A fluorescence lifetime-based detection method and device, and a readable storage medium
Patent Information
- Application Number
- CN202410969215.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2026-10-30
AI Technical Summary
整个检测过程需要病理科医生全程实施，人力成本高，处理时间较长，易受经验性和主观性影响，而且目前是从细胞形态进行区分，对肿瘤存在的更早期的微环境的改变，比如细胞形态尚未发生明显改变的微环境，还无法甄别
[0041]1、本发明通过染色、荧光寿命成像和机器学习分类预测完成对检测样本的整个检测过程，相对传统方法，检测效率更高，成本更低。
Smart Images
Figure CN123016424A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of optical detection, and more particularly to a detection method and apparatus based on fluorescence lifetime, and a readable storage medium. Background Technology
[0002] The incidence of endometrial cancer is increasing year by year, making early screening and diagnosis crucial. Currently, pathological diagnosis is the gold standard for diagnosing endometrial cancer. In clinical practice, endometrial tissue can be collected through three methods: aspiration biopsy, curettage, and hysteroscopy. Aspiration biopsy is widely used clinically due to its less invasive nature, simplicity, and lower cost.
[0003] For example, a biopsy is performed to obtain endometrial tissue, which is then tested, including sequentially performing H&E staining, immunohistochemistry, optical microscopy, slide interpretation, and finally diagnosis. The entire testing process requires a pathologist to perform it, resulting in high labor costs, long processing time, and susceptibility to experience and subjectivity. Furthermore, current methods differentiate tumors based on cell morphology, failing to identify earlier changes in the tumor microenvironment, such as microenvironments where cell morphology has not yet undergone significant alterations.
[0004] Therefore, there is an urgent need for a faster and more cost-effective testing method to assist doctors in making diagnoses. Summary of the Invention
[0005] The purpose of this invention is to provide a fluorescence lifetime-based detection method and apparatus, and a readable storage medium, to solve at least some of the problems existing in the current detection technology.
[0006] The technical solution provided by this invention is as follows:
[0007] A fluorescence lifetime-based detection method includes: acquiring a test sample and staining it with a viscosity fluorescent probe;
[0008] Fluorescence lifetime imaging was performed on the stained test samples to obtain fluorescence lifetime images of the test samples;
[0009] The fluorescence lifetime image is preprocessed;
[0010] The preprocessed fluorescence lifetime image is input into a classification model to classify and predict the detected samples; the classification model is based on a deep learning network.
[0011] In some embodiments, obtaining the test sample and staining it with a viscosity fluorescent probe includes:
[0012] A fixative is added to fix the contents of the test sample;
[0013] Add the first preset amount of antioxidant;
[0014] After centrifugation to remove the supernatant, a viscosity-sensitive fluorescent probe was added to stain the contents.
[0015] Centrifuge again to remove the supernatant, and use the fixative to wash the precipitate by blowing and washing. Repeat the centrifugation-washing process a preset number of times to obtain the stained cell sample to be tested.
[0016] In some embodiments, before centrifuging again to remove the supernatant after staining, the process includes:
[0017] Add a second preset amount of anti-quenching agent.
[0018] In some embodiments, the preprocessing of the fluorescence lifetime image includes:
[0019] The contrast of the fluorescence lifetime image is improved by using grayscale broadening and / or grayscale histogram equalization.
[0020] Cell segmentation was performed on the fluorescence lifetime images with improved contrast.
[0021] In some embodiments, cell segmentation of the contrast-enhanced fluorescence lifetime image includes:
[0022] An improved first U-Net network is used to denoise the fluorescence lifetime image with improved contrast;
[0023] Cell segmentation is performed on the denoised image;
[0024] The first U-Net network is based on the base U-Net network, replacing the standard convolutional layers in the base U-Net network with residual blocks, and adding an image style representation layer between model downsampling and upsampling. After the downsampling process, the image style representation layer performs global average pooling and normalization on the obtained feature map to obtain a 256-dimensional image style vector. The image style vector is projected onto a dimension that matches the number of channels of the feature map of the current upsampling layer, and superimposed into the input of each upsampling layer through a broadcast mechanism.
[0025] In some embodiments, the cell segmentation of the denoised image includes:
[0026] An improved second U-Net network is used to perform cell segmentation on the denoised image;
[0027] The first U-Net network and the second U-Net network have the same structure but different parameters.
[0028] In some embodiments, training the second U-Net network includes:
[0029] The unlabeled sample images are segmented using the trained base U-Net model, and the segmentation results are used as segmentation labels.
[0030] The second U-Net network is trained using sample images labeled with the segmentation tags.
[0031] In some embodiments, the construction of the classification model includes:
[0032] Prepare a sample dataset containing several samples, each of which is a fluorescence lifetime image carrying coarse labels, including high-risk and low-risk labels;
[0033] Build deep learning networks;
[0034] The deep learning network was used to perform cross-validation on the sample dataset.
[0035] Based on the results of the cross-validation, the prediction accuracy is calculated, a confidence learning algorithm is used to identify and delete samples containing incorrect labels, the sample deletion rate is calculated, and the sample dataset is updated.
[0036] The deep learning network is retrained based on the updated sample dataset. The process of cross-validation and removing samples containing incorrect labels is repeated until the sample removal rate no longer decreases and the prediction accuracy no longer increases compared to the previous training results.
[0037] The trained deep learning network is used as the classification model.
[0038] The present invention also provides a fluorescence lifetime-based detection device, comprising: a memory for storing a computer program; and a processor for implementing the fluorescence lifetime-based detection method described in any of the foregoing embodiments when running the computer program.
[0039] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the fluorescence lifetime-based detection method described in any of the foregoing embodiments.
[0040] The fluorescence lifetime-based detection method, apparatus, and readable storage medium provided by this invention can bring at least the following beneficial effects:
[0041] 1. This invention completes the entire detection process of the sample by staining, fluorescence lifetime imaging and machine learning classification prediction. Compared with traditional methods, it has higher detection efficiency and lower cost.
[0042] 2. This invention uses a viscosity fluorescent probe to stain the test sample and identifies intracellular viscosity differences based on the difference in fluorescence lifetime. Compared with the traditional method of identification based on cell morphology, it can detect changes in the microenvironment earlier and improve the efficiency of early screening.
[0043] 3. This invention improves the accuracy of cell segmentation by first denoising the fluorescence lifetime image and then performing cell segmentation.
[0044] 4. This invention improves the U-Net network by replacing the standard convolutional layers in the basic U-Net network with residual blocks and adding an image style representation layer between model downsampling and upsampling, thereby enhancing the capacity, performance, and generalization of the model used for cell segmentation.
[0045] 5. This invention improves the quality of training data and increases the classification accuracy of the classification model by eliminating erroneous labeled samples through confidence learning. Attached Figure Description
[0046] The preferred embodiments will now be described in a clear and easy-to-understand manner, with reference to the accompanying drawings, to further explain the above-mentioned characteristics, technical features, advantages, and implementation methods of a fluorescence lifetime-based detection method and apparatus, and a readable storage medium.
[0047] Figure 1 This is a flowchart of an embodiment of a fluorescence lifetime-based detection method of the present invention;
[0048] Figure 2 This is a schematic diagram of FLIM images of benign and malignant tissues;
[0049] Figure 3 This is a schematic diagram of the first U-Net network structure;
[0050] Figure 4 This is a flowchart of an embodiment of building a classification model;
[0051] Figure 5 This is a schematic diagram of an embodiment of a fluorescence lifetime-based detection device according to the present invention;
[0052] Figure 6 This is a schematic diagram of another embodiment of a fluorescence lifetime-based detection device of the present invention. Detailed Implementation
[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.
[0054] To keep the drawings concise, only the parts relevant to the invention are shown schematically in each figure, and they do not represent the actual structure of the product. Furthermore, for ease of understanding, in some figures, components with the same structure or function are shown only schematically, or only one is labeled. In this document, "a" can mean not only "only one" but also "more than one".
[0055] One embodiment of the present invention, such as Figure 1 As shown, a detection method based on fluorescence lifetime includes:
[0056] Step S100: Obtain the test sample and stain it using a viscosity fluorescent probe;
[0057] Step S200 involves performing fluorescence lifetime imaging processing on the stained test sample to obtain a fluorescence lifetime image of the test sample.
[0058] Step S300 involves preprocessing the fluorescence lifetime image;
[0059] Step S400 inputs the preprocessed fluorescence lifetime image into the classification model to classify and predict the detected samples; the classification model is based on a deep learning network.
[0060] Specifically, this embodiment uses endometrial tissue samples as an example. Endometrial tissue samples can be obtained through endometrial aspiration biopsy, and these samples can be used to assess the health status of the endometrium.
[0061] In other embodiments, the test sample may also be a cell or test sample that can reflect other types of cancer, for example, it can be extended to the examination of other gynecological cancers, such as vaginal cancer, vulvar cancer, fallopian tube cancer, etc.
[0062] Viscosity fluorescent probes are fluorescent labels used to measure the viscosity of liquids. The fluorescence lifetime of these fluorescent labels is sensitive to changes in viscosity within the microenvironment. 9-(2,2-dicyanvinyl)jugolide (DCVJ) is one such viscosity fluorescent probe. Other fluorescent probes with similar functions include ANS, Bis-ANS, Nile Red, and CCVJ.
[0063] Since the viscosity of cancer cells and normal cells differs, viscosity fluorescent probes can be used to stain cells in a test sample, and the difference in the fluorescence lifetime of intracellular fluorescent markers can be used to identify cells in the process of cancer progression.
[0064] The steps for obtaining fluorescence lifetime imaging (FLIM) of the test sample include: preparing a cell smear from the stained test sample; scanning the cell smear to obtain raw fluorescence lifetime data; and performing multi-e exponential fitting statistics on the raw fluorescence lifetime data to obtain the fluorescence lifetime image of the test sample.
[0065] The process of scanning cell smears may include: projecting excitation light onto the stained test sample, which causes a viscoelastic fluorescent probe to fluoresce; collecting the fluorescence signal generated by the viscoelastic fluorescent probe; and obtaining time-based fluorescence lifetime data based on the fluorescence signal.
[0066] Fluorescence lifetime is defined as the time required for the fluorescence intensity peak to decay to 1 / e. When only one fluorescent molecule is present, the fluorescence lifetime exhibits a single-e exponential decay. When multiple components of the fluorophore are present, the fluorescence decay is the sum of the decays of the multiple exponential components, fitted using a multi-e exponential method.
[0067] The average fluorescence lifetime t of each pixel in the fluorescence lifetime image m : Among them, t i For the fluorescence lifetime of different components, a i It refers to the percentage of different components.
[0068] like Figure 2 As shown, the left side is a FLIM image of malignant endometrial tissue (Malignant), and the right side is a FLIM image of benign endometrial tissue (Benign), reflecting the distribution of average fluorescence lifetime values of benign and malignant tissues under FLIM microscopy. m The average fluorescence lifetime is represented by a corresponding color in the image, with different fluorescence lifetime values represented by different colors. FLIM images not only provide visual information about biological tissues or cells, but also reveal changes in the microenvironment of lesion areas through fluorescence lifetime values, which is beneficial for early screening.
[0069] To further enhance the differences between benign and malignant FLIM images, FLIM images need to be preprocessed, such as by grayscale broadening and / or grayscale histogram equalization to enhance image contrast, making details in the image clearer and thus improving the image's usability.
[0070] Gray-level widening enhances image contrast by stretching the range of gray levels. Gray-level histogram equalization (HQE) redistributes the gray levels of an image so that each gray level has a similar number of pixels, making the image's gray-level histogram as flat as possible, thus improving image contrast. Gray-level widening is more suitable for image processing that requires adjustment of specific gray levels, while gray-level histogram equalization is more suitable for scenarios that enhance overall contrast.
[0071] The preprocessed fluorescence lifetime image is input into a classification model based on a deep learning network. The classification model uses this image to predict whether the cellular tissue in the detection sample is benign or malignant, or the corresponding probability, etc.
[0072] In this embodiment, the entire detection process of the test sample is completed through staining, fluorescence lifetime imaging, and machine learning classification prediction. Compared with traditional methods, it has higher detection efficiency, lower cost, and higher repeatability, and avoids subjective bias. By using a viscosity fluorescent probe to stain the test sample, the viscosity difference within the cell is identified based on the difference in fluorescence lifetime. Compared with the traditional method of identification based on cell morphology, changes in the microenvironment can be detected earlier, improving the efficiency of early screening.
[0073] In one embodiment, step S100 includes:
[0074] Step S110: Add a fixative to fix the contents of the test sample;
[0075] Step S120: Add a first preset amount of antioxidant;
[0076] After centrifugation to remove the supernatant in step S130, a viscosity fluorescent probe is added to stain the contents.
[0077] In step S150, centrifuge again to remove the supernatant, and use a fixative to agitate and wash the precipitate.
[0078] Repeat step S150 a preset number of times to obtain the stained cell sample to be tested.
[0079] Specifically, 4% paraformaldehyde can be added to fix the contents of the test sample. Then, an appropriate amount of antioxidant, such as glutathione (GSH), is added. Antioxidants can scavenge free radicals and maintain the cell sample in a better state during preservation, while reducing the chance of staining agent quenching, thereby improving staining efficiency and the accuracy and reliability of the experiment. After centrifugation to remove the supernatant, a viscosity fluorescent probe, such as 50-150 μM DCVJ dilution buffer (diluent is 4% paraformaldehyde), is added to stain the contents for 20-50 minutes. Afterward, centrifugation is performed again to remove the supernatant, and the precipitate is washed with 4% paraformaldehyde by pipetting. This centrifugation-washing process is repeated a predetermined number of times (e.g., 1-5 times) to obtain the final cell sample to be tested.
[0080] In one embodiment, between step S130 and step S150, the following further step is included:
[0081] Step S140: Add a second preset amount of anti-quenching agent.
[0082] Specifically, an appropriate amount of anti-quenching agent is added, such as 1,4-diazobicyelo(2,2,2)-octane (DABCO). Anti-quenching agents can protect stained specimens and stabilize the luminescence of fluorescent molecules, thereby improving the reliability of the experiment.
[0083] In one embodiment, the preprocessing of the fluorescence lifetime image in step S300 includes:
[0084] Step S310: Remove defective images;
[0085] Step S320 improves the contrast of the qualified image;
[0086] Step S330 involves cell segmentation of the contrast-enhanced fluorescence lifetime image.
[0087] Specifically, methods such as grayscale broadening and / or grayscale histogram equalization can be used to improve image contrast, and U-Net neural networks can be used to perform cell segmentation on fluorescence lifetime images with improved contrast. Cell segmentation refers to separating individual cells or cell clusters from the background of an image, filtering out irrelevant background objects such as blood and mucus, so that only cells are visible in the image.
[0088] In one embodiment, cell segmentation in step S330 includes:
[0089] Step S331 uses an improved first U-Net network to denoise the fluorescence lifetime image with improved contrast;
[0090] Step S332 involves cell segmentation of the denoised image.
[0091] Among them, such as Figure 3As shown, the first U-Net network includes a downsampling part 1, an upsampling part 2, and an image style representation layer. The first U-Net network is based on the base U-Net network, replacing the standard convolutional layers (which consist of a 3x3 convolutional layer, batch normalization, and ReLU activation) with residual blocks, and adding an image style representation layer between the downsampling part 1 and the upsampling part 2. After the downsampling process, the image style representation layer performs global average pooling and normalization on the output feature map to obtain a 256-dimensional image style vector. This image style vector is then projected onto a dimension matching the number of channels in the feature map of the current upsampling layer and superimposed into the input of each upsampling layer via a broadcast mechanism.
[0092] like Figure 3 As shown, both the upsampling and downsampling parts consist of four spatial scales (corresponding to four different layers in the figure). Each spatial scale consists of two cascaded residual blocks, and each residual block includes two 3*3 convolutional layers. Thus, each spatial scale has four convolutional layers, while the basic U-Net network only has two convolutional layers per spatial scale. This improvement doubles the number of convolutional layers, thereby increasing the model capacity and performance.
[0093] After the downsampling process is completed, the image style features obtained by global average pooling (i.e., the features reflected by the image style vector) are superimposed on all inputs of the upsampling layer, so that the model can perform adaptive calculation according to the style of different images, thereby improving the generalization of the model.
[0094] In one embodiment, the image style vector is superimposed onto the input of the second convolutional layer of the first residual block of each upsampling layer. For example... Figure 3 As shown, the arrows in the image style vectors point to the second convolutional layer of the first residual block of each upsampling layer.
[0095] This allows for the acquisition of low-dimensional features of the image through the processing of the first convolutional layer of the first residual block of each upsampling layer, followed by the combination of the image style vector obtained from the overall feature map for the next convolution, avoiding excessive extraction of low-level features while controlling the scale of the computational parameters.
[0096] In one embodiment, cell segmentation in step S330 includes:
[0097] Step S331 uses an improved first U-Net network to denoise the fluorescence lifetime image with improved contrast;
[0098] Step S333 uses an improved second U-Net network to perform cell segmentation on the denoised image;
[0099] The first U-Net network and the second U-Net network have the same structure, only their parameters differ. The structure of the first U-Net network is as follows: Figure 3 As shown.
[0100] In one embodiment, training the second U-Net network includes:
[0101] The unlabeled sample images are segmented using the trained base U-Net model, and the segmentation results are used as segmentation labels.
[0102] The second U-Net network is trained using sample images labeled with the segmentation tag.
[0103] This allows for self-supervised learning transfer on new datasets based on a pre-trained base U-Net model, generating training supervision terms for an improved U-Net network without the need for manual labeling of segmentation tags, thus reducing the manual labeling of training samples for the improved U-Net network.
[0104] In one embodiment, such as Figure 4 As shown, the construction of the classification model includes:
[0105] Step S500: Prepare the sample dataset, which contains several samples. Each sample is a fluorescence lifetime image with coarse labels, including high-risk and low-risk labels; initialize the sample deletion rate and prediction accuracy.
[0106] Step S510: Construct the deep learning network;
[0107] Step S520 uses a deep learning network to perform cross-validation on the sample dataset;
[0108] Step S530: Based on the results of cross-validation, calculate the new prediction accuracy, use the confidence learning algorithm to identify and delete samples containing incorrect labels, calculate the sample deletion rate, and update the sample dataset.
[0109] In step S540, relative to the previous training results, determine whether the new sample deletion rate no longer decreases and whether the new prediction accuracy no longer increases. If so, proceed to step S550; otherwise, proceed to step S520.
[0110] Step S550 uses the trained deep learning network as a classification model.
[0111] Specifically, a classification model is built based on a deep learning network. The sample dataset is used to train this deep learning network, and the trained deep learning network serves as the classification model.
[0112] The samples can be derived from case studies, and each sample is labeled. The label design principle is relatively simple; for example, all FLIM images from cancer patients are labeled as high-risk, and all FLIM images from benign cases are labeled as low-risk. However, this method of labeling may not be accurate. To improve the accuracy and robustness of the classification model, mislabeled samples (i.e., noise samples) need to be removed.
[0113] Cross-validation, such as K-fold cross-validation where K is a positive integer, provides the probability that each sample in the dataset is predicted to belong to each class, as well as the class corresponding to the highest probability (denoted as the predicted class). Based on the labeled class and predicted class of each sample in the dataset, the prediction accuracy is calculated. Prediction accuracy is the proportion of samples whose labeled class matches the predicted class during internal cross-validation out of the total number of samples.
[0114] Confidence learning, based on the results of cross-validation and a confidence threshold, identifies noisy samples (i.e., mislabeled samples) and removes them to improve the accuracy of the sample dataset. The sample deletion rate is then calculated based on the results of confidence learning. The sample deletion rate is the proportion of deleted samples out of the total number of samples.
[0115] Retrain the deep learning network using the updated sample dataset, and continue to use cross-validation on the updated sample dataset to obtain a new prediction accuracy and a new sample deletion rate. If, relative to the previous results, the new sample deletion rate no longer decreases and the new prediction accuracy no longer increases, then the sample dataset is no longer cleaned. The deep learning network trained at this point is considered a well-trained network and can be used as a classification model.
[0116] Deep learning networks can be residual convolutional neural networks such as ResNet, GoogLeNet, AlexNet, VGG16, or Inception.
[0117] By combining confidence learning with deep learning through the above steps, we can effectively clean up erroneous labels in the dataset, build a more accurate dataset, and train a more robust and accurate classification model, thereby improving the accuracy and reliability of medical image classification.
[0118] One embodiment of the present invention, such as Figure 5 As shown, a fluorescence lifetime-based detection device includes:
[0119] The staining module 100 is used to acquire the test sample and stain it with a viscosity fluorescent probe;
[0120] The imaging module 200 is used to perform fluorescence lifetime imaging processing on the stained test sample to obtain a fluorescence lifetime image of the test sample.
[0121] The preprocessing module 300 is used to preprocess the fluorescence lifetime image;
[0122] The classification module 400 is used to input the preprocessed fluorescence lifetime image into the classification model to classify and predict the detected samples; the classification model is based on a deep learning network.
[0123] In one embodiment, the staining module is further configured to: add a fixative to fix the contents of the test sample; add a first preset amount of antioxidant; centrifuge to remove the supernatant and then add a viscosity fluorescent probe to stain the contents; centrifuge again to remove the supernatant and use a fixative to blow and wash the precipitate; repeat the centrifugation-washing process a preset number of times to obtain the stained cell sample to be tested.
[0124] In one embodiment, the staining module is further configured to: add a fixative to fix the contents of the test sample; add a first preset amount of antioxidant; after centrifugation to remove the supernatant, add a viscosity fluorescent probe to stain the contents; add a second preset amount of anti-quenching agent; centrifuge again to remove the supernatant, and use the fixative to blow and wash the precipitate; repeat the centrifugation-washing process a preset number of times to obtain the stained cell sample to be tested.
[0125] In one embodiment, the preprocessing module includes:
[0126] The initial screening unit is used to remove unqualified images;
[0127] The contrast enhancement unit is used to enhance the contrast of a qualified image.
[0128] A cell segmentation unit is used to segment cells in contrast-enhanced fluorescence lifetime images.
[0129] In one embodiment, the cell segmentation unit is further configured to denoise the fluorescence lifetime image with improved contrast using an improved first U-Net network; and to perform cell segmentation on the denoised image.
[0130] In one embodiment, the cell segmentation unit is further configured to denoise the contrast-enhanced fluorescence lifetime image using an improved first U-Net network; and to perform cell segmentation on the denoised image using an improved second U-Net network.
[0131] In one embodiment, the cell segmentation unit is further configured to segment unlabeled sample images using a trained base U-Net model, and use the resulting segmentation as a segmentation label; and use the sample images labeled with the segmentation label to train a second U-Net network.
[0132] In one embodiment, a building module is also included.
[0133] The module is used to: prepare a sample dataset containing several samples, each a fluorescence lifetime image with coarse labels, including high-risk and low-risk labels; initialize the sample deletion rate and prediction accuracy; build a deep learning network; perform cross-validation on the sample dataset using the deep learning network; calculate the prediction accuracy based on the cross-validation results, use a confidence learning algorithm to identify and delete samples containing incorrect labels, calculate the sample deletion rate again, and update the sample dataset; retrain the deep learning network with the updated sample dataset, repeating the above process of cross-validation and deleting samples containing incorrect labels until the sample deletion rate no longer decreases and the prediction accuracy no longer increases compared to the previous training results; and use the trained deep learning network as a classification model.
[0134] It should be noted that the embodiments of the fluorescence lifetime-based detection device provided by this invention and the embodiments of the fluorescence lifetime-based detection method provided above are all based on the same inventive concept and can achieve the same technical effects. Therefore, other specific details of the embodiments of the fluorescence lifetime-based detection device can be referred to the description of the embodiments of the fluorescence lifetime-based detection method described above.
[0135] Another embodiment of the present invention, such as Figure 6 As shown, a fluorescence lifetime-based detection device includes:
[0136] Memory 10 is used to store computer program 20;
[0137] The processor 30 is used to implement the fluorescence lifetime-based detection method described in any of the foregoing embodiments when running the computer program 20.
[0138] The memory 10 can be any internal storage unit and / or external storage device capable of storing data and programs. For example, the memory 10 can be a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD card), or a flash memory card.
[0139] Depending on the requirements, the processor 10 may be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a general-purpose processor, or other logic devices.
[0140] In one embodiment of the present invention, a computer-readable storage medium stores a computer program thereon. When executed by a processor, the computer program can implement the fluorescence lifetime-based detection method described in the foregoing embodiments. That is, when part or all of the technical solutions contributing to the prior art in the foregoing embodiments of the present invention are embodied in a computer software product, the aforementioned computer software product is stored in a computer-readable storage medium. The computer-readable storage medium can be any physical device or apparatus capable of carrying computer program code. For example, the computer-readable storage medium can be a USB flash drive, a portable hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, etc.
[0141] It should be noted that the above embodiments can be freely combined as needed. The above description is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A detection method based on fluorescence lifetime, characterized in that, include: Obtain the test sample and stain it using a viscosity fluorescent probe; Fluorescence lifetime imaging was performed on the stained test samples to obtain fluorescence lifetime images of the test samples; The fluorescence lifetime image is preprocessed; The preprocessed fluorescence lifetime image is input into the classification model to perform classification prediction on the detected samples; The classification model is based on a deep learning network.
2. The detection method based on fluorescence lifetime according to claim 1, characterized in that, The process of obtaining the test sample and staining it with a viscosity fluorescent probe includes: A fixative is added to fix the contents of the test sample; Add the first preset amount of antioxidant; After centrifugation to remove the supernatant, a viscosity-sensitive fluorescent probe was added to stain the contents. Centrifuge again to remove the supernatant, and use the fixative to wash the precipitate by blowing and washing. Repeat the centrifugation-washing process a preset number of times to obtain the stained cell sample to be tested.
3. The detection method based on fluorescence lifetime according to claim 2, characterized in that, After staining and before centrifugation to remove the supernatant, the process includes: Add a second preset amount of anti-quenching agent.
4. The detection method based on fluorescence lifetime according to claim 1, characterized in that, The preprocessing performed on the fluorescence lifetime image includes: The contrast of the fluorescence lifetime image is improved by using grayscale broadening and / or grayscale histogram equalization. Cell segmentation was performed on the fluorescence lifetime images with improved contrast.
5. The detection method based on fluorescence lifetime according to claim 4, characterized in that, The cell segmentation of the contrast-enhanced fluorescence lifetime image includes: An improved first U-Net network is used to denoise the fluorescence lifetime image with improved contrast; Cell segmentation is performed on the denoised image; The first U-Net network is based on the base U-Net network, replacing the standard convolutional layers in the base U-Net network with residual blocks, and adding an image style representation layer between model downsampling and upsampling. After the downsampling process, the image style representation layer performs global average pooling and normalization on the obtained feature map to obtain a 256-dimensional image style vector. The image style vector is projected onto a dimension that matches the number of channels of the feature map of the current upsampling layer, and superimposed into the input of each upsampling layer through a broadcast mechanism.
6. The fluorescence lifetime-based detection method according to claim 5, characterized in that, The cell segmentation of the denoised image includes: An improved second U-Net network is used to perform cell segmentation on the denoised image; The first U-Net network and the second U-Net network have the same structure.
7. The fluorescence lifetime-based detection method according to claim 6, characterized in that, Training the second U-Net network includes: The unlabeled sample images are segmented using the trained base U-Net model, and the segmentation results are used as segmentation labels. The second U-Net network is trained using sample images labeled with the segmentation tags.
8. The detection method based on fluorescence lifetime according to claim 1, characterized in that, The construction of the classification model includes: Prepare a sample dataset containing several samples, each of which is a fluorescence lifetime image carrying coarse labels, including high-risk and low-risk labels; Build deep learning networks; The deep learning network was used to perform cross-validation on the sample dataset. Based on the results of the cross-validation, the prediction accuracy is calculated, a confidence learning algorithm is used to identify and delete samples containing incorrect labels, the sample deletion rate is calculated, and the sample dataset is updated. The deep learning network is retrained based on the updated sample dataset. The process of cross-validation and removing samples containing incorrect labels is repeated until the sample removal rate no longer decreases and the prediction accuracy no longer increases compared to the previous training results. The trained deep learning network is used as the classification model.
9. A detection device based on fluorescence lifetime, characterized in that, include: Memory, used to store computer programs; A processor for implementing the fluorescence lifetime-based detection method as described in any one of claims 1 to 8 when running the computer program.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the fluorescence lifetime-based detection method as described in any one of claims 1-8.