nnunet segmentation method for zebrafish larva whole brain vasculature based on self-contained dataset training

Abstract

The application discloses a kind of nnUNet zebra fish juvenile whole brain vascular system segmentation methods based on autonomous data set training, it is related to high-resolution imaging technology, image processing and medical image segmentation field, the method makes full use of zebra fish live transparency and fluorescent label advantage, obtains high-resolution whole brain three-dimensional vascular image data, and constructs high-quality segmentation truth value database by semi-automatic segmentation and artificial correction, training is carried out using nnU-Net deep learning model, realize the three-dimensional automatic segmentation of zebra fish brain vascular system signal.The application method significantly improves the degree of automation and precision of image segmentation, effectively solves the problems of low efficiency, high artificial dependence and poor repeatability of traditional brain vascular segmentation.The method is suitable for large-scale high-throughput data processing, can provide efficient, standardized image processing scheme for zebra fish brain vascular development mechanism and brain vascular disease model research, and has wide application prospect.

Description

Technical Field

[0001] This invention belongs to the field of biomedical image processing and computer vision technology, specifically relating to a method for segmenting the whole brain vascular system of juvenile zebrafish using nnUNet trained on an autonomous dataset. Background Technology

[0002] Quantitative assessment of human cerebral vascular phenotypes relies on visual observation and manual measurement. Achieving a balance between imaging quality and image segmentation efficiency in 3D reconstruction and quantification requires technological breakthroughs. 1-2 Therefore, by selecting suitable model animals and establishing a complete platform for the identification, reconstruction, and quantitative analysis of the cerebrovascular system, a comprehensive description of the physiological structure and disease phenotypes of cerebrovascular systems can be achieved, providing technical strategies for cerebrovascular research and breaking through the current technical bottlenecks in cerebrovascular research.

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

[0004] Zebrafish, as a vertebrate model organism, has significant advantages in studying mechanisms related to vascular system development. Firstly, the short breeding cycle and high-throughput operations of zebrafish facilitate genetic studies such as phenotypic identification and mutant construction using CRISPR / Cas9 gene editing and Morpholino-mediated gene knockdown techniques. Secondly, compared to other species that require complex procedures such as tissue clearing and special staining of brain specimens, or angiography or tissue exposure before in vivo imaging, zebrafish offer a more efficient and efficient learning experience. 5-6 The fully transparent nature of zebrafish during development allows for dynamic observation of the entire brain vascular development cycle through non-invasive in vivo tissue imaging, giving zebrafish an irreplaceable advantage in in vivo experimental imaging techniques. Furthermore, zebrafish possess blood-brain barrier structural components similar to those found in mice and humans.7 Furthermore, various transgenic zebrafish strains with fluorescently labeled vascular systems enable high-resolution in vivo imaging of the vascular system and even super-resolution imaging of subcellular structures, contributing to a deeper understanding of the mechanisms of angiogenesis at the cellular biology level. Therefore, zebrafish are currently the ideal model animal for high-throughput, non-invasive, in vivo dynamic imaging studies of the entire brain's blood vessels.

[0005] Based on the current research status, it is feasible to use live zebrafish to achieve precise and efficient whole-brain vascular reconstruction and quantification, which is an important technological breakthrough for carrying out research on encephalopathy and cerebral vascular phenotypes.

[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 biomedicalimage segmentation[C].International Conference on Medical image computing andcomputer-assisted intervention.Springer,Cham,2015:234-241.

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

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

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

[0013] 7.O'Brown,NMet al.Bridging barriers: a comparative look at the blood-brain barrier across organisms[J].Genes Dev,2018,32:466-478. Summary of the Invention

[0014] Purpose of the invention: To address the problems of low efficiency, insufficient accuracy, difficulty in identifying deep blood vessels, and high dependence on manual intervention in existing zebrafish cerebral vascular image segmentation methods, this invention proposes an automatic segmentation method for the whole cerebral vascular system of juvenile zebrafish, in order to improve the accuracy and processing efficiency of three-dimensional image segmentation and promote the high-throughput and standardized development of cerebral vascular research.

[0015] Technical solution: To achieve the objectives of this invention, the technical solution adopted is as follows:

[0016] A method for segmenting the whole-brain vascular system of juvenile zebrafish based on an nnUNet model trained on a self-developed dataset is proposed. A standardized image segmentation and analysis workflow is constructed to achieve automatic three-dimensional segmentation and digital reconstruction of the intracranial vascular system of zebrafish. The method includes the following steps:

[0017] Step 1: High-resolution three-dimensional imaging of cerebral blood vessels with Z-axis compensation was performed using transgenic zebrafish with endothelial cell-specific fluorescent labeling.

[0018] Step 2: Combine semi-automatic image segmentation and correction strategies to obtain real-label images of the intracranial vascular system and construct a standardized vascular segmentation dataset for training.

[0019] Step 3: Based on the constructed dataset, the nnUNet architecture is applied to train the model, realize the fully supervised segmentation and prediction of zebrafish cerebral blood vessels, and obtain 3D mask results for digital reconstruction of the whole brain blood vessels.

[0020] Furthermore, step 1 includes:

[0021] First, during the process of high-resolution imaging of the cranial vascular system using transgenic zebrafish with live endothelial cells, confocal microscopy was used to perform whole-brain fluorescence imaging of the zebrafish to obtain signals of deep brain blood vessels.

[0022] Agarose gel was used to fix zebrafish in a dorsal position, and the anteroposterior diameter of the brain was set under a 20x water microscope: the snout side included the nasal artery-nasal vein transition area in the olfactory region, and the tail side included the first ring vessel at the junction of the basilar artery and spinal cord vessels.

[0023] Set the Z-axis range: the dorsal long vein is used as a landmark on the dorsal side, and the point where the first and second branchial arches converge into the primitive internal jugular vein is used as a landmark on the ventral side. Set the Z-axis shooting step length.

[0024] In Z-axis correction mode, the top, middle and bottom layers are automatically identified based on the current full Z-axis range, and the imaging parameters of the three layers are set accordingly.

[0025] During the imaging process, the fluorescence signal of the whole brain blood vessels was imaged in a mode that automatically incremented to the target value. By establishing a dynamic parameter increment setting for the fluorescence excitation intensity of the Z-axis, high-resolution imaging of the whole brain blood vessels of zebrafish was achieved.

[0026] Furthermore, using the dorsal long vein as the imaging starting layer, the top layer excitation light intensity ELI was first set. Top The value range is then determined, and the imaging parameters for 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] Among them, ELI Top and HV Top These represent the excitation light intensity and the high voltage value of the top layer, respectively. Middle and HVMiddle These represent the excitation light intensity and high voltage value of the intermediate layer, respectively. Botton and HV Botton These represent the excitation light intensity and the high voltage value of the underlying layer, respectively. a, b, c, mt1, mt2, bm1, and bm2 are user-defined parameters.

[0031] Furthermore, step 2 applies a semi-automatic combined with manual correction image segmentation steps to obtain the ground truth and dataset of intracranial vascular system segmentation, including:

[0032] The Fiji_Labkit plugin was used to perform semi-automatic segmentation of cerebral vascular signals from Z-axis corrected imaging. Sparse vascular markers were manually added to the original signal, including marking the vascular filling and vessel walls as the foreground, while sparse background signals were provided. Interactive image label iterative training was performed to obtain a binarized segmented image.

[0033] ITK_SNAP was applied to correct the cerebral blood vessel filling status of all layers of the binarized segmented image. After discrimination and removal of extra-central vascular signals, a stack of true-value segmented images of the intracranial vascular system was obtained.

[0034] Furthermore, in step 3, the nnUNet dynamic adaptation network structure is adopted, which includes convolutional layers, 3D max pooling, deconvolutional layers, and skip connections.

[0035] 3D patch input is used, and data augmentation is performed through various data augmentation methods. The preprocessed data is divided into blocks and nnUNet is trained. The optimizer is SGD with Nesterov momentum.

[0036] Set the initial learning rate and training cap, and adjust them using a poly decay strategy. Automatically save and validate the model with the best performance. The trained model is used for image prediction and outputs the merged 3D reconstruction results.

[0037] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0038] This invention provides the first standardized three-dimensional image segmentation process for the whole brain vascular system of juvenile zebrafish, a model animal. It effectively improves the imaging quality and recognition accuracy of deep blood vessels. Through self-labeled data, it achieves the removal of extracerebral signals and the complete extraction of cerebral vascular structures, significantly improving the accuracy of the dataset. Based on the self-labeled dataset, it achieves efficient and automated image segmentation, providing phenotypic research technical support for cerebral vascular development and disease research. Attached Figure Description

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

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

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

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

[0043] Figure 5 The training model nnUNet_flt1 is established and its prediction performance is evaluated through transfer learning strategies. Detailed Implementation

[0044] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0045] This invention provides a method for segmenting the whole-brain vascular system of juvenile zebrafish based on an nnUNet model trained on an autonomous dataset, enabling high-precision automatic segmentation of the whole-brain vascular system in juvenile zebrafish. A standardized image segmentation and analysis pipeline (Zebrafish Intracranial Vasculature Segmentation and Standardization Pipeline, ZICVesSP) is constructed based on the three-dimensional imaging features of zebrafish intracranial vessels, as follows: Figure 1 , Figure 1 Image A shows Z-axis corrected confocal 3D imaging of the cerebral blood vessels in a live zebrafish with endothelial transgenic cells. Figure 1 B in the example demonstrates the application of a semi-automatic combined manual correction step to obtain the ground truth and dataset of intracranial vascular system segmentation. Figure 1 The example C demonstrates how to achieve fully supervised 3D mask prediction based on the nnUNet model through deep learning training using a constructed dataset (ZICVesDataset). The entire process autonomously implements the transposition of image formats (.ims / .tif / .nii) between software via code, achieving an automated processing framework from importing raw images to fine segmentation of vascular structures. It achieves automated 3D segmentation and digital reconstruction of the intracranial vascular system in zebrafish. Using transgenic zebrafish with endothelial cell-specific fluorescent labels, high-resolution 3D imaging of cerebral vessels with Z-axis compensation is performed. Combined with semi-automatic image segmentation and correction strategies, true labeled images of the intracranial vascular system are obtained, and a standardized vascular segmentation dataset is constructed for training. Based on the constructed dataset, the nnUNet architecture is applied for model training to achieve fully supervised segmentation prediction of zebrafish cerebral vessels, obtaining 3D mask results for whole-brain vascular digital reconstruction.

[0046] First, during high-resolution imaging of the cranial vascular system using the transgenic zebrafish line Tg(kdrl:ras-mCherry) with live endothelial cells, confocal microscopy was used to perform whole-brain fluorescence imaging to enhance the acquisition of deep brain vascular signals. Anesthetized zebrafish were fixed in a dorsal position on agarose gel. Under 20x water microscopy, the anteroposterior diameter of the brain imaging was set: the rostral side included the olfactory region where the nasal artery and nasal vein transitioned, and the caudal side included the first ring vessel at the junction of the basilar artery and spinal cord vessels; the Z-axis range was: dorsal side marked by the long dorsal vein, and ventral side marked by the point where the first and second branchial arches converge into the primitive internal jugular vein, with a Z-axis step size of 2 / 3 μm. In Z-axis correction mode, the system automatically identified the top, middle, and bottom layers based on the current full Z-axis range. Imaging parameters such as excitation light intensity (ELI) and high voltage (HV) values ​​for the three layers were set: using the long dorsal vein as the initial imaging layer, the excitation light intensity (ELI) of the top layer was first set. Top The value range is 20-25mV, and then the imaging parameters for the three layers are set 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; where ELI Top and HV Top These represent the excitation light intensity and the high voltage value of the top layer, respectively. Middle and HV Middle These represent the excitation light intensity and high voltage value of the intermediate layer, respectively. Botton and HV Botton These represent the excitation light intensity and high voltage value of the underlying layer, respectively, with user-defined parameters a=5, b=5, c=5, mt1=100, mt2=130, bm1=100, bm2=140. Pinhole: 1.2AU; Frame Average: 4x. During imaging, the system automatically increments to the target values ​​to image the whole-brain vascular fluorescence signal. Figure 1 As shown in Figure A, high-resolution imaging of the whole brain blood vessels in zebrafish was achieved by establishing a dynamic parameter increment setting for the Z-axis fluorescence excitation intensity.

[0050] Figure 2 Figure A shows the spatial distribution of fluorescent vascular signal intensity obtained by signal intensity topography (Interactive 3D Surface Plot) before and after excitation correction. Fluorescence intensity topography is used to display the spatial distribution of fluorescent vascular signal intensity obtained by the control group and after Z-axis correction imaging. The entire Z-axis range is selected for visualization, including the superficial layer (layer 6 / Z6), intermediate layer (layer 50 / Z50), and deep layer (layer 100 / Z100). After correction imaging, quantitative results of the fluorescence signal area show that more signal was obtained in the intermediate and deep vascular locations.

[0051] Figure 2 The mean fluorescence intensity of different optical slice layers after B-axis detection control and Z-axis correction imaging (6 cases / group) was measured. Data are expressed as mean ± standard error. Statistical significance was determined by two-way ANOVA combined with Tukey's post-hoc test (*P<0.05, ****P<0.0001). In Z-axis correction mode, the mean fluorescence gray value increased in the middle of the imaging range (Z50), and the mean fluorescence gray value in the deeper layers (Z75, Z100) was significantly compensated. Furthermore, the fluorescence signal intensity of the deep endothelium was similar to that of the superficial layer.

[0052] To obtain segmented images of the cranial vascular system Figure 1 B in the diagram is a schematic diagram of obtaining the true value and dataset of intracranial vascular system segmentation by applying a semi-automatic combined manual correction image segmentation step.

[0053] Figure 3 The process involved manually correcting the cerebral vascular filling status of all layers using the combined medical image processing software ITK_SNAP. Two medical experts then manually removed extra-central vascular signals such as those from the eyes, gill arches, heart, and fins to obtain a true-value segmentation image stack of the intracranial vascular system. Figure 3 Figure A demonstrates the application of the shallow learning tool Fiji_Labkit plugin to perform semi-automatic segmentation of cerebral vascular signals from Z-axis corrected imaging. By manually adding sparse vascular markers (green) to the original signal (left image - bottom left), including the filling of the blood vessels (green) and the vessel walls (yellow) as foreground, and simultaneously providing a sparse background signal (gray) (left image - bottom right), the binary segmented image (right image) is obtained through interactive iterative training of image labels (middle image). Figure 3 The B application uses ITK-SNAP to perform layer-by-layer manual evaluation and correction of the binarized segmented image to obtain the true segmentation. Under the Zhoumian perspective, non-intracranial vascular signals of the optical slice image are removed (yellow western box). The overall effect of fluorescence signal segmentation (lower left) and intracranial vascular system (lower right) correction is shown in 3D view.

[0054] To improve the efficiency of image segmentation, this invention uses the original sample image and the ground truth image as a pair of data sets. Figure 1 A, 1B (dashed boxes) Construct a zebrafish juvenile brain vascular dataset, which consists of 105 data sets covering various developmental stages and physiological and pathological model samples, as shown in Table 1.

[0055] Table 1. Training dataset for the nnUNet_kdrl model

[0056]

[0057]

[0058] Five groups of wild-type samples from each developmental stage were randomly selected as the prediction set for training and model evaluation. 100 samples were trained using a 5-fold cross-validation strategy, with 80 samples used as the training set and 20 samples used as the validation set to ensure that all samples participated in the validation process. Figure 1 C in the example demonstrates how to achieve fully supervised 3D mask prediction of brain blood vessels based on the nnUNet model by constructing a dataset for deep learning training.

[0059] Figure 4 In the diagram, 'A' represents the preprocessing steps and architecture of the official nnUNet (v2.2.0), dynamically adapting the network structure and patch size, using 3D patch input, and integrating multiple data augmentation methods (rotation, translation, noise reduction, blurring, brightness, and contrast). The optimizer is SGD with Nesterov momentum, with an initial learning rate of 0.01, adjusted using a poly decay strategy. The training duration is capped at 1000 epochs per fold, and the optimal model is automatically saved for validation. The training process is automated via command line to ensure reproducibility and standardization of the experiments.

[0060] Figure 4 Figure A illustrates the fluorescence image processing and prediction workflow of the nnUNe model. The process begins with image preprocessing, including grayscale conversion, font transformation, and data augmentation techniques. The preprocessed data is then divided into blocks and used to train the nnUNet architecture, which includes convolutional layers, 3D max pooling, deconvolutional layers, and skip connections to achieve efficient feature extraction. The trained model is then used for image prediction and outputs merged 3D reconstruction results.

[0061] Figure 4The training curve of nnUNet_kdrl, represented by B, shows that during 1000 epochs of training, the accuracy curve gradually increases and then stabilizes, while the training loss and validation loss correspondingly decrease. The Dice coefficient stabilizes as training progresses, indicating improved model segmentation performance and convergence. Model training was completed on a high-performance workstation equipped with an NVIDIA RTX 4090 GPU, running in Python 3.8 and PyTorch 1.10.

[0062] Figure 4 B in the diagram shows the training curves of the nnUNet_kdrl model, displaying the loss curves: loss training (loss_tr: blue line) and training set loss validation (loss_val: red line), and validation set Directly calculated Dice coefficient (pseudo dice(mov.age.): green line).

[0063] The nnUNet_kdrl training model is a segmentation model built for red vascular fluorescent fish strains using kdrl as the endothelial promoter. The nnUNet_kdrl model is applied to predict the red fluorescence signal of vascular endothelium using kdrl as the promoter. The model prediction results are visualized as follows: Figure 5 In the model, B, B1-B3 (B, B1-B3, nnUNet_kdrl model predicts Tg(kdrl:ras-mCherry) channel cerebral vascular signals), the vascular segmentation signal is shown in yellow, including missing signals (white boxes) and misidentified signals (blue boxes). The prediction results for 5 samples / groups in the prediction set were evaluated using Dice similarity coefficient, intersection-over-union (IOU) value, and Hausdorff distance (HD95), as shown in Table 3. Because transgenic fish strains with different promoters exhibit differences in the localization and intensity of fluorescence signals on endothelial cells, the nnUNet_kdrl model was used to predict the green fluorescence signal of vascular endothelium using flt1 as the promoter. Figure 5The visualization results for B and C1-C3 show over-recognition of ocular and cardiac signals (C1-C3, nnUNet_kdrl model predicts cerebral vascular signals in the Tg(flt1BCA:YFP) channel; D1-D3, nnUNet_flt1 model predicts cerebral vascular signals in the Tg(kdrl:ras-mCherry) channel; E1-E3, nnUNet_flt1 model predicts cerebral vascular signals in the Tg(flt1BCA:YFP) channel. The vascular segmentation signals are shown in yellow, including missing signals (white boxes) and misidentified signals (blue boxes). Table 3 shows a decrease in the Dice coefficient and IOU, and a significant increase in HD95, indicating reduced segmentation quality and suggesting insufficient model generalization ability.

[0064] By formulating a transfer learning strategy and leveraging the high-throughput dual-fluorescent marker construction advantage of zebrafish, and based on the established nnUNet_kdrl training model, efficient construction of datasets and training models for different endothelial fish strains can be achieved, compensating for generalization limitations. First, a dual-transgenic fish strain Tg(kdrl:ras-mCherry::flt1BCA:YFP) is constructed, enabling a single fish to simultaneously express both red and green endothelial fluorescence signals. Using the established self-developed dataset prediction platform nnUNet_kdrl, a segmentation map of red endothelial cerebral vessels is predicted. Then, on the same fish, this segmentation map is used, with a green signal as the base map, for correction, to obtain a matching segmentation map for the green signal, thus establishing a segmentation database for green endothelial signals. Figure 5 In the diagram, A1-A3 represent further obtained dual-channel confocal imaging of the double-transgenic zebrafish line Tg(kdrl:ras-mCherry::flt1BCA:YFP) (A, A1-A3, endothelial transgenic zebrafish line Tg(kdrl:ras-mCherry::flt1BCA:YFP confocal dual-channel imaging). For the red fluorescent cerebral blood vessel signals located by the kdrl promoter, the nnUNet_kdrl model was applied for segmentation prediction. Using this segmentation map as a template, two medical professionals corrected the green fluorescent cerebral blood vessel signals located by the flt1 promoter in the same sample, obtaining the ground truth segmentation values ​​for the intracranial vascular system (excluding non-cerebral blood vessels) located by the flt1 promoter. The training dataset is shown in Table 2, and the nnUNet_flt1 segmentation prediction model was then obtained.

[0065] Table 2 shows the training dataset for the nnUNet_flt1 model.

[0066]

[0067] Figure 5In the model, B and D1-D3 represent the red fluorescence signals of vascular endothelium predicted by the nnUNet_flt1 model with kdrl as the promoter; E1-E3 represent the green fluorescence signals of vascular endothelium in the brain predicted by the nnUNet_flt1 model with flt1 as the promoter. Figure 5 C in the figure represents the obtained nnUNet_flt1 training curve (blue line is the loss curve loss training(loss_tr), red line is the training set loss validation(loss_val), and green line is the validation set directly calculated Dicecoefficient(pseudo dice(mov.age.)). During the 1000 epochs of training, the accuracy curve gradually increases and then stabilizes, while the training loss and validation loss show a corresponding decreasing trend. As training progresses, the Dice coefficient tends to stabilize, indicating that the model's segmentation performance is improving and its convergence is good.

[0068] Table 3. Evaluation of nnUNet Model Predictions

[0069]

[0070] This embodiment provides two fish-type intracranial vessel prediction models: the nnUNet_kdrl model and the nnUNet_flt1 model. Table 3 shows the segmentation prediction performance evaluation based on the prediction results of 5 samples / groups in the prediction set. Compared with the nnUNet_kdrl model, the nnUNet_flt1 model improves the Dice coefficient and IOU, and decreases the HD95, indicating that the nnUNet_flt1 model can efficiently obtain accurate segmentation results for the flt1 fish type through transfer learning strategy.

Claims

1. A method for segmenting the whole brain vascular system of juvenile zebrafish using nnUNet trained on an independent dataset, characterized in that... Includes the following steps: Step 1: High-resolution three-dimensional imaging of cerebral blood vessels with Z-axis compensation was performed using transgenic zebrafish with endothelial cell-specific fluorescent labeling. Specifically: First, during the process of high-resolution imaging of the cranial vascular system using transgenic zebrafish with live endothelial cells, confocal microscopy was used to perform whole-brain fluorescence imaging of the zebrafish to obtain signals of deep brain blood vessels. Agarose gel was used to fix zebrafish in a dorsal position, and the anteroposterior diameter of the brain was set under a 20x water microscope: the snout side included the nasal artery-nasal vein transition area in the olfactory region, and the tail side included the first ring vessel at the junction of the basilar artery and spinal cord vessels. Set the Z-axis range: the dorsal long vein is used as a landmark on the dorsal side, and the point where the first and second branchial arches converge into the primitive internal jugular vein is used as a landmark on the ventral side. Set the Z-axis shooting step length. In Z-axis correction mode, the top, middle, and bottom layers are automatically identified based on the current full Z-axis range, and the imaging parameters for these three layers are set accordingly. Specifically, using the dorsal long vein as the imaging starting layer, the excitation light intensity of the top layer is first set. The value range is then determined, and the imaging parameters for the three layers are set as follows: ； ， ； ， ； in, and These represent the excitation light intensity and the high voltage value of the top layer, respectively. and These represent the excitation light intensity and the high voltage value of the intermediate layer, respectively. and These represent the excitation light intensity and the high voltage value of the underlying layer, respectively, a, b, c, , , , These are custom parameters; During the imaging process, the fluorescence signal of the whole brain blood vessels was imaged in a mode that automatically incremented to the target value. By establishing a dynamic parameter increment setting for the fluorescence excitation intensity of the Z-axis, high-resolution imaging of the whole brain blood vessels of zebrafish was achieved. Step 2: Combine semi-automatic image segmentation and correction strategies to obtain real-label images of the intracranial vascular system and construct a standardized vascular segmentation dataset for training. Step 3: Based on the constructed dataset, the nnUNet architecture is applied to train the model, realize the fully supervised segmentation and prediction of zebrafish cerebral blood vessels, and obtain 3D mask results for digital reconstruction of the whole brain blood vessels.

2. The method for segmenting the whole brain vascular system of juvenile zebrafish based on nnUNet trained on an autonomous dataset as described in claim 1, characterized in that, Step 2 applies a semi-automatic combined with manual correction image segmentation steps to obtain the ground truth and dataset of intracranial vascular system segmentation, including: The Fiji_Labkit plugin was used to perform semi-automatic segmentation of cerebral vascular signals from Z-axis corrected imaging. Sparse vascular markers were manually added to the original signal, including marking the vascular filling and vessel walls as the foreground, while sparse background signals were provided. Interactive image label iterative training was performed to obtain a binarized segmented image. ITK_SNAP was applied to correct the cerebral blood vessel filling status of all layers of the binarized segmented image. After discrimination and removal of extra-central vascular signals, a stack of true-value segmented images of the intracranial vascular system was obtained.

3. The method for segmenting the whole brain vascular system of juvenile zebrafish based on nnUNet trained on an autonomous dataset as described in claim 1, characterized in that, In step 3, the nnUNet dynamic adaptation network structure is adopted, which includes convolutional layers, 3D max pooling, deconvolutional layers, and skip connections. 3D patch input is used, and data augmentation is performed through various data augmentation methods. The preprocessed data is divided into blocks and nnUNet is trained. The optimizer is SGD with Nesterov momentum. Set the initial learning rate and training cap, and adjust them using a poly decay strategy. Automatically save and validate the model with the best performance. The trained model is used for image prediction and outputs the merged 3D reconstruction results.

Images