Autonomous data set training-based nnUNet zebra fish juvenile fish whole cerebral vessel system segmentation method

Abstract

The invention discloses an nnUNet zebra fish juvenile fish whole-brain blood vessel system segmentation method based on autonomous data set training, and relates to the technical field of high-resolution imaging technology, image processing and medical image segmentation. According to the method, high-resolution whole-brain three-dimensional blood vessel image data is obtained by fully utilizing the transparency and fluorescence labeling advantages of a zebra fish living body; and a high-quality segmentation truth value database is constructed through semi-automatic segmentation and manual correction, and an nnU-Net deep learning model is adopted for training, so that three-dimensional automatic segmentation of zebra fish cerebrovascular system signals is realized. According to the method, the automation degree and precision of image segmentation are remarkably improved, and the problems of low efficiency, high manual dependency, poor repeatability and the like of traditional cerebral vessel segmentation are effectively solved. The method is suitable for large-scale high-throughput data processing, can provide an efficient and standardized image processing scheme for zebra fish cerebrovascular development mechanism and cerebrovascular disease model research, and has a wide application prospect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of biomedical image processing and computer vision, and particularly relates to a nnUNet zebrafish larva whole brain blood vessel system segmentation method based on autonomous data set training. BACKGROUND

[0002] The quantitative evaluation of human brain vascular phenotype research relies on naked eye observation and manual measurement, and three-dimensional reconstruction and quantification are difficult to achieve a balance in imaging quality and image segmentation efficiency, and a breakthrough in the technical level is needed 1-2 . Therefore, by screening suitable model animals, establishing a complete brain blood vessel system recognition and reconstruction and quantitative analysis platform, and comprehensively describing the physiological structure and disease phenotype of brain blood vessels, a technical strategy can be provided for brain blood vessel research to break through the current technical bottleneck of brain blood vessel research.

[0003] Whole brain blood vessel imaging and quantitative analysis rely on high-resolution microscopic imaging technology, combined with blood vessel lumen recognition, filling and structure reconstruction technology to obtain three-dimensional data. In recent years, the application of deep machine learning in medical image processing has significantly improved the accuracy of blood vessel segmentation. Among them, nnUNet (no-new-U-Net) is an automated medical image segmentation framework proposed by Isensee et al. in 2018, which is based on U-Net and has undergone a series of optimizations, making it perform well in various medical image segmentation tasks 3-4 . Blood vessel three-dimensional reconstruction relies on high-precision blood vessel signal segmentation, but due to signal noise interference and tissue complexity, existing methods still have limitations in non-target signal filtering. Although deep machine learning models can improve medical image segmentation efficiency, the identification of low-contrast blood vessels still has challenges, and the training process relies on high-quality labeled data. In addition, after whole brain blood vessel reconstruction, professionals with anatomical knowledge are still needed to perform manual evaluation and correction, making the data processing process complex and time-consuming.

[0004] Zebrafish, as a vertebrate model organism, has great advantages in the study of vascular system development-related mechanisms. First, the short breeding cycle and single high-throughput operation of zebrafish are conducive to genetic research such as phenotype identification and mutant construction through CRISPR / Cas9 gene editing technology and Morpholino-mediated gene knockdown technology. Second, compared to other species that require complex procedures such as tissue transparency and special staining for brain specimens, or angiography or tissue exposure before live imaging 5-6 . The whole transparent nature of zebrafish at the development stage allows for dynamic observation of whole brain blood vessel development through non-invasive live tissue imaging, making zebrafish an irreplaceable advantage in the application of live experimental imaging technology. At the same time, zebrafish has the same blood-brain barrier structure components as mice and humans7 And the multiple zebrafish fluorescent protein specific labeling of vascular system transgenic fish line can achieve high-resolution live imaging of vascular system and even subcellular structure, which helps to understand the mechanism of vascular formation at the level of cell biology. Therefore, zebrafish is an ideal model animal for current high-throughput non-invasive dynamic imaging of whole brain vasculature.

[0005] Based on the current research status, it is feasible to use live zebrafish to achieve precise and efficient reconstruction of whole brain vasculature and quantification, which is an important technical breakthrough for brain disease and cerebral vascular phenotype research.

[0006] REFERENCES

[0007] 1. Barak, T. et al. PPIL4 is essential for brain angiogenesis and implicated in intracranial aneurysms in humans [J]. Nat Med, 2021, 27: 2165-2175.

[0008] 2. Kirst, C. et al. Mapping the Fine-Scale Organization and Plasticity of the Brain Vasculature [J]. Cell, 2020, 180: 780-795.e25.

[0009] 3. Ronneberger, O. et al. U-net: Convolutional networks for biomedical image segmentation [C]. International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015: 234-241.

[0010] 4. Isensee, F. et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation [J]. Nat Methods, 2021, 18: 203-211.

[0011] 5. Zhang, Z. Z. et al. Multi-Scale Light-Sheet Fluorescence Microscopy for Fast Whole Brain Imaging[J]. Front Neuroanat, 2021, 15: 732464.

[0012] 6. Kiemen, A. L. et al. CODA: quantitative 3D reconstruction of large tissues at cellular resolution[J]. Nat Methods, 2022, 19: 1490-1499.

[0013] 7. O'Brown, N. M. et al. Bridging barriers: a comparative look at the blood-brain barrier across organisms[J]. Genes Dev, 2018, 32: 466-478. SUMMARY

[0014] The present application aims to solve the problems of low efficiency, insufficient accuracy, difficulty in recognizing deep blood vessels and high dependence on artificial methods in existing zebrafish brain blood vessel image segmentation methods. The present application proposes an automatic segmentation method for zebrafish larva whole brain blood vessel system to improve the accuracy and processing efficiency of three-dimensional image segmentation and promote the high-throughput and standardized development of brain blood vessel research.

[0015] Technical solution: To achieve the purpose of the present application, the technical solution adopted by the present application is:

[0016] A zebrafish larva whole brain blood vessel system segmentation method based on nnUNet model trained by autonomous data set, a standardized image segmentation analysis process is constructed to realize three-dimensional automatic segmentation and digital reconstruction of intracranial blood vessel system of zebrafish. The method comprises the following steps:

[0017] Step 1, using endothelial cell specific fluorescent labeled transgenic zebrafish, performing Z-axis compensation high resolution three-dimensional imaging of intracranial blood vessels;

[0018] Step 2, combining semi-automatic image segmentation and correction strategy, obtaining real label map of intracranial blood vessel system, and constructing standardized blood vessel segmentation data set for training;

[0019] Step 3, based on the constructed dataset, the nnUNet architecture is applied for model training to realize zebrafish brain blood vessel full-supervised segmentation prediction, and 3D mask results for whole brain blood vessel digital reconstruction are obtained.

[0020] Further, step 1 comprises:

[0021] Firstly, during the process of applying in vivo endothelial cell transgenic line zebrafish to perform high-resolution imaging on the cranial vascular system, confocal microscopy is used to perform fluorescence imaging on the whole brain of the zebrafish, and deep brain blood vessel signals are obtained;

[0022] The zebrafish is fixed in a dorsal position by agarose gel, and the brain imaging front-to-back diameter is set under a 20x water lens: the rostral side includes the olfactory region, the nasal artery, and the nasal vein transition area, and the caudal side includes the first annular blood vessel at the junction of the basilar artery and the spinal cord blood vessels;

[0023] The Z-axis range is set: the dorsal side is marked by the dorsal longitudinal vein, and the ventral side is marked by the first two branchial arches merging into the primitive internal jugular vein, and the Z-axis shooting step is set;

[0024] In the Z-axis correction mode, the top, middle and bottom positions are automatically identified according to the current full Z-axis range, and the imaging parameters of the three layers are set;

[0025] During the shooting process, the full brain blood vessel fluorescence signal is imaged in an automatic incremental mode to the target value, and by establishing a dynamic parameter incremental setting of Z-axis fluorescence excitation intensity, high-resolution imaging of the whole brain blood vessels of zebrafish is realized.

[0026] Further, the dorsal longitudinal vein is used as the starting layer of imaging, and the top layer excitation light intensity ELI Top is first set, and then the imaging parameters of the three layers are set as follows:

[0027] HV Top = ELI Top + a;

[0028] ELI Middle = ELI Top / mt1*mt2, HV Middle = ELI Middle + b;

[0029] ELI Botton = ELI Middle / bm1*bm2, HV Botton = ELI Botton + c;

[0030] Wherein, ELI Top and HV Top are the top layer excitation light intensity and the top layer high voltage value, respectively, ELI Middle and HVMiddle respectively are the middle layer excitation light intensity and the middle layer high voltage value, ELI Botton and HV Botton respectively are the bottom layer excitation light intensity and the bottom layer high voltage value, a, b, c, mt1, mt2, bm1, bm2 are self-defined parameters.

[0031] Further, step 2 applies a semi-automatic joint manual correction image segmentation step to obtain an intracranial blood vessel system segmentation true value and a data set, including:

[0032] The Fiji_Labkit plug-in is applied to semi-automatically segment the brain blood vessel signal of the Z-axis corrected imaging, sparse blood vessel markers are manually added to the original signal, including filling and tube wall marking of the blood vessel as foreground, and sparse background signal is given, interactive image label iterative training is performed, and a binary segmentation image is obtained;

[0033] ITK_SNAP is applied to correct the filling state of the brain blood vessel of all layers of the binary segmentation image, central blood vessel signals are removed through discrimination, and a full true value segmentation image stack of the intracranial blood vessel system is obtained.

[0034] Further, in step 3, the nnUNet dynamic adaptive network structure is adopted, and the network includes a convolutional layer, a 3D maximum pooling layer, a deconvolutional layer and a skip connection;

[0035] 3D patch input is adopted, data enhancement is performed in multiple data enhancement modes, the preprocessed data is blocked and nnUNet is trained, and the optimizer is an SGD with Nesterov momentum;

[0036] The initial learning rate and the upper limit of training are set, the poly decay strategy is adopted for adjustment, the model with the optimal verification performance is automatically saved, and the trained model is used for image prediction and output of the merged 3D reconstruction result.

[0037] Beneficial effects: compared with the prior art, the technical scheme of the present application has the following beneficial technical effects:

[0038] The method of the present application first realizes the standardized three-dimensional image segmentation process of the whole brain blood vessel system of the model animal zebrafish larvae, effectively improves the deep blood vessel imaging quality and recognition accuracy, realizes the brain external signal removal and brain blood vessel structure complete extraction through self-labeling data, significantly improves the data set accuracy, realizes the efficient and automatic processing of image segmentation based on the self-data set, and provides phenotype research technical support for brain blood vessel development and disease research. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a schematic diagram of the total process of ZICVesSP.

[0040] Figure 2 is a schematic diagram of Z-axis correction imaging effect.

[0041] Figure 3 is a schematic diagram of semi-automatic segmentation and correction.

[0042] Figure 4 is a schematic diagram of nnUNet_kdrl training model and prediction process.

[0043] Figure 5 is the evaluation of nnUNet_flt1 training model and prediction effect established by transfer learning strategy. DETAILED DESCRIPTION

[0044] The technical solutions of the present application will be further described below in combination with the drawings and examples.

[0045] The present application provides a zebrafish larva whole brain vascular system segmentation method based on nnUNet model trained by autonomous dataset, which is used to realize high-precision automatic segmentation of zebrafish larva whole brain vascular system. A standardized image segmentation analysis process (Zebrafish Intracranial Vasculature Segmentation and Standardization Pipeline, ZICVesSP) is constructed according to the three-dimensional imaging characteristics of zebrafish intracranial blood vessels, as shown in A of Figure 1 , Figure 1 A shows Z-axis correction confocal 3D imaging of endothelial transgenic live zebrafish intracranial blood vessels. Figure 1 B shows that the intracranial vascular system segmentation ground truth and dataset are obtained by applying semi-automatic combined with manual correction image segmentation steps. Figure 1 C shows that through the construction of dataset (ZICVesDataset), deep learning training is realized, and full supervision 3D mask prediction based on nnUNet model is realized. The whole process realizes the automatic processing framework from the import of original image to the fine segmentation of vascular structure by realizing the image format (.ims / .tif / .nii) conversion between software through the code. Three-dimensional automatic segmentation and digital reconstruction of zebrafish intracranial vascular system are realized. Using endothelial cell specific fluorescently labeled transgenic zebrafish, Z-axis compensation high-resolution three-dimensional imaging of intracranial blood vessels is carried out; combined with semi-automatic image segmentation and correction strategy, the real label map of intracranial vascular system is obtained, and the standardized vascular segmentation dataset for training is constructed; based on the constructed dataset, nnUNet architecture is applied for model training, zebrafish brain vascular full supervision segmentation prediction is realized, and 3D mask result for whole brain vascular digital reconstruction is obtained.

[0046] First, in the process of high-resolution imaging of the cranial vasculature using the live endothelial cell transgenic line zebrafish Tg(kdrl:ras-mCherry), in order to enhance the acquisition of deep brain vascular signals, confocal microscopy was used to perform whole-brain fluorescence imaging of zebrafish. After anesthesia, the zebrafish was fixed as a dorsal side by agarose gel, and the brain imaging front-to-back diameter was set under a 20x water lens: the rostral side included the olfactory region, the nasal artery-nasal vein transition area, and the caudal side included the first annular blood vessels at the junction of the basilar artery and spinal cord blood vessels; the Z-axis range: the dorsal side was marked by the dorsal longitudinal vein, and the ventral side was marked by the first two gill arches merging into the primitive internal jugular vein, and the Z-axis shooting step was 2 / 3 pm. In the Z-axis correction mode, the system automatically identifies the top, middle and bottom positions according to the current full Z-axis range. The imaging parameters such as excitation light intensity (ELI) and high voltage (HV) value were set: taking the dorsal longitudinal vein as the starting layer of imaging, first set the top excitation light intensity ELI Top value range is 20-25 mV, and then set the three-layer imaging parameters as follows:

[0047] HV Top = ELI Top + a;

[0048] ELI Middle = ELI Top / mt1*mt2, HV Middle = ELI Middle + b;

[0049] ELI Botton = ELI Middle / bm1*bm2, HV Botton = ELI Botton + c; wherein ELI Top and HV Top are the top excitation light intensity and the top high voltage value, ELI Middle and HV Middle are the middle excitation light intensity and the middle high voltage value, ELI Botton and HV Botton are the bottom excitation light intensity and the bottom high voltage value, and the custom parameters a=5, b=5, c=5, mt1=100, mt2=130, bm1=100, bm2=140. Pinhole: 1.2 AU; Frame Average: 4x. During the shooting process, the system images the whole brain vascular fluorescence signal in the mode of automatically increasing to the target value. As shown in A of Figure 1 , by establishing a dynamic parameter incremental setting of Z-axis fluorescence excitation intensity, high-resolution imaging of the whole brain vasculature of zebrafish is achieved.

[0050] Figure 2 A in FIG. 15 shows the signal intensity topography (Interactive 3D Surface Plot) of the fluorescence vascular signal intensity spatial distribution after excitation light correction and correction imaging. The fluorescence vascular signal intensity spatial distribution of the control group and the Z-axis correction imaging was displayed by the fluorescence intensity topography, and the shallow layer (6th layer / Z6), the middle layer (50th layer / Z50), and the deep layer (100th layer / Z100) of the Z-axis full range were visualized. After correction imaging, the fluorescence signal area quantitative results showed that more signals were obtained in the middle and deep vascular running positions.

[0051] Figure 2 B in FIG. 15 shows the average fluorescence intensity of different optical section layers (6 cases / group) of the control group and the Z-axis correction imaging. The data is expressed as mean ± standard error, and the statistical significance is determined by two-way ANOVA combined with Tukeys' post-test. (*P<0.05, ****P<0.0001). The average fluorescence gray value of the Z-axis correction mode increased in the middle of the imaging range (Z50), the average fluorescence gray value of the deep layer (Z75, Z100) was significantly compensated, and the deep endothelial fluorescence signal intensity was similar to that of the shallow layer.

[0052] In order to obtain the segmentation image of the cranial vascular system, Figure 1 B in FIG. 15 is a schematic diagram of obtaining the segmentation true value and data set of the intracranial vascular system by applying semi-automatic combined with manual correction image segmentation steps.

[0053] Figure 3 A in FIG. 15 is a manual operation of the medical image processing software ITK_SNAP to correct the filling state of all layers of brain blood vessels, and two medical experts manually remove the central blood vessel signals such as eyes, gill arch, heart, and fin after identification to obtain a full true value segmentation image stack of the intracranial vascular system. Figure 3 A in FIG. 15 shows the semi-automatic segmentation of the brain blood vessel signal of the Z-axis correction imaging by applying the shallow learning tool Fiji_Labkit plug-in. The original signal (left graph-bottom left) is manually added with sparse blood vessel markers (green), including the filling of the blood vessels (green) and the tube wall (yellow), which are marked as foreground, and the sparse background signal (gray) (left graph-bottom right) is given, and the binary segmentation image (right graph) is obtained by interactive image label iterative training (middle graph). Figure 3 B in FIG. 15 applies ITK-SNAP to manually evaluate and correct the binary segmentation image layer by layer to obtain the segmentation true value, removes the non-intracranial vascular signal of the optical section image under the free perspective (yellow line box), and displays the overall effect of the fluorescence signal segmentation (bottom left) and the intracranial vascular system (bottom right) under the 3D perspective.

[0054] In order to improve the efficiency of image segmentation operation, the application takes the sample original image and the full true value image as a pair of data groups Figure 1 A, 1B-dotted line box) to construct a zebrafish larva whole brain blood vessel data set, the data set has 105 data groups, covering each development cycle and physiological and pathological model samples, as shown in Table 1.

[0055] Table 1 nnUNet_kdrl model training data set

[0056]

[0057]

[0058] Randomly select 5 groups of wild type samples of each development cycle as a prediction set for training model evaluation, 100 groups of samples are trained by 5-fold cross validation strategy, 80 groups are used as a training set, and 20 groups are used as a verification set, to ensure that all samples participate in the verification process. Figure 1 C in the figure shows that the deep learning training is carried out by constructing the data set, and the full supervision brain blood vessel 3D mask prediction based on the nnUNet model is realized.

[0059] Figure 4 A in the figure is the official nnUNet (v2.2.0) preprocessing step and architecture, which adapts the network structure and patch size, uses 3D patch input, integrates various data enhancement methods (rotation, translation, noise reduction, blur, brightness and contrast). The optimizer is SGD with Nesterov momentum, the initial learning rate is 0.01, and the poly decay strategy is used for adjustment. The upper limit of each fold training is 1000 epochs, and the model with the optimal validation performance is automatically saved. The training process is realized by command line automation, which ensures the reproducibility and standardization of the experiment.

[0060] Figure 4 A in the figure shows the fluorescence image processing and prediction workflow of the nnUNet model. The process starts from image preprocessing, including grayscale conversion, font transformation and data enhancement techniques. The preprocessed data is blocked and used for training of the nnUNet architecture, which includes convolutional layers, 3D maximum pooling, deconvolutional layers and skip connections to achieve efficient feature extraction. The trained model is used for image prediction and outputs the merged 3D reconstruction results.

[0061] Figure 4B in FIG. 6B shows the training curve of nnUNet_kdrl, which shows that during the training process of 1000 epochs, the accuracy curve gradually rises and gradually stabilizes, the training loss and validation loss correspondingly show a downward trend, and the Dice coefficient tends to be stable, indicating that the model segmentation performance improves and converges. The model training is completed on a high-performance workstation equipped with NVIDIA RTX4090 GPU, and the running environment is Python 3.8, PyTorch 1.10.

[0062] Figure 4 B in FIG. 6B shows the training curve of nnUNet_kdrl, which shows that during the training process of 1000 epochs, the accuracy curve gradually rises and gradually stabilizes, the training loss and validation loss correspondingly show a downward trend, and the Dice coefficient tends to be stable, indicating that the model segmentation performance improves and converges. The model training is completed on a high-performance workstation equipped with NVIDIA RTX4090 GPU, and the running environment is Python 3.8, PyTorch 1.10.

[0063] The nnUNet_kdrl training model is a segmentation model constructed for the red fluorescent fish line with kdrl as the endothelial promoter. The nnUNet_kdrl model is used to predict the red fluorescent signal of the vascular endothelium with kdrl as the promoter, and the model prediction effect visualization is as shown in Figure 5 B in FIG. 6B shows the training curve of nnUNet_kdrl, which shows that during the training process of 1000 epochs, the accuracy curve gradually rises and gradually stabilizes, the training loss and validation loss correspondingly show a downward trend, and the Dice coefficient tends to be stable, indicating that the model segmentation performance improves and converges. The model training is completed on a high-performance workstation equipped with NVIDIA RTX4090 GPU, and the running environment is Python 3.8, PyTorch 1.10. Figure 5B, C1-C3 visualization results in B, C1-C3 show over-identification of eye and heart signals (C1-C3, nnUNet_kdrl model predicts Tg(flt1BCA:YFP) channel brain blood vessel signals; D1-D3, nnUNet_flt1 model predicts Tg(kdrl:ras-mCherry) channel brain blood vessel signals; E1-E3, nnUNet_flt1 model predicts Tg(flt1BCA:YFP) channel brain blood vessel signals. The segmented blood vessel signals are yellow, including missing signals (white line frame) and misidentified signals (blue line frame)), and Table 3 shows that the Dice coefficient and IOU decrease, and the HD95 significantly increases, indicating that the segmentation quality decreases, suggesting that the model has insufficient generalization ability.

[0064] By developing a transfer learning strategy, taking advantage of the high-throughput construction of zebrafish double fluorescent markers, based on the nnUNet_kdrl training model that has been constructed, different endothelial fish dataset and training model can be efficiently constructed to compensate for the generalization ability: first, construct the double transgenic fish line Tg(kdrl:ras-mCherry::flt1BCA:YFP), so that one fish can express red and green endothelial fluorescent signals. Apply the nnUNet_kdrl model based on the established self-built dataset prediction platform to predict the red endothelial brain blood vessel signal segmentation map, then use the segmentation map on the same fish to correct the green signal as the bottom map to obtain the matched segmentation map of the green signal, and establish the segmentation database of the green endothelial signal. Figure 5 A, A1-A3 in A, A1-A3 are further obtained double transgenic zebrafish line Tg(kdrl:ras-mCherry::flt1BCA:YFP) double-channel confocal imaging (A, A1-A3, endothelial transgenic zebrafish line Tg(kdrl:ras-mCherry::flt1BCA:YFP) confocal double-channel imaging). For the red fluorescent brain blood vessel signal labeled by the kdrl promoter, the nnUNet_kdrl model is used for segmentation prediction. And use this segmentation map as a template, two medical professionals correct the green fluorescent brain blood vessel signal labeled by the flt1 promoter for the same sample, obtain the intracranial blood vessel system (remove non-brain blood vessels) segmentation true value of the flt1 promoter, establish the training dataset as shown in Table 2, and further obtain the nnUNet_flt1 segmentation prediction model.

[0065] Table 2 nnUNet_flt1 model training dataset

[0066]

[0067] Figure 5B, D1-D3 are the nnUNet_flt1 model predicted kdrl as promoter vascular endothelial red fluorescence signal; E1-E3 are the nnUNet_flt1 model predicted brain flt1 as promoter vascular endothelial green fluorescence signal. Figure 5 C in the figure is the obtained nnUNet_flt1 training curve shows (blue line is the loss curve loss training (loss_tr), the red line is the training set loss validation (loss_val), the green line is the validation set Directly calculated Dice coefficient (pseudo dice (mov.age.)), during the training process of 1000 epochs, the accuracy curve gradually rises and gradually stabilizes, the training loss and the validation loss correspondingly present the downward trend, with the training of the Dice coefficient tends to be stable, prompting the model segmentation performance and convergence.

[0068] Table 3 nnUNet model prediction evaluation

[0069]

[0070] Two fish series intracranial vascular prediction models are provided in this embodiment, which are nnUNet_kdrl model and nnUNet_flt1 model respectively. The prediction results of the prediction set 5 examples / group samples are shown in Table 3 for segmentation prediction effect evaluation. Compared with the segmentation prediction results of flt1 fish series by applying nnUNet_kdrl model, the Dice coefficient and IOU of the segmentation prediction results of flt1 fish series by applying nnUNet_flt1 model are improved, and the HD95 is reduced, prompting that through the transfer learning strategy, the nnUNet_flt1 model for realizing accurate segmentation of flt1 fish series can be efficiently obtained.

Claims

1. A method for training nnUNet based on autonomous dataset, which is characterized in that, The method comprises the following steps: Step 1: Z-axis compensation high-resolution three-dimensional imaging of the cerebral vascular system of a zebrafish is performed by using an endothelial cell-specific fluorescently labeled transgenic zebrafish; Step 2: Real label images of the intracranial vascular system are obtained by combining a semi-automatic image segmentation and correction strategy, and a standardized vascular segmentation data set for training is constructed; Step 3: Based on the constructed data set, a model is trained by using an nnUNet architecture, full-supervised segmentation prediction of the zebrafish brain blood vessels is realized, and 3D mask results for digital reconstruction of the whole brain blood vessels are obtained.

2. The nnUNet zebrafish larvae whole brain vasculature segmentation method based on autonomous dataset training according to claim 1, characterized in that, Step 1 comprises the following steps: First, during the process of high-resolution imaging of the cerebral vascular system of a zebrafish by using an in vivo endothelial cell transgenic zebrafish, the zebrafish is subjected to fluorescence imaging in the whole brain range by using a confocal microscope, and deep brain vascular signals are obtained; The zebrafish is fixed in a dorsal position by using an agarose gel, and under a 20-fold water lens, the brain imaging front-to-back diameter is set as follows: the rostral side includes the transition region of the nasal artery and the nasal vein in the olfactory region, and the caudal side includes the first annular blood vessel at the junction of the basilar artery and the spinal cord blood vessels; The Z-axis range is set as follows: the dorsal side is marked by the dorsal longitudinal vein, the ventral side is marked by the first second gill arch merging into the original internal jugular vein, and the Z-axis shooting step is set; In the Z-axis correction mode, the top, middle and bottom positions are automatically identified according to the current full Z-axis range, and the imaging parameters of the three layers are set; During the shooting process, the fluorescence signals of the whole brain blood vessels are imaged in an automatic incremental mode to the target value, the dynamic parameter incremental setting of the Z-axis fluorescence excitation intensity is established, and high-resolution imaging of the whole brain blood vessels of the zebrafish is realized.

3. The nnUNet zebrafish larvae whole brain vasculature segmentation method trained based on autonomous dataset according to claim 2, characterized in that, With the dorsal long saphenous vein as the imaging starting layer, first set the top layer excitation light intensity ELI Top The value range, then set three layers of imaging parameters as follows: HV Top = ELI Top + a; ELI Middle = ELI Top / mt1*mt2, HV Middle = ELI Middle + b; ELI Botton = ELI Middle / bm1*bm2, HV Botton = ELI Botton + c; Wherein, ELI Top and HV Top are the top layer excitation light intensity and the top layer high voltage value, respectively, ELI Middle and HV Middle are the middle layer excitation light intensity and the middle layer high voltage value, respectively, ELI Botton and HV Botton are the bottom layer excitation light intensity and the bottom layer high voltage value, respectively, and a, b, c, mt1, mt2, bm1, bm2 are self-defined parameters.

4. The nnUNet zebrafish larvae whole brain vasculature segmentation method based on autonomous dataset training according to claim 1 or 2, characterized in that, In step 2, the intracranial vascular system segmentation true value and data set are obtained by using a semi-automatic combined manual correction image segmentation step, which comprises the following steps: The Z-axis correction imaging brain blood vessel signals are semi-automatically segmented by using the Fiji_Labkit plug-in, the original signals are manually added with sparse blood vessel markers, the blood vessel filling and tube wall markers are marked as foreground, and sparse background signals are given, and the image label iterative training is interactively performed to obtain a binary segmentation image; The brain blood vessel filling state of all layers of the binary segmentation image is corrected by using ITK_SNAP, the central blood vessel signals are removed by identification, and a full true value segmentation image stack of the intracranial vascular system is obtained.

5. The nnUNet zebrafish larvae whole brain vasculature segmentation method based on autonomous dataset training according to claim 1, characterized in that, In step 3, the nnUNet adaptive network structure is used, the network comprises a convolutional layer, a 3D maximum pooling layer, a deconvolutional layer and a skip connection; 3D patch input is used, data enhancement is performed by using multiple data enhancement methods, the preprocessed data is divided and the nnUNet is trained, and the optimizer is an SGD with Nesterov momentum; The initial learning rate and the upper limit of training are set, the poly decay strategy is used for adjustment, the model with the optimal verification performance is automatically saved, the trained model is used for image prediction, and the merged 3D reconstruction result is output.

Images