data acquisition module for acquiring feature data of a detection sample, the feature data including multiple measurement points located on the detection sample and feature curves of signal light generated by each measurement point; an imaging module for acquiring optical imaging data of tissue slices of the detection sample at several intervals; an image segmentation module for performing image segmentation on the optical image of each tissue slice based on the optical imaging data to obtain multiple segmented images; a block label acquisition module for acquiring a block label for each segmented image in the tissue slice; a classification label acquisition module for acquiring a classification label corresponding to the measurement point based on the positional correspondence between the segmented image and the measurement point, and the block label of the segmented image, wherein the measurement point, the corresponding classification label, and the signal light feature curve constitute a first training dataset; a model building module for constructing an initial detection model; a training module for training the initial detection model using the first training dataset to obtain a trained detection model; and a prediction module for using the trained detection model to predict the measurement point and the corresponding signal light feature curve to obtain the corresponding classification label.

[0010] Accordingly, embodiments of the present invention also provide a device, including at least one memory and at least one processor, wherein the memory stores one or more computer instructions, and the one or more computer instructions are executed by the processor to implement the detection method provided in the embodiments of the present invention.

[0011] Accordingly, embodiments of the present invention also provide a storage medium that stores one or more computer instructions, which are executed by the processor to implement the detection method provided in the embodiments of the present invention.

[0012] FDU-PA230002-CN

[0013] Compared with the prior art, the technical solution of the embodiments of the present invention has the following advantages:

[0014] In the detection method provided by this invention, multiple segmented images are obtained by image segmentation of the optical image of each tissue slice, and a block label is obtained for each segmented image in the tissue slice. Based on the positional correspondence between the segmented images and the measurement points, and the block label of the segmented images, a classification label corresponding to the measurement points is obtained. Compared with the optical image of each tissue slice, the segmented images and measurement points are smaller in size and have higher precision. Therefore, in the step of training the initial detection model using the first training dataset to obtain the trained detection model, the precision of the detection model in classification judgment can be improved, which is conducive to improving the accuracy and precision of detection. It is especially suitable for the detection and judgment of multifocal diseases. Moreover, by obtaining multiple measurement points on the detection sample and the signal light feature curves of the measurement points, it is also conducive to improving the efficiency of obtaining feature data. Furthermore, the detection model can be used to perform prediction processing based on the original data (i.e., measurement points and corresponding signal light feature curves) to obtain the corresponding classification label without the need for fitting analysis of the signal light feature curves, which is conducive to improving detection efficiency.

[0015] The detection device provided in this embodiment of the invention includes an optical fiber detection end connected to the light source. The optical fiber detection end is used as a detection end for detecting measurement points on the test sample. The detection light is irradiated onto the measurement point through the optical fiber detection end, causing the test sample at the measurement point to generate signal light. This provides a basis for the detection module to detect the signal light and obtain the signal light characteristic curve of the signal point. Compared with staining and optical detection of tissue sections, measuring the measurement point through the optical fiber detection end eliminates the need for sampling, fixing, and staining of the test sample. Instead, it directly detects each measurement point on the surface of the test sample to obtain the signal light characteristic curve, simplifying the process, improving the convenience of detection, and reducing the time spent on each measurement point, thus significantly increasing the detection speed. Furthermore, the optical fiber detection end is suitable for detecting samples such as cells, tissues, skin on the human body surface, and internal devices, thus expanding the applicability of the test samples. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of an embodiment of the detection method of the present invention;

[0017] Figure 2 A schematic diagram of a detection sample and the obtained feature data is shown;

[0018] Figure 3 A schematic diagram is shown of slicing the test sample at several intervals.

[0019] Figure 4 It shows Figure 1 A flowchart illustrating an embodiment of step S4;

[0020] FDU-PA230002-CN

[0021] Figure 5 Two examples of optical images of tissue sections are shown;

[0022] Figure 6 The diagrams illustrate how predictive processing is performed on the segmented image corresponding to the optical image to obtain the block labels of the corresponding segmented image;

[0023] Figure 7 This is a schematic diagram of the structure of an embodiment of the detection device of the present invention;

[0024] Figure 8 This is a functional block diagram of an embodiment of the detection model construction system of the present invention;

[0025] Figure 9 This is a hardware structure diagram of an embodiment of the device provided by the present invention. Detailed Implementation

[0026] As the background technology indicates, current detection efficiency, precision, and accuracy still need improvement. The following analysis, combining several detection methods, explains why these aspects still require further improvement.

[0027] One detection method involves inserting an optical fiber and using an endoscope to reach the lesion site for image acquisition. Alternatively, fluorescence lifetime imaging can be performed on tissue sections. This method determines surgical margins based on the average fluorescence lifetime value of the images. The average value is compared to a threshold; values ​​higher or lower than the threshold indicate malignancy or normal / benign conditions. However, this method requires fluorescence microscopy, resulting in high system costs, slow processing speed for large areas, and insufficient accuracy.

[0028] Another detection method involves examining autofluorescence (FLIM) images of a large number of samples and then using an unsupervised clustering algorithm to predict cancer development. This method is effective for predicting cancer development in single-species cells, but it is not suitable for complex tissue analysis or in vivo detection.

[0029] Another detection method is applied to Mohs microsurgery, which aims to remove all tumors while preserving as much normal tissue as possible. This method involves first removing the visible tumor, then removing a thin layer of tissue at the margins, mapping the location, and then sectioning and staining each section of this thin layer for microscopic examination. If all tissue sections are negative under the microscope, the surgery is complete. If a positive margin is observed, the procedure continues, removing thin layers of tissue layer by layer in the area of ​​remaining cancerous tissue, repeating the section staining and pathological examination until complete removal. However, this procedure is very time-consuming, potentially taking hours or even days. Pathological examination involves numerous points, typically requiring the examination of thousands or even tens of thousands of tissue sections. Excessive spacing between sections can also lead to missed residual tumor, resulting in low detection efficiency and accuracy.

[0030] Another detection method is to use fiber-optic point scanning FLIM (Fluorescence Lifetime FDU-PA230002-CN).

[0031] Intraoperative FLIM (Fluorescence Lifetime Imaging) system integration into the da Vinci Surgical System enables intraoperative guidance. This method first acquires data points on the patient's tissue and then uses supervised machine learning algorithms to predict the probability of cancer. However, the labels on the training set are derived from the experience of physicians and pathology teams, and are only suitable for labeling large lesions. For multifocal cancerous tissues, the labels may be too coarse or incorrect, or require a very large amount of pathological work. Furthermore, it is subjective and subject to sampling errors. Even small-sized areas with incorrect labels can affect the overall accuracy of the model.

[0032] In summary, the efficiency, precision, and accuracy of current testing still need to be improved.

[0033] To address the aforementioned technical problem, embodiments of the present invention provide a detection method. Figure 1 This is a flowchart illustrating an embodiment of the detection method of the present invention.

[0034] like Figure 1 As shown, in this embodiment, the detection method includes the following basic steps:

[0035] Step S1: Obtain feature data of the test sample, the feature data including multiple measurement points located on the test sample and the signal light feature curve of each measurement point;

[0036] Step S2: Obtain optical imaging data of tissue sections of the test sample at several intervals;

[0037] Step S3: Based on the optical imaging data, perform image segmentation on the optical image of each tissue slice to obtain multiple segmented images;

[0038] Step S4: Obtain the block label of each segmented image in the tissue slice;

[0039] Step S5: Based on the positional correspondence between the segmented image and the measurement point, and the block label of the segmented image, obtain the classification label corresponding to the measurement point. The measurement point, the corresponding classification label, and the signal light feature curve are used to form the first training dataset.

[0040] Step S6: Construct the initial detection model;

[0041] Step S7: Train the initial detection model using the first training dataset to obtain a trained detection model;

[0042] Step S8: Using the trained detection model, predict the measurement points and the corresponding signal light feature curves to obtain the corresponding classification labels.

[0043] In the method for constructing the detection model provided in this embodiment of the invention, multiple segmented images are obtained by performing image segmentation on the optical image of each tissue slice, and a block label is obtained for each segmented image in the tissue slice. Based on the positional correspondence between the segmented images and the measurement points, and the block labels of the segmented images, a classification label corresponding to the measurement points is obtained, and the FDU-PA230002-CN of each tissue slice is also obtained.

[0044] Compared to optical images, segmented images and measurement points are smaller and more precise. This allows for improved precision in classification during the initial training process using the first training dataset to train the detection model, thus enhancing accuracy and precision, particularly for multifocal diseases. Furthermore, obtaining multiple measurement points and their signal light characteristic curves on the sample improves the efficiency of feature data acquisition. The detection model can then be used to predict and obtain classification labels based on the original data (i.e., measurement points and their corresponding signal light characteristic curves) without requiring fitting analysis of the signal light characteristic curves, further improving detection efficiency.

[0045] To make the above-mentioned objects, features and advantages of the embodiments of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] refer to Figure 1 Step S1: Obtain feature data of the test sample, the feature data including multiple measurement points located on the test sample and the feature curve of the signal light generated by each measurement point.

[0047] The feature data of the detection samples are obtained so that the classification label corresponding to the measurement point can be obtained based on the positional correspondence between the segmented image and the measurement point and the block label of the segmented image. This allows the measurement point, the corresponding classification label, and the signal light feature curve to form the first training dataset, which can then be used to train the initial detection model.

[0048] The test sample is the sample to be tested in order to obtain a diagnostic result.

[0049] The test sample is equipped with multiple measurement points, each of which serves as a measurement location. This facilitates the detection of the signal light characteristic curve generated at each measurement location. Compared to staining and optical detection of tissue sections, this embodiment obtains the signal light characteristic curve by detecting each measurement point, eliminating the need for sampling, fixing, and staining of the test sample. Instead, the signal light characteristic curve is obtained directly at each measurement point on the surface of the test sample. This simplifies the process, improves the convenience of detection, and reduces the time spent on each measurement point, thus significantly increasing the detection speed.

[0050] Reference Figure 2 , Figure 2 A schematic diagram of a detection sample and the obtained feature data is shown. As an example, Figure 2 (a) shows a schematic diagram of a test sample 10 and multiple measurement points 11 on the test sample 10.

[0051] As an example, the test sample is used as a test sample for extramammary Paget's disease. In other embodiments, the test sample may also be a test sample that can reflect other cancer types, such as FDU-PA230002-CN.

[0052] It can be extended to other multifocal skin cancers, including basal cell carcinoma, squamous cell carcinoma, melanoma, and dermatofibrosarcoma. The detection of these cancers can be performed directly at the lesion site using a fiber optic detection device.

[0053] In other embodiments, the test samples can also be those capable of reflecting other types of cancer, including gastrointestinal cancers (oral cancer, esophageal cancer, stomach cancer, intestinal cancer, etc.), urinary system cancers (urethral cancer, bladder cancer, ureteral cancer, kidney cancer, etc.), respiratory system cancers (tracheal cancer, bronchial cancer, lung cancer, etc.), and reproductive system cancers (cervical cancer, vaginal cancer, vulvar cancer, fallopian tube cancer, ovarian cancer, etc.). These cancers can be detected using an endoscope combined with a fiber optic detection end.

[0054] In this embodiment, the step of obtaining feature data of the test sample includes: setting the measurement point on the test sample based on the detection requirements; detecting the measurement point to obtain feature data of the measurement point.

[0055] It should be noted that, Figure 2 The size, location, number, density, and arrangement of the measurement points shown in (a) are merely examples. In actual implementation, the size, location, number, density, and arrangement of the measurement points are determined based on the actual detection needs.

[0056] The characteristic curve of the signal light generated by the detection light on the test sample at each measurement point is obtained. The characteristic curve of the signal light can reflect the characteristics of the signal light generated by different substances in the test sample when excited by the detection light. Then, the measurement information and state of different substances in the test sample can be obtained from the characteristics of the signal light, which makes it easier to predict the block label of the test sample at different measurement points.

[0057] Specifically, in actual testing, detection light (e.g., laser) is irradiated onto the measurement point of the test sample through the fiber optic detection end, so that the test sample at the measurement point location generates signal light, thereby enabling the detection of the signal light generated by the test sample at the measurement point location, and thus obtaining the characteristic curve of the signal light.

[0058] In this embodiment, the step of obtaining feature data of the test sample involves obtaining multi-channel feature data of the test sample, where each channel acquires a different wavelength range of optical signal. By obtaining multi-channel feature data of the test sample, the received signal light can cover a certain spectral width, thereby increasing the dimensionality and comprehensiveness of the detection, and correspondingly improving the accuracy of the detection.

[0059] As an example, in this embodiment, characteristic curves of signal light generated by two coenzymes, collagen, and protoporphyrin IX (PpIX) in the test sample are obtained. In this embodiment, the two coenzymes are nicotinamide adenine dinucleotide (phosphate) (NAD(P)H, reduced form of nicotinamide-adenine dinucleotide (phosphate)) and flavin adenine dinucleotide (FAD).

[0060] Specifically, in this embodiment, for NAD(P)H, a picosecond laser of 330nm to 410nm (e.g., a 405nm laser) or a femtosecond laser of 660nm to 820nm can be provided as the detection light to receive fluorescence in the signal light band of 420nm to 540nm. For FAD, a picosecond laser of 350nm to 490nm (e.g., a 488nm laser) or a femtosecond laser of 700nm to 980nm can be provided as the detection light to receive fluorescence in the signal light band of 500nm to 620nm. For collagen, a picosecond laser of 380nm to 520nm or a femtosecond laser of 760nm to 1040nm can be provided as the detection light. The received signal light band can be selected in the range of 350-600nm depending on the wavelength of the detection light, and can receive the generated fluorescence or the second harmonic light signal generated by femtosecond light excitation. For PpIX, picosecond lasers of 390nm to 440nm or femtosecond lasers of 780nm to 880nm can be provided as detection light, and the wavelength of the received signal light is fluorescence of 550nm to 680nm.

[0061] Accordingly, in this embodiment, the signal light includes fluorescence and second harmonic light signals. In other embodiments, based on the type of substance in the actual detection sample and the type of detection light, the signal light may also include other types of light signals. Correspondingly, in other embodiments, the number of channels of the obtained feature data may also be other numbers.

[0062] In other words, in this embodiment, in the step of obtaining the characteristic data of the test sample, multiple channels respectively acquire the fluorescence generated by NAD(P)H, the fluorescence generated by FAD, the fluorescence or second harmonic light signal generated by collagen, and the fluorescence generated by PpIX, and obtain the characteristic curve of the signal light corresponding to the fluorescence or second harmonic light signal.

[0063] Accordingly, in this embodiment, one channel is used to acquire the characteristic curve of the second harmonic optical signal generated by the test sample. The second harmonic optical signal can reflect the collagen content in the test sample, and the collagen content can reflect the lesion status of the test sample, thereby obtaining more dimensions of information for assessing the lesion status, and thus improving the comprehensiveness and accuracy of the detection.

[0064] As an example, the signal light characteristic curve includes a fluorescence lifetime decay curve or a spectral curve.

[0065] The fluorescence lifetime decay curve contains the intensity ratio between different channels, reflecting the content ratio of the substances corresponding to different channels in the detected sample. It also includes fluorescence lifetime information and fluorescence spectral intensity. Therefore, by obtaining the fluorescence lifetime decay curve at the measurement point, multi-dimensional characteristic information of the signal light can be obtained, thus improving the accuracy of the judgment. Furthermore, by obtaining the fluorescence lifetime decay curve, exponential fitting analysis is not required to obtain fluorescence lifetime information; instead, it can be directly obtained from the fluorescence... (FDU-PA230002-CN)

[0066] Fluorescence lifetime information is obtained from the optical lifetime decay curve. Subsequently, the raw data of the fluorescence lifetime decay curve can be directly used to form the first training dataset for training the initial detection model, as well as the input for the detection model, which is beneficial to further improve the detection speed.

[0067] The spectral curve can reflect the intensity of the received signal light, and thus reflect the content of substances contained in the sample at the measurement point. Moreover, the spectral curve is obtained quickly, which helps to improve the detection speed.

[0068] As an example, Figure 2 (b) and Figure 2 (c) Schematic diagrams showing the spectral curves obtained at measurement point 12 and the fluorescence lifetime decay curves of multiple channels obtained at measurement point 13 are shown respectively. As an example, in this embodiment, four channels (e.g.) are obtained. Figure 2 (c) The fluorescence lifetime decay curves corresponding to the signal lights of CH1, CH2, CH3, and CH4 shown are illustrated as an example. In other embodiments, the number of fluorescence lifetime decay curves obtained varies depending on the actual number of channels.

[0069] refer to Figure 2 Step S2: Obtain optical imaging data of tissue sections of the test sample at several intervals.

[0070] Optical imaging data of tissue slices at several intervals of the test sample are obtained so that the optical images of the tissue slices can be segmented to obtain multiple segmented images based on the optical imaging data, and the block labels of the segmented images can be obtained. Then, the classification label corresponding to the measurement point can be obtained by using the positional correspondence between the segmented images and the measurement points.

[0071] Reference Figure 3 This illustrates a schematic diagram of slicing the test sample at several intervals. For example... Figure 3 As shown, as an example, optical imaging data of tissue sections of the test sample at 8 intervals were obtained.

[0072] In other embodiments, the number, location, and spacing of tissue slices can be flexibly adjusted based on actual detection needs.

[0073] In this embodiment, fluorescence lifetime imaging data is used as an example for illustration. In other embodiments, the optical imaging data may also be stimulated Raman scattering microscopy (SRS) data, multi-photon excitation (MPE) data, autofluorescence imaging data, histochemical staining image data, or second harmonic imaging data.

[0074] Continue to refer to Figure 1 Step S3: Based on the optical imaging data, slice each of the tissues. (FDU-PA230002-CN)

[0075] The optical image of the film is segmented to obtain multiple segmented images.

[0076] Image segmentation is performed on the optical image of each tissue slice to obtain multiple segmented images, so as to obtain the block label of each segmented image in the tissue slice. This allows the classification label of the measurement point to be obtained based on the positional correspondence between the segmented images and the measurement points.

[0077] Moreover, compared with the optical image of each tissue slice, the segmented image and measurement points are smaller in size and more precise. This is beneficial for improving the precision of the detection model in classification judgment during the subsequent step of training the initial detection model using the first training dataset to obtain a trained detection model, thereby improving the accuracy and precision of detection.

[0078] In this embodiment, the step of performing image segmentation on the optical image of each tissue slice to obtain multiple segmented images based on the optical imaging data includes: performing image segmentation on the optical image of each tissue slice based on detection requirements to obtain multiple segmented images.

[0079] Specifically, based on the detection requirements, the number of image segments to be performed on the optical image of each tissue slice is set to obtain segmented images of corresponding sizes.

[0080] It should be noted that, in this embodiment, in the step of obtaining the feature data of the test sample, the measurement point is located on the surface of the test sample, and the characteristic curve of the signal light generated by the surface tissue of the test sample at the measurement point is measured accordingly. Accordingly, in this embodiment, the segmented image near the surface of the tissue slice in the optical image of each tissue slice corresponds to the measurement point.

[0081] In practice, OpenCV (Open Source Computer Vision Library) can be used to segment the optical image of each tissue slice to obtain multiple segmented images. More specifically, the optical image can be uniformly cropped to obtain multiple segmented images.

[0082] Continue to refer to Figure 1 Step S4: Obtain the block labels for each segmented image in the tissue slice.

[0083] Obtain the block labels of the segmented images in the tissue slices so that the classification labels corresponding to the measurement points can be obtained subsequently based on the positional correspondence between the segmented images and the measurement points, as well as the block labels of the segmented images.

[0084] In this embodiment, the block labels of the segmented image are described as normal or malignant. In other embodiments, the block labels of the segmented image may also include benign. Alternatively, in other embodiments, based on actual detection needs and the classification of detection results, the block labels of the segmented image may include other types.

[0085] FDU-PA230002-CN

[0086] refer to Figure 4 , showed Figure 1 A flowchart illustrating an embodiment of step S4. The following is in conjunction with... Figure 4 The specific steps S4 for obtaining the block labels of each segmented image in the tissue slice in this embodiment will be described in detail.

[0087] like Figure 4As shown, step S41 is performed to obtain the pathological diagnosis results of a partial number of tissue sections.

[0088] Obtain the pathological diagnosis results of a portion of the tissue sections so that, based on the pathological diagnosis results, the block labels of each segmented image in the portion of the tissue sections can be obtained.

[0089] In this embodiment, only a portion of the tissue sections are obtained for pathological diagnosis, thereby reducing the number of tissue sections required for pathological diagnosis and thus improving the detection rate.

[0090] Specifically, in this embodiment, the step of obtaining the pathological diagnosis results of a portion of the tissue sections may include: staining the portion of the tissue sections and performing pathological examination, with a pathology expert making a judgment to obtain the pathological diagnosis results.

[0091] In practice, the tissue sections of the aforementioned number can be stained with hematoxylin and eosin (H&E), followed by pathological examination and judgment by a pathologist to obtain a pathological diagnosis.

[0092] like Figure 4 As shown, step S42: Based on the pathological diagnosis results, labeling is performed on each segmented image in the partial number of tissue slices to obtain the block labels of the segmented images. The segmented images and corresponding block labels in the partial number of tissue slices are used to form the second training dataset.

[0093] The second training dataset is used for subsequent training of the classification model.

[0094] Specifically, based on the pathological diagnosis results of the tissue section, each segmented image of the current tissue section is labeled with a corresponding block label. For example, if the pathological diagnosis result of the current tissue section is malignant, then each segmented image of the current tissue section is labeled as malignant; if the pathological diagnosis result of the current tissue section is normal, then each segmented image of the current tissue section is labeled as normal. In other words, the segmentation label corresponding to each segmented image of the current tissue section is consistent with the pathological diagnosis result of the tissue section.

[0095] Reference Figure 5 Two examples of optical images 100 of tissue sections are shown. Optical image 110 shows a tissue section with a pathological diagnosis of malignancy, and the corresponding segmentation image 130 is also labeled as malignant; optical image 120 shows a tissue section with a pathological diagnosis of normal, and the corresponding segmentation image 130 is also labeled as normal.

[0096] like Figure 4As shown, step S43: Construct a classification model. Construct a classification model so that it can be used subsequently for the classification of FDU-PA230002-CN.

[0097] The classification model is trained using two training datasets.

[0098] As an example, the classification model includes a Residual Neural Network (ResNet) model. Residual Neural Networks employ "shortcut connections" to address degradation, significantly reducing the training difficulties associated with excessively deep neural networks. As an example, the ResNet model is a ResNet 50 model.

[0099] In other embodiments, the classification model may be other types of classification models based on actual needs.

[0100] It should be noted that continued reference is necessary. Figure 4 As an example, step S4, which obtains the block label of each segmented image in the tissue slice, further includes: after obtaining the block label of each segmented image in the partial number of tissue slices based on the pathological diagnosis result, and before training the classification model using the second training dataset, step S46: performing image recognition processing on the segmented images in the second training dataset to obtain image recognition results; step S47: removing segmented images with a background area ratio greater than a preset value from the second training dataset based on the image recognition results.

[0101] The segmented images in the second training dataset are processed by image recognition. Based on the image recognition results, the segmented images with a background area ratio greater than a preset value are removed from the second training dataset. This prevents the segmented images with an excessively large background area ratio from being mixed in the second training dataset, which helps to prevent interference with the training of subsequent classification models and reduces the probability of false positives or false negatives predicted by the subsequently trained classification models.

[0102] like Figure 4 As shown, step S44 is executed: the classification model is trained using the second training dataset to obtain a trained classification model.

[0103] The classification model is trained using the second training dataset, enabling it to analyze and learn the relationship between the segmented images and the block labels. Subsequently, the classification model can be used to predict the segmented images of the remaining tissue slices to obtain the corresponding block labels.

[0104] Reference Figure 5In this embodiment, a second training dataset is constructed using only tissue sections 110 with pathological diagnoses of malignancy and 120 with pathological diagnoses of normal tissue sections, as an example to illustrate the training process for the classification model. In other embodiments, based on actual detection needs, a second training dataset can also be constructed using tissue sections with pathological diagnoses of malignancy, normal tissue sections, and benign tissue sections, as well as pathological diagnoses of benign tissue sections, to train the classification model. FDU-PA230002-CN

[0105] like Figure 4 As shown, step S45 is executed: after obtaining the trained classification model, the classification model is used to predict the segmented images in the remaining number of tissue slices to obtain the block labels of the segmented images in the remaining number of tissue slices.

[0106] A classification model is used to predict the segmentation images of the remaining tissue slices to obtain corresponding block labels. Compared with the optical images of tissue slices, the segmentation images are smaller in size, which is conducive to obtaining more refined labels. Moreover, using a classification model to predict the segmentation images of the remaining tissue slices not only improves the accuracy of obtaining block labels for the segmentation images, but also saves a lot of annotation time and increases the speed of obtaining annotations for the segmentation images.

[0107] like Figure 6 The diagram illustrates the prediction processing performed on the segmented images corresponding to optical images 100a, 100b, and 100c to obtain the block labels of the corresponding segmented image 130. Optical image 100a represents a completely malignant tissue section, optical image 100b represents a tissue section of a mixed region, and optical image 100c represents a completely normal tissue section. Figure 6 In the segmented image, the color intensity of the segmented image is used to distinguish between malignant and normal conditions.

[0108] Among them, "all benign," "all malignant," or "all normal" refer to the fact that the block labels of all segmented images of the current tissue slice are benign, malignant, or normal, respectively.

[0109] It should be noted that this embodiment uses the example of obtaining block labels for segmented images in the remaining number of tissue slices by training a classification model. However, the methods for obtaining block labels for segmented images in the remaining number of tissue slices are not limited to this.

[0110] In other embodiments, the step of obtaining block labels for segmented images in the remaining number of tissue sections may further include: labeling each segmented image in the remaining number of tissue sections with a corresponding block label based on the pathological diagnosis results of the remaining number of tissue sections. Specifically, the segmentation label of each segmented image is consistent with the pathological diagnosis result of the corresponding tissue section.

[0111] It should also be noted that, in some other embodiments, the step of obtaining the block label of each segmented image in the tissue slice may further include: after obtaining the block labels of the segmented images in the remaining number of tissue slices, performing cleaning and correction processing on the second training dataset; after performing cleaning and correction processing on the second training dataset, retraining the classification model using the second training dataset; and using the retrained classification model to perform prediction processing on the segmented images to obtain the block label of each segmented image.

[0112] The second training dataset was cleaned and corrected, and then the classification model was retrained. (FDU-PA230002-CN)

[0113] This improves the accuracy of training samples, minimizes the interference of erroneous samples on model training, and enhances the reliability and robustness of the classification model obtained through training, thereby increasing the accuracy of the block labels obtained during prediction processing.

[0114] In practice, the second training dataset can be cleaned and corrected based on cross-validation combined with confidence learning, fidelity-based weighted learning, or sampling detection.

[0115] In some embodiments, the step of cleaning and correcting the second training dataset based on cross-validation combined with confidence learning may include: obtaining the joint distribution of the pathological diagnosis results and the block labels of the segmented images in the second training dataset; and based on the joint distribution, deleting samples from the second training dataset whose probability is lower than the current class mean and whose probability of another class is higher than the other class mean, according to the mean of the probability of each class.

[0116] By obtaining the joint distribution of the pathological diagnosis results and the block labels of the segmented images in the second training dataset, and based on the joint distribution, according to the mean of the probability of each class, samples with a probability lower than the current class mean and a probability of another class higher than the other class mean are deleted from the second training dataset, thereby achieving the purpose of data cleaning, and the class weights in the second training dataset are readjusted, and then the classification model is retrained to improve the accuracy of the classification model.

[0117] For example, in specific implementation, based on the joint distribution of pathological diagnosis results and block labels, and according to the mean of the normal and malignant class probabilities, samples with a probability lower than the mean of the normal class and a probability higher than the mean of the malignant class are deleted from the second training dataset.

[0118] In some embodiments, the step of cleaning and correcting the second training dataset based on fidelity-weighted learning may include: using the classification model as a first classification model; training a student network using the optical images of the tissue slices and the corresponding pathological diagnosis results; training a teacher network using the segmented images and the corresponding block labels; and generating soft labels for all segmented images and the corresponding block labels using the teacher network to obtain a soft dataset.

[0119] Accordingly, the student network is subjected to adversarial training using the soft dataset to obtain a second classification model; and the second classification model is used to predict all segmented images to obtain the block labels of segmented images in all tissue slices.

[0120] By training both the student and teacher networks, obtaining a soft dataset using the teacher network, and then using the soft dataset to perform adversarial training on the student network, an FDU-PA230002-CN model that approximates the first classification model is obtained.

[0121] The second classification model can retain the noise data in the optical image and the corresponding pathological diagnosis results during adversarial training, and can perform fidelity-preserving weighted learning, thereby improving the accuracy and predictive authenticity of the second classification model.

[0122] In some embodiments, the step of cleaning and correcting the second training dataset based on sampling detection may include: dividing the second training dataset into multiple sub-data sets, and using a classification model to perform sampling prediction on each sub-data set; after performing sampling prediction on each sub-data set, cleaning the second training dataset to re-label or delete data whose prediction results are inconsistent with the classification results obtained based on pathological diagnosis results; retraining the classification model using the cleaned second training dataset; and after retraining the classification model, using the classification model to predict the segmented images in the remaining number of tissue slices to obtain the block labels of the segmented images in the remaining number of tissue slices.

[0123] By cleaning the second training dataset, data whose prediction results are inconsistent with the block labels obtained based on pathological diagnosis results are re-labeled or deleted from the second training dataset, and the classification model is retrained, thereby preventing erroneous sample data from interfering with model training and improving the accuracy of classification model prediction.

[0124] It should be noted that in this embodiment, after obtaining the block labels of the segmented image corresponding to the optical image of each tissue slice, a positive rate can be given for a large range of images based on the proportion of different block labels in the segmented image of each tissue slice. As an example, the accuracy of the finally trained classification model can reach 95%, and it can give a positive rate for large range of images for slices of malignant, mixed, and normal regions. The positive rate for malignant regions is close to 100%, the positive rate for normal regions is 0%, while the positive rate for mixed regions varies depending on the specific cancer type. Taking extramammary Paget's disease as an example, the positive rate for mixed regions measured in this embodiment is approximately 30% to 60%. Therefore, in this embodiment, the prediction accuracy of the classification model is high.

[0125] Continue to refer to Figure 1 Step S5: Based on the positional correspondence between the segmented image and the measurement point, and the block label of the segmented image, obtain the classification label corresponding to the measurement point. The measurement point, the corresponding classification label, and the signal light feature curve are used to form the first training dataset.

[0126] Obtain the classification labels corresponding to the measurement points in order to construct the first training dataset, which will then facilitate the subsequent training of the initial detection model using the first training dataset.

[0127] Compared to optical images of each tissue slice, segmented images and measurement points are smaller in size and have higher precision. (FDU-PA230002-CN)

[0128] This allows for higher precision in classification judgment during the initial detection model training process using the first training dataset, thereby improving the accuracy and precision of detection. This is particularly suitable for the detection and judgment of multifocal diseases. Furthermore, by obtaining multiple measurement points and their signal light feature curves on the detection sample, the efficiency of obtaining feature data is improved. The detection model can also obtain the corresponding classification label based on the original data (i.e., measurement points and corresponding signal light feature curves) without the need for fitting analysis of the signal light feature curves, which is beneficial for improving detection efficiency.

[0129] In this embodiment, the segmented image near the surface of the tissue slice in the optical image of each tissue slice corresponds to the measurement point. Accordingly, the block label of the segmented image near the surface of the tissue slice is used as the classification label of the measurement point at the corresponding position.

[0130] Continue to refer to Figure 1 Step S6: Construct an initial detection model so that the initial detection model can be trained using the first training dataset to obtain a trained detection model.

[0131] As an example, the steps for constructing the initial detection model include: constructing the initial detection model based on logistic regression, decision tree, random forest, or support vector machine (SVM).

[0132] In this embodiment, the initial detection model built based on support vector machines is used as an example for illustration. Support vector machines are a novel few-shot learning method with a solid theoretical foundation. They do not involve probability measures or the law of large numbers, and their final decision function is determined by only a few support vectors. The computational complexity depends on the number of support vectors, rather than the dimension of the sample space, which avoids the "curse of dimensionality" in a sense.

[0133] In other embodiments, the initial detection model may be constructed based on other suitable classification models.

[0134] It should be noted that, in this embodiment, the method further includes: after obtaining the classification label corresponding to the measurement point based on the positional correspondence between the segmented image and the measurement point, and the block label of the segmented image, and before training the initial detection model using the first training dataset, step S9 is performed: data augmentation processing is performed on the first training dataset.

[0135] The first training dataset is augmented to increase the sample size.

[0136] As an example, the steps for data augmentation of the first training dataset include any one or more of the following: linearly superimposing several signal light characteristic curves of the same category using a random scaling factor that sums to 1; randomly shifting FDU-PA230002-CN along the horizontal axis within a limited range of spectral line shift fluctuations.

[0137] Signal light characteristic curve; add random Gaussian white noise of varying intensity to the signal light characteristic curve.

[0138] Continue to refer to Figure 1Step S7: Use the first training dataset to train the initial detection model to obtain a trained detection model.

[0139] In this embodiment, multiple segmented images are obtained by image segmentation of the optical image of each tissue slice, and block labels are obtained for each segmented image in the tissue slice. Based on the positional correspondence between the segmented images and the measurement points, and the block labels of the segmented images, classification labels corresponding to the measurement points are obtained. Compared with the optical image of each tissue slice, the segmented images and measurement points are smaller in size and have higher precision. Therefore, in the step of training the initial detection model using the first training dataset to obtain the trained detection model, the precision of the detection model in classification judgment can be improved, which is conducive to improving the accuracy and precision of detection, especially suitable for the detection and judgment of multifocal diseases. Moreover, by obtaining multiple measurement points on the detection sample and the signal light feature curves of the measurement points, it is also beneficial to improve the efficiency of obtaining feature data. Furthermore, the detection model can be used to perform prediction processing based on the original data (i.e., measurement points and corresponding signal light feature curves) to obtain the corresponding classification labels without the need for fitting analysis of the signal light feature curves, which is beneficial to improving detection efficiency.

[0140] Continue to refer to Figure 1 Step S8: Using the trained detection model, predict the feature curves of the measurement point and the signal light to obtain the corresponding classification label.

[0141] By utilizing a trained detection model, the characteristic curves of the measurement points and the signal light are predicted to obtain corresponding classification labels, thereby improving the accuracy and precision of detection. This is particularly suitable for the detection and judgment of multifocal diseases. Furthermore, by obtaining multiple measurement points on the test sample and the signal light characteristic curves of the measurement points, the efficiency of obtaining feature data can be improved. Moreover, the corresponding classification labels can be obtained based on the original data (i.e., measurement points and corresponding signal light characteristic curves) using the detection model, without the need for fitting analysis of the signal light characteristic curves, which is beneficial to improving detection efficiency.

[0142] Furthermore, in this embodiment, the classification labels corresponding to the measurement points are first obtained. The measurement points, the corresponding classification labels, and the signal light feature curve constitute the first training dataset. The initial detection model is then trained using the first training dataset to obtain a trained detection model. That is, this embodiment is a supervised clustering algorithm, and the number of clusters is relatively clear (2 or 3), so that the detection model can easily achieve high accuracy, thereby facilitating the detection of complex tissues.

[0143] In other embodiments, the detection method may further include: obtaining the positive rate of the test sample based on the classification labels of all measurement points on the same test sample. Specifically, this is achieved by using the same test sample FDU-PA230002-CN

[0144] The proportion of measurement points with different classification labels is used to obtain the positive rate of the test samples.

[0145] Accordingly, the present invention also provides a detection device. Figure 7 This is a schematic diagram of the structure of an embodiment of the detection device of the present invention.

[0146] In this embodiment, the detection device is used to obtain the feature data of the detection sample in the detection method of the aforementioned embodiment.

[0147] In this embodiment, the detection device includes: a light source for providing detection light; an optical fiber detection end 300 connected to the light source, the optical fiber detection end 300 serving as a detection end for detecting measurement points on the detection sample; wherein the detection light is irradiated onto the measurement point through the optical fiber detection end 300, causing the detection sample at the measurement point location to generate signal light; and a detection module 310 connected to the optical fiber detection end 300, the detection module 310 being used to detect the signal light generated by the detection sample at the measurement point location and obtain the signal light characteristic curve of each measurement point.

[0148] The detection device includes an optical fiber detection end 300 connected to the light source. The optical fiber detection end 300 serves as a detection end for detecting measurement points on the test sample. The detection light shines through the optical fiber detection end onto the measurement point, causing the test sample at that point to generate signal light. This provides a basis for the detection module 310 to detect the signal light and obtain the signal light characteristic curve of the signal point. Compared to staining and optical detection of tissue sections, measuring the measurement points through the optical fiber detection end eliminates the need for sampling, fixing, and staining of the test sample. Instead, it directly detects each measurement point on the surface of the test sample to obtain the signal light characteristic curve, simplifying the process, improving the convenience of detection, and reducing the time spent on each measurement point, thus significantly increasing the detection speed. Furthermore, the optical fiber detection end 300 facilitates the detection of samples such as cells, tissues, skin on the human body surface, and internal devices, expanding the applicability of the test samples.

[0149] In one embodiment, the detection module 311 includes: multiple sets of detection units (not shown), each set of detection units detecting optical signals with different wavelength ranges; each set of detection units includes a filter and a detector connected to each other, the filter being connected to the optical fiber detection end.

[0150] The detection module 311 includes multiple sets of detection units, and each set of detection units detects optical signals with different wavelength ranges, thereby obtaining multi-channel feature data. More specifically, in this embodiment, each set of detection units is used to obtain the signal light feature curve of the current channel.

[0151] In this embodiment, the filter is used to filter optical signals from different channels so that each channel can acquire FDU-PA230002-CN.

[0152] Data of optical signals in the corresponding wavelength band; the detector is used to collect data of optical signals from different channels.

[0153] In this embodiment, by setting up multiple sets of detection units, the characteristic data of the channel can be detected simultaneously, which helps to improve detection efficiency.

[0154] In another embodiment, the detection module 312 includes: a filter turntable connected to the optical fiber detection end, wherein a plurality of filters are disposed on the filter turntable; and a detector connected to the filter turntable.

[0155] By setting up a filter turntable and selecting different filters on the turntable to work with the detector, multi-channel feature data can be obtained.

[0156] In yet another embodiment, the detection module 313 includes a spectrometer connected to the optical fiber detection end for obtaining a spectral curve.

[0157] Accordingly, the present invention also provides a system for constructing a detection model. Figure 8 This is a functional block diagram of an embodiment of the detection model construction system of the present invention.

[0158] refer to Figure 8In this embodiment, the detection model construction system 400 includes: a data acquisition module 401, used to acquire feature data of the detection sample, the feature data including multiple measurement points located on the detection sample and the feature curve of the signal light generated by each measurement point; an imaging module 402, used to acquire optical imaging data of tissue slices of the detection sample at several intervals; an image segmentation module 403, used to perform image segmentation on the optical image of each tissue slice based on the optical imaging data to obtain multiple segmented images; a block label acquisition module 404, used to acquire a block label for each segmented image in the tissue slice; and a classification label acquisition module. Module 405 is used to obtain the classification label corresponding to the measurement point based on the positional correspondence between the segmented image and the measurement point, and the block label of the segmented image. The measurement point, the corresponding classification label, and the signal light feature curve are used to form a first training dataset. Model building module 406 is used to build an initial detection model. Training module 407 is used to train the initial detection model using the first training dataset to obtain a trained detection model. Prediction module 408 is used to use the trained detection model to predict the feature curve of the measurement point and the corresponding signal light to obtain the corresponding classification label.

[0159] The data acquisition module 401 obtains the feature data of the detection sample so that it can obtain the classification label corresponding to the measurement point based on the positional correspondence between the segmented image and the measurement point and the block label of the segmented image. This enables the measurement point, the corresponding classification label, and the signal light feature curve to form the first training dataset, which can then be used to train the initial detection model.

[0160] The test sample is the sample to be tested in order to obtain a diagnostic result.

[0161] FDU-PA230002-CN

[0162] The test sample is equipped with multiple measurement points, each of which serves as a measurement location. This facilitates the detection of the signal light characteristic curve generated at each measurement location. Compared to staining and optical detection of tissue sections, this embodiment obtains the signal light characteristic curve by detecting each measurement point, eliminating the need for sampling, fixing, and staining of the test sample. Instead, the signal light characteristic curve is obtained directly at each measurement point on the surface of the test sample. This simplifies the process, improves the convenience of detection, and reduces the time spent on each measurement point, thus significantly increasing the detection speed.

[0163] As an example, Figure 2(a) shows a schematic diagram of a test sample 10 and multiple measurement points 11 on the test sample 10.

[0164] As an example, the test sample is used as an external Paget's disease test sample. In other embodiments, the test sample can also be a test sample that can reflect other cancer types, for example, it can be extended to other multifocal skin cancers, including: basal cell carcinoma, squamous cell carcinoma, melanoma, dermatofibrosarcoma, etc. The detection of these cancer types can all be achieved by directly detecting the lesion site through the fiber optic detection end.

[0165] In other embodiments, the test samples can also be those capable of reflecting other types of cancer, including gastrointestinal cancers (oral cancer, esophageal cancer, stomach cancer, intestinal cancer, etc.), urinary system cancers (urethral cancer, bladder cancer, ureteral cancer, kidney cancer, etc.), respiratory system cancers (tracheal cancer, bronchial cancer, lung cancer, etc.), and reproductive system cancers (cervical cancer, vaginal cancer, vulvar cancer, fallopian tube cancer, ovarian cancer, etc.). These cancers can be detected using an endoscope combined with a fiber optic detection end.

[0166] In this embodiment, obtaining the feature data of the test sample includes: setting the measurement point on the test sample based on the detection requirements; detecting the measurement point to obtain the feature data of the measurement point.

[0167] It should be noted that, Figure 2 The size, location, number, density, and arrangement of the measurement points shown in (a) are merely examples. In actual implementation, the size, location, number, density, and arrangement of the measurement points are determined based on the actual detection needs.

[0168] The characteristic curve of the signal light generated by the detection light on the test sample at each measurement point is obtained. The characteristic curve of the signal light can reflect the characteristics of the signal light generated by different substances in the test sample when excited by the detection light. Then, the measurement information and state of different substances in the test sample can be obtained from the characteristics of the signal light, which makes it easier to predict the block label of the test sample at different measurement points.

[0169] Specifically, in actual testing, detection light (e.g., laser) is projected onto the measurement point of the test sample through the fiber optic detection end, causing the test sample at the measurement point to generate signal light, thereby triggering the FDU-PA230002-CN signal.

[0170] It can detect the signal light generated by the test sample at the measurement point location, and then obtain the characteristic curve of the signal light.

[0171] In this embodiment, the step of obtaining feature data of the test sample involves obtaining multi-channel feature data of the test sample, where each channel acquires a different wavelength range of optical signal. By obtaining multi-channel feature data of the test sample, the received signal light can cover a certain spectral width, thereby increasing the dimensionality and comprehensiveness of the detection, and correspondingly improving the accuracy of the detection.

[0172] As an example, in this embodiment, characteristic curves of signal light generated by two coenzymes, collagen, and protoporphyrin IX (PpIX) in the test sample are obtained. In this embodiment, the two coenzymes are nicotinamide adenine dinucleotide (phosphate) (NAD(P)H, reduced form of nicotinamide-adenine dinucleotide (phosphate)) and flavin adenine dinucleotide (FAD).

[0173] Specifically, in this embodiment, for NAD(P)H, a picosecond laser of 330nm to 410nm (e.g., a 405nm laser) or a femtosecond laser of 660nm to 820nm can be provided as the detection light to receive fluorescence in the signal light band of 420nm to 540nm. For FAD, a picosecond laser of 350nm to 490nm (e.g., a 488nm laser) or a femtosecond laser of 700nm to 980nm can be provided as the detection light to receive fluorescence in the signal light band of 500nm to 620nm. For collagen, a picosecond laser of 380nm to 520nm or a femtosecond laser of 760nm to 1040nm can be provided as the detection light. The received signal light band can be selected in the range of 350-600nm depending on the wavelength of the detection light, and can receive the generated fluorescence or the second harmonic light signal generated by femtosecond light excitation. For PpIX, picosecond lasers of 390nm to 440nm or femtosecond lasers of 780nm to 880nm can be provided as detection light, and the wavelength of the received signal light is fluorescence of 550nm to 680nm.

[0174] Accordingly, in this embodiment, the signal light includes fluorescence and second harmonic light signals. In other embodiments, based on the type of substance in the actual detection sample and the type of detection light, the signal light may also include other types of light signals. Correspondingly, in other embodiments, the number of channels of the obtained feature data may also be other numbers.

[0175] In other words, in this embodiment, in the step of obtaining the characteristic data of the test sample, multiple channels respectively acquire the fluorescence generated by NAD(P)H, the fluorescence generated by FAD, the fluorescence or second harmonic light signal generated by collagen, and the fluorescence generated by PpIX, and obtain the characteristic curve of the signal light corresponding to the fluorescence or second harmonic light signal.

[0176] FDU-PA230002-CN

[0177] Accordingly, in this embodiment, one channel is used to acquire the characteristic curve of the second harmonic optical signal generated by the test sample. The second harmonic optical signal can reflect the collagen content in the test sample, and the collagen content can reflect the lesion status of the test sample, thereby obtaining more dimensions of information for assessing the lesion status, and thus improving the comprehensiveness and accuracy of the detection.

[0178] As an example, the signal light characteristic curve includes a fluorescence lifetime decay curve or a spectral curve.

[0179] The fluorescence lifetime decay curve contains the intensity ratio between different channels, reflecting the content ratio of the substances corresponding to different channels in the detection sample. It also contains fluorescence lifetime information and fluorescence spectral intensity. Therefore, by obtaining the fluorescence lifetime decay curve of the measurement point, multi-dimensional characteristic information of the signal light can be obtained, which is conducive to improving the accuracy of judgment. In addition, by obtaining the fluorescence lifetime decay curve, there is no need to perform exponential fitting analysis to obtain fluorescence lifetime information. Instead, fluorescence lifetime information can be obtained directly from the fluorescence lifetime decay curve. Subsequently, the raw data of the fluorescence lifetime decay curve can be directly used to form the first training dataset for training the initial detection model and as the input for the detection model, which is conducive to further improving the detection speed.

[0180] The spectral curve can reflect the intensity of the received signal light, and thus reflect the content of substances contained in the sample at the measurement point. Moreover, the speed of obtaining the spectral curve is fast, which is also beneficial to improving the detection speed.

[0181] As an example, Figure 2 (b) and Figure 2 (c) Schematic diagrams showing the spectral curves obtained at measurement point 12 and the fluorescence lifetime decay curves of multiple channels obtained at measurement point 13, respectively. As an example, four channels (e.g.) are obtained. Figure 2 (c) The fluorescence lifetime decay curves corresponding to the signal lights of CH1, CH2, CH3, and CH4 shown are illustrated as an example. In other embodiments, the number of fluorescence lifetime decay curves obtained varies depending on the actual number of channels.

[0182] In this embodiment, when the signal light characteristic curve is a fluorescence lifetime decay curve, obtaining the characteristic data of the detection sample includes: obtaining the original fluorescence lifetime decay curve corresponding to each measurement point; and preprocessing the original fluorescence lifetime decay curve to remove the curve before the peak value in the fluorescence lifetime decay curve.

[0183] The imaging module 402 is used to obtain optical imaging data of tissue slices of the test sample at several intervals, so that the optical images of the tissue slices can be segmented to obtain multiple segmented images based on the optical imaging data, and the block labels of the segmented images can be obtained. Then, the classification label corresponding to the measurement point can be obtained by utilizing the positional correspondence between the segmented images and the measurement points.

[0184] FDU-PA230002-CN

[0185] Reference Figure 3 This illustrates a schematic diagram of slicing the test sample at several intervals. For example... Figure 3 As shown, as an example, optical imaging data of tissue sections of the test sample at 8 intervals were obtained.

[0186] In other embodiments, the number, location, and spacing of tissue slices can be flexibly adjusted based on actual detection needs.

[0187] In this embodiment, fluorescence lifetime imaging data is used as an example for illustration. In other embodiments, the optical imaging data may also be stimulated Raman scattering microscopy (SRS) data, multi-photon excitation (MPE) data, autofluorescence imaging data, histochemical staining image data, or second harmonic imaging data.

[0188] The image segmentation module 403 is used to perform image segmentation on the optical image of each tissue slice based on the optical imaging data to obtain multiple segmented images.

[0189] Image segmentation is performed on the optical image of each tissue slice to obtain multiple segmented images, so as to obtain the block label of each segmented image in the tissue slice. This allows the classification label of the measurement point to be obtained based on the positional correspondence between the segmented images and the measurement points.

[0190] Moreover, compared with the optical image of each tissue slice, the segmented image and measurement points are smaller in size and more precise. This is beneficial for improving the precision of the detection model in classification judgment during the subsequent step of training the initial detection model using the first training dataset to obtain a trained detection model, thereby improving the accuracy and precision of detection.

[0191] In this embodiment, the step of performing image segmentation on the optical image of each tissue slice to obtain multiple segmented images based on the optical imaging data includes: performing image segmentation on the optical image of each tissue slice based on detection requirements to obtain multiple segmented images.

[0192] Specifically, based on the detection requirements, the number of image segments to be performed on the optical image of each tissue slice is set to obtain segmented images of corresponding sizes.

[0193] It should be noted that, in this embodiment, during the step of obtaining the feature data of the test sample, the measurement point is located on the surface of the test sample. The characteristic curve of the signal light generated by the surface tissue of the test sample at the measurement point is measured accordingly. Accordingly, in this embodiment, the segmented image near the surface of each tissue slice in the optical image corresponds to the measurement point. FDU-PA230002-CN

[0194] In practice, OpenCV (Open Source Computer Vision Library) can be used to segment the optical image of each tissue slice to obtain multiple segmented images. More specifically, the optical image can be uniformly cropped to obtain multiple segmented images.

[0195] The block label acquisition module 404 is used to obtain the block label of each segmented image in the tissue slice, so as to obtain the classification label corresponding to the measurement point based on the positional correspondence between the segmented image and the measurement point, and the block label of the segmented image.

[0196] In this embodiment, the block labels of the segmented image are described as normal or malignant. In other embodiments, the block labels of the segmented image may also include benign. Alternatively, in other embodiments, based on actual detection needs and the classification of detection results, the block labels of the segmented image may include other types.

[0197] As an example, the block label acquisition module 404 includes: a pathological diagnosis unit (not shown), used to obtain the pathological diagnosis results of a portion of the tissue slices; a dataset acquisition unit (not shown), used to perform label annotation processing on each segmented image in the portion of the tissue slices based on the pathological diagnosis results, to obtain the block labels of the segmented images, wherein the segmented images and corresponding block labels in the portion of the tissue slices are used to constitute a second training dataset; a classification model construction unit (not shown), used to construct a classification model; a classification model training unit (not shown), used to train the classification model using the second training dataset to obtain a trained classification model; and a prediction unit (not shown), used to perform prediction processing on the segmented images in the remaining portion of the tissue slices using the classification model, to obtain the block labels of the segmented images in the remaining portion of the tissue slices.

[0198] The pathological diagnosis unit obtains the pathological diagnosis results of a portion of the tissue sections, so that, based on the pathological diagnosis results, it can subsequently obtain the block labels of each segmented image in the portion of the tissue sections.

[0199] In this embodiment, only a portion of the tissue sections are obtained for pathological diagnosis, thereby reducing the number of tissue sections required for pathological diagnosis and thus improving the detection rate.

[0200] Specifically, in this embodiment, the step of obtaining the pathological diagnosis results of a portion of the tissue sections may include: staining the portion of the tissue sections and performing pathological examination, with a pathology expert making a judgment to obtain the pathological diagnosis results.

[0201] In practice, the tissue sections of the aforementioned number can be stained with hematoxylin and eosin (H&E), followed by pathological examination and judgment by a pathologist to obtain a pathological diagnosis.

[0202] FDU-PA230002-CN

[0203] The dataset acquisition unit obtains the patch labels for each segmented image from a subset of tissue slices.

[0204] Specifically, based on the pathological diagnosis results of the tissue section, each segmented image of the current tissue section is labeled with a corresponding block label. For example, if the pathological diagnosis result of the current tissue section is malignant, then each segmented image of the current tissue section is labeled as malignant; if the pathological diagnosis result of the current tissue section is normal, then each segmented image of the current tissue section is labeled as normal. In other words, the segmentation label corresponding to each segmented image of the current tissue section is consistent with the pathological diagnosis result of the tissue section.

[0205] Reference Figure 5 Two examples of optical images 100 of tissue sections are shown. Optical image 110 shows a tissue section with a pathological diagnosis of malignancy, and the corresponding segmentation image 130 is also labeled as malignant; optical image 120 shows a tissue section with a pathological diagnosis of normal, and the corresponding segmentation image 130 is also labeled as normal.

[0206] The classification model building unit constructs a classification model so that the classification model training unit can use the second training dataset to train the classification model.

[0207] As an example, the classification model includes a Residual Neural Network (ResNet) model. Residual Neural Networks employ "shortcut connections" to address degradation, significantly reducing the training difficulties associated with excessively deep neural networks. As an example, the ResNet model is a ResNet 50 model.

[0208] In other embodiments, the classification model may be other types of classification models based on actual needs.

[0209] As an example, the block label acquisition module 404 may further include: a recognition unit (not shown), used to perform image recognition processing on the segmented images in the second training dataset after obtaining the block label of each segmented image in the partial number of tissue slices based on the pathological diagnosis result, and before training the classification model using the second training dataset, to obtain image recognition results, and based on the image recognition results, to remove segmented images with a background area ratio greater than a preset value from the second training dataset.

[0210] The segmented images in the second training dataset are processed by image recognition. Based on the image recognition results, the segmented images with a background area ratio greater than a preset value are removed from the second training dataset. This prevents the segmented images with an excessively large background area ratio from being mixed in the second training dataset, which helps to prevent interference with the training of subsequent classification models and reduces the probability of false positives or false negatives predicted by the subsequently trained classification models.

[0211] FDU-PA230002-CN

[0212] The classification model training unit uses the second training dataset to train the classification model, so that the classification model can analyze and learn the relationship between the segmented image and the block label. Subsequently, the classification model can be used to predict the segmented images of the remaining number of tissue slices to obtain the corresponding block labels.

[0213] Reference Figure 5 In this embodiment, a second training dataset is constructed using only tissue sections 110 with pathological diagnoses of malignancy and 120 with pathological diagnoses of normal tissue sections, as an example to illustrate the training process for the classification model. In other embodiments, based on actual detection needs, a second training dataset can also be constructed using tissue sections with pathological diagnoses of malignancy, normal tissue sections, and benign tissue sections, as well as pathological diagnoses of benign tissue sections, to train the classification model.

[0214] The prediction unit is used to predict the segmented images of the remaining number of tissue slices using a classification model to obtain corresponding block labels. Compared with the optical images of tissue slices, the segmented images are smaller in size, which is beneficial for obtaining more refined labels. Moreover, predicting the segmented images of the remaining number of tissue slices using a classification model not only improves the accuracy of obtaining block labels for the segmented images, but also saves a lot of annotation time and increases the speed of obtaining annotations for the segmented images.

[0215] like Figure 6 The diagram illustrates the prediction processing performed on the segmented images corresponding to optical images 100a, 100b, and 100c to obtain the block labels of the corresponding segmented image 130. Optical image 100a represents a completely malignant tissue section, optical image 100b represents a tissue section of a mixed region, and optical image 100c represents a completely normal tissue section. Figure 6 In the segmented image, the color intensity of the segmented image is used to distinguish between malignant and normal conditions.

[0216] Among them, "all benign," "all malignant," or "all normal" means that the block labels of all segmented images of the current tissue slice are benign, malignant, or normal.

[0217] It should be noted that in this embodiment, the example used is to train the classification model using a classification model training unit and then use a prediction unit to obtain the block labels of the segmented images in the remaining number of tissue slices. However, the methods for obtaining the block labels of the segmented images in the remaining number of tissue slices are not limited to this.

[0218] In other embodiments, the segmentation label can be assigned to each segmented image in the remaining number of tissue sections based on the pathological diagnosis results. Specifically, the segmentation label of each segmented image is consistent with the pathological diagnosis result of the corresponding tissue section.

[0219] It should also be noted that, in some other embodiments, the block tag acquisition module may further include: FDU-PA230002-CN

[0220] The cleaning and correction unit (not shown) is used to clean and correct the second training dataset; the retraining unit (not shown) is used to retrain the classification model using the second training dataset after cleaning and correcting it; and the re-prediction unit (not shown) is used to predict the segmented images using the retrained classification model to obtain the block label of each segmented image.

[0221] The data cleaning and correction unit cleans and corrects the second training dataset, and the retraining unit retrains the classification model to improve the accuracy of the training samples. This minimizes the interference of erroneous samples on model training, thereby improving the reliability and robustness of the classification model obtained through training, and consequently improving the accuracy of the block labels obtained during prediction processing.

[0222] In practice, the second training dataset can be cleaned and corrected based on cross-validation combined with confidence learning, fidelity-based weighted learning, or sampling detection.

[0223] In some embodiments, cleaning and correcting the second training dataset based on cross-validation combined with confidence learning may include: obtaining the joint distribution of the pathological diagnosis results and the block labels of the segmented images in the second training dataset; and based on the joint distribution, deleting samples from the second training dataset whose probability is lower than the current class mean and whose probability of another class is higher than the other class mean, according to the mean of the probability of each class.

[0224] By obtaining the joint distribution of the pathological diagnosis results and the block labels of the segmented images in the second training dataset, and based on the joint distribution, according to the mean of the probability of each class, samples with a probability lower than the current class mean and a probability of another class higher than the other class mean are deleted from the second training dataset, thereby achieving the purpose of data cleaning, and the class weights in the second training dataset are readjusted, and then the classification model is retrained to improve the accuracy of the classification model.

[0225] For example, in specific implementation, based on the joint distribution of pathological diagnosis results and block labels, and according to the mean of the normal and malignant class probabilities, samples with a probability lower than the mean of the normal class and a probability higher than the mean of the malignant class are deleted from the second training dataset.

[0226] In some embodiments, cleaning and correcting the second training dataset based on fidelity-weighted learning may include: using the classification model as a first classification model; training a student network using the optical images of the tissue slices and the corresponding pathological diagnosis results; training a teacher network using the segmented images and the corresponding block labels; and generating soft labels for all segmented images and the corresponding block labels using the teacher network to obtain a soft dataset (Soft FDU-PA230002-CNDataset).

[0227] Accordingly, the student network is subjected to adversarial training using the soft dataset to obtain a second classification model; and the second classification model is used to predict all segmented images to obtain the block labels of segmented images in all tissue slices.

[0228] By training a student network and a teacher network, and using the teacher network to obtain a soft dataset, and then using the soft dataset to perform adversarial training on the student network, a second classification model that approximates the first classification model is obtained. Furthermore, during the adversarial training process, the noise data in the optical images and corresponding pathological diagnosis results can be preserved, and corresponding fidelity-preserving weighted learning can be performed, thereby improving the accuracy and predictive realism of the second classification model.

[0229] In some embodiments, cleaning and correcting the second training dataset based on sampling detection may include: dividing the second training dataset into multiple sub-data sets, and using a classification model to perform sampling prediction on each sub-data set; after performing sampling prediction on each sub-data set, cleaning the second training dataset to re-label or delete data whose prediction results are inconsistent with the classification results obtained based on pathological diagnosis results; retraining the classification model using the cleaned second training dataset; and after retraining the classification model, using the classification model to predict the segmented images in the remaining number of tissue slices to obtain the block labels of the segmented images in the remaining number of tissue slices.

[0230] By cleaning the second training dataset, data whose prediction results are inconsistent with the block labels obtained based on pathological diagnosis results are re-labeled or deleted from the second training dataset, and the classification model is retrained, thereby preventing erroneous sample data from interfering with model training and improving the accuracy of classification model prediction.

[0231] It should be noted that in this embodiment, after obtaining the block labels of the segmented image corresponding to the optical image of each tissue slice, a positive rate can be assigned to images with a large size range based on the proportion of different block labels in the segmented image of each tissue slice. As an example, the accuracy of the finally trained classification model can reach 95%, and it can assign positive rates to images with a large size range for slices of malignant, mixed, and normal regions. The positive rate for malignant regions is close to 100%, the positive rate for normal regions is 0%, while the positive rate for mixed regions varies depending on the specific cancer type. Taking extramammary Paget's disease as an example, the positive rate for mixed regions measured in this embodiment is approximately 30% to 60%. Therefore, in this embodiment, the prediction accuracy of the classification model is high.

[0232] FDU-PA230002-CN

[0233] The classification label acquisition module 405 obtains the classification label corresponding to the measurement point based on the positional correspondence between the segmented image and the measurement point, as well as the block label of the segmented image. The measurement point, the corresponding classification label, and the signal light feature curve are used to form a first training dataset, which facilitates the subsequent training of the initial detection model using the first training dataset.

[0234] Compared to the optical images of each tissue slice, the segmented images and measurement points are smaller and more precise. This improves the precision of the detection model in classification during the training process using the first training dataset to obtain a trained detection model, thus enhancing the accuracy and precision of detection. This is particularly suitable for the detection and assessment of multifocal diseases. Furthermore, by obtaining multiple measurement points and their signal light feature curves on the test sample, the efficiency of obtaining feature data is improved. The detection model can also obtain corresponding classification labels based on the original data (i.e., measurement points and corresponding signal light feature curves) without the need for fitting analysis of the signal light feature curves, thereby improving detection efficiency.

[0235] In this embodiment, the segmented image near the surface of the tissue slice in the optical image of each tissue slice corresponds to the measurement point. Accordingly, the block label of the segmented image near the surface of the tissue slice is used as the classification label of the measurement point at the corresponding position.

[0236] The model building module 406 is used to build an initial detection model so that the training module 407 can use the first training dataset to train the initial detection model and obtain a trained detection model.

[0237] As an example, the model building module 406 builds the initial detection model based on logistic regression, decision tree, random forest, or support vector machine (SVM).

[0238] In this embodiment, the initial detection model constructed by the model building module 406 based on the support vector machine is used as an example for illustration. The support vector machine is a novel few-sample learning method with a solid theoretical foundation. It basically does not involve probability measures and the law of large numbers, and its final decision function is determined by only a few support vectors. The computational complexity depends on the number of support vectors, rather than the dimension of the sample space, which avoids the "curse of dimensionality" in a sense.

[0239] In other embodiments, the model building module may also build the initial detection model based on other suitable classification models.

[0240] It should be noted that, in this embodiment, the detection model construction system 400 may further include: a data augmentation module (not shown), used to obtain the classification label corresponding to the measurement point based on the positional correspondence between the segmented image and the measurement point, and the block labels of the segmented image, and then... (FDU-PA230002-CN)

[0241] Furthermore, before training the initial detection model using the first training dataset, the first training dataset is augmented.

[0242] The data augmentation module augments the first training dataset to increase the sample size.

[0243] As an example, the data augmentation module includes any one or more of the following: a superposition unit (not shown) for linearly superimposing several signal light characteristic curves of the same category using a random scaling factor that sums to 1; a translation unit (not shown) for randomly translating the signal light characteristic curve along the horizontal axis within a limited range of spectral line shift fluctuations; and a noise addition unit (not shown) for adding random Gaussian white noise of varying intensity to the signal light characteristic curve.

[0244] The training module 407 is used to train the initial detection model using the first training dataset to obtain a trained detection model.

[0245] The detection model is used to obtain corresponding classification labels based on measurement points and corresponding signal light feature curves.

[0246] In this embodiment, multiple segmented images are obtained by image segmentation of the optical image of each tissue slice, and block labels are obtained for each segmented image in the tissue slice. Based on the positional correspondence between the segmented images and the measurement points, and the block labels of the segmented images, classification labels corresponding to the measurement points are obtained. Compared with the optical image of each tissue slice, the segmented images and measurement points are smaller in size and have higher precision. Therefore, in the step of training the initial detection model using the first training dataset to obtain the trained detection model, the precision of the detection model in classification judgment can be improved, which is conducive to improving the accuracy and precision of detection, especially suitable for the detection and judgment of multifocal diseases. Moreover, by obtaining multiple measurement points on the detection sample and the signal light feature curves of the measurement points, it is also beneficial to improve the efficiency of obtaining feature data. Furthermore, the detection model can be used to perform prediction processing based on the original data (i.e., measurement points and corresponding signal light feature curves) to obtain the corresponding classification labels without the need for fitting analysis of the signal light feature curves, which is beneficial to improving detection efficiency.

[0247] The prediction module 408 is used to perform prediction processing on the feature curves of the measurement point and the signal light using the trained detection model to obtain the corresponding classification label.

[0248] By utilizing a trained detection model, the characteristic curves of the measurement points and the signal light are predicted to obtain corresponding classification labels, thereby improving the accuracy and precision of detection. This is particularly suitable for the detection and diagnosis of multifocal diseases. Furthermore, obtaining multiple measurement points on the test sample and the signal light characteristic curves of these points improves the efficiency of obtaining feature data and enables the processing of raw data based on the original data.

[0249] The corresponding classification label can be obtained by using the detection model based on the measurement point and the corresponding signal light characteristic curve, without the need for fitting analysis of the signal light characteristic curve, which is beneficial to improving detection efficiency.

[0250] Furthermore, in this embodiment, the classification labels corresponding to the measurement points are first obtained. The measurement points, the corresponding classification labels, and the signal light feature curve constitute the first training dataset. The initial detection model is then trained using the first training dataset to obtain a trained detection model. That is, this embodiment is a supervised clustering algorithm, and the number of clusters is relatively clear (2 or 3), so that the detection model can easily achieve high accuracy, thereby facilitating the detection of complex tissues.

[0251] In other embodiments, the detection method may further include: obtaining the positive rate of the test sample based on the classification labels of all measurement points on the same test sample. Specifically, the positive rate of the test sample is obtained based on the proportion of measurement points with different classification labels on the same test sample.

[0252] This invention also provides a device that can implement the detection method provided by this invention by using the above-described detection model construction method in the form of a loading program.

[0253] An optional hardware structure for the terminal device provided in this embodiment of the invention can be as follows: Figure 9 As shown, it includes: at least one processor 01, at least one communication interface 02, at least one memory 03, and at least one communication bus 04.

[0254] In this embodiment of the invention, the number of processor 01, communication interface 02, memory 03, and communication bus 04 is at least one, and processor 01, communication interface 02, and memory 03 communicate with each other through communication bus 04; optionally, communication interface 02 can be an interface of a communication module for network communication, such as the interface of a GSM module; processor 01 may be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement this embodiment of the invention.

[0255] Memory 03 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0256] The memory 03 stores one or more computer instructions, which are executed by the processor 01 to implement the detection method provided in this embodiment of the invention.

[0257] It should be noted that the aforementioned terminal device may also include other devices (not shown) that may not be essential to understanding the content disclosed in the embodiments of the present invention; given that these other devices may not be essential for understanding the content disclosed in the embodiments of the present invention, the embodiments of the present invention will not describe them one by one.

[0258] FDU-PA230002-CN

[0259] Accordingly, embodiments of the present invention also provide a storage medium that stores one or more computer instructions, which are executed by the processor to implement the detection method provided in the embodiments of the present invention.

[0260] The embodiments of the present invention described above are combinations of elements and features of the present invention. Unless otherwise stated, the elements or features described are optional. Individual elements or features may be practiced without combination with other elements or features. Furthermore, embodiments of the present invention may be constructed by combining some elements and / or features. The order of operations described in the embodiments of the present invention may be rearranged. Some constructions of any embodiment may be included in another embodiment and may be replaced by corresponding constructions of another embodiment. It will be apparent to those skilled in the art that claims in the appended claims that are not expressly referenced to each other may be combined to form embodiments of the present invention, or may be included as new claims in amendments made after the filing of this application.

[0261] Embodiments of the present invention can be implemented by various means, such as hardware, firmware, software, or combinations thereof. In a hardware configuration, the method according to an exemplary embodiment of the present invention can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, etc.

[0262] In firmware or software configuration, embodiments of the present invention can be implemented in the form of modules, processes, functions, etc. Software code can be stored in a memory unit and executed by a processor. The memory unit is located inside or outside the processor and can send data to and receive data from the processor via various known means.

[0263] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is accorded the widest scope consistent with the principles and novel features disclosed herein.

[0264] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A detection method, characterized in that, include: The feature data of the test sample is obtained, the feature data including multiple measurement points located on the test sample and the feature curve of the signal light generated by each measurement point; Obtain optical imaging data of tissue sections of the test sample at several intervals; Based on the optical imaging data, the optical image of each tissue slice is segmented to obtain multiple segmented images; Obtain the block labels for each segmented image in the tissue slice; Based on the positional correspondence between the segmented image and the measurement point, and the block label of the segmented image, the classification label corresponding to the measurement point is obtained. The measurement point, the corresponding classification label, and the signal light feature curve are used to form the first training dataset. Construct an initial detection model; The initial detection model is trained using the first training dataset to obtain a trained detection model; Using the trained detection model, the characteristic curves of the measurement points and the corresponding signal light are predicted to obtain the corresponding classification labels; The steps for obtaining feature data of a test sample include: setting the measurement points on the test sample based on the testing requirements; detecting the measurement points to obtain feature data of the measurement points; The step of performing image segmentation on the optical image of each tissue slice to obtain multiple segmented images based on the optical imaging data includes: performing image segmentation on the optical image of each tissue slice based on detection requirements to obtain multiple segmented images; The segmented image near the surface of the tissue slice in the optical image of each tissue slice corresponds to the measurement point.

2. The detection method as described in claim 1, characterized in that, The steps for obtaining the block labels of each segmented image in the tissue slice include: Obtain pathological diagnostic results for a portion of the tissue sections; Based on the pathological diagnosis results, each segmented image in the specified number of tissue slices is labeled to obtain the block labels of the segmented images. The segmented images and corresponding block labels in the specified number of tissue slices are used to form the second training dataset. Build a classification model; Using the second training dataset, the classification model is trained to obtain a trained classification model; The trained classification model is used to predict the segmentation images in the remaining number of tissue slices to obtain the block labels of the segmentation images in the remaining number of tissue slices.

3. The detection method as described in claim 2, characterized in that, The step of obtaining the block label of each segmented image in the tissue slice further includes: obtaining the block label of each segmented image in the specified number of tissue slices based on the pathological diagnosis result; and before training the classification model using the second training dataset, performing image recognition processing on the segmented images in the second training dataset to obtain image recognition results; based on the image recognition results, removing segmented images with a background area ratio greater than a preset value from the second training dataset; and / or, The step of obtaining the block label of each segmented image in the tissue slices further includes: after obtaining the block labels of the segmented images in the remaining number of tissue slices, cleaning and correcting the second training dataset based on cross-validation combined with confidence learning, fidelity-weighted learning, or sampling detection; after cleaning and correcting the second training dataset, retraining the classification model using the second training dataset; and using the retrained classification model to predict the segmented images to obtain the block label of each segmented image.

4. The detection method as described in claim 1, characterized in that, The characteristic curves of the signal light include fluorescence lifetime decay curves or spectral curves.

5. The detection method as described in claim 1, characterized in that, In the step of obtaining feature data of the test sample, multi-channel feature data of the test sample is obtained, and the wavelength range of the optical signal acquired by each channel is different.

6. The detection method as described in claim 1, characterized in that, The method further includes: after obtaining the classification label corresponding to the measurement point based on the positional correspondence between the segmented image and the measurement point, and the block label of the segmented image, and before training the initial detection model using the first training dataset, performing data augmentation processing on the first training dataset; The steps for augmenting the first training dataset include any one or more of the following: linearly superimposing several signal light feature curves of the same category using a random scaling factor that sums to 1; randomly shifting the signal light feature curves along the horizontal axis within a limited range of spectral line shift fluctuations; and adding random Gaussian white noise of varying intensity to the signal light feature curves.

7. The detection method as described in claim 1, characterized in that, The optical imaging data includes fluorescence lifetime imaging data, stimulated Raman scattering microscopy imaging data, multiphoton excitation fluorescence imaging data, autofluorescence imaging data, histochemical staining image data, or second harmonic imaging data.

8. The detection method as described in claim 1, characterized in that, The detection method further includes: obtaining the positive rate of the test sample based on the classification labels of all measurement points on the same test sample.

9. A detection device, characterized in that, The detection device is used to obtain feature data of the detection sample in the detection method according to any one of claims 1 to 8; the detection device includes: A light source, used to provide detection light; An optical fiber detection end is connected to the light source and is used as a detection end to detect measurement points on the test sample; wherein, the detection light shines on the measurement point through the optical fiber detection end, causing the test sample at the measurement point to generate signal light; The detection module is connected to the optical fiber detection end. The detection module is used to detect the signal light generated by the detection sample at the measurement point and obtain the signal light characteristic curve of each measurement point.

10. The detection device as described in claim 9, characterized in that, The detection module includes: Multiple detection units, each detecting optical signals within a different wavelength range; each detection unit includes interconnected filters and detectors, the filters being connected to the optical fiber detection end; or, A filter turntable is connected to the optical fiber detection end, and multiple filters are provided on the filter turntable; a detector is connected to the filter turntable.

11. The detection device as described in claim 10, characterized in that, The detection module includes a spectrometer connected to the optical fiber detection end for obtaining spectral curves.

12. A detection system, characterized in that, include: The data acquisition module is used to obtain feature data of the test sample, the feature data including multiple measurement points located on the test sample and the feature curve of the signal light generated by each measurement point; The imaging module is used to acquire optical imaging data of tissue sections of the test sample at several intervals. An image segmentation module is used to perform image segmentation on the optical image of each tissue slice based on the optical imaging data to obtain multiple segmented images; The block label acquisition module is used to obtain the block label of each segmented image in the tissue slice; The classification label acquisition module is used to obtain the classification label corresponding to the measurement point based on the positional correspondence between the segmented image and the measurement point, and the block label of the segmented image. The measurement point, the corresponding classification label, and the signal light feature curve are used to form the first training dataset. The model building module is used to build the initial detection model; The training module is used to train the initial detection model using the first training dataset to obtain a trained detection model. The prediction module is used to use the trained detection model to predict the feature curves of the measurement point and the corresponding signal light to obtain the corresponding classification label. The data acquisition module is also used to set the measurement points on the test sample based on the detection requirements; to detect the measurement points and obtain the feature data of the measurement points; The image segmentation module is also used to perform image segmentation on the optical image of each tissue slice based on detection requirements to obtain multiple segmented images; The segmented image near the surface of the tissue slice in the optical image of each tissue slice corresponds to the measurement point.

13. A device, characterized in that, It includes at least one memory and at least one processor, the memory storing one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the detection method as described in any one of claims 1-8.

14. A storage medium, characterized in that, The storage medium stores one or more computer instructions, which are used to implement the detection method as described in any one of claims 1-8.

Images