A cerebellar purkinje neuron rapid identification system and method based on multi-modal deep learning

Abstract

The application provides a cerebellar Purkinje neuron rapid identification system and method based on multi-modal deep learning, relates to the cross technical field of biomedical engineering and computer vision, and fuses specific fluorescence features and multi-dimensional morphological features, constructs a special multi-modal deep learning model, solves the interference problem of other neurons in a mixed culture system, and greatly reduces the false positive rate; a small sample learning architecture solves the industry pain point of a small amount of biological sample labeled data, improves the model generalization capability, and adapts to multi-scene sample identification. The identification system fuses Purkinje neuron specific fluorescence labeling features and cell morphological features, constructs a multi-modal deep learning recognition model based on small sample learning, is matched with a full-automatic microscopic imaging and analysis system, realizes rapid and high-accuracy identification of Purkinje neurons under a mixed culture system, does not need a complex cell purification step, greatly shortens the experimental period, and reduces the technical threshold.

Description

Technical Field

[0001] This application relates to the field of cell recognition technology, and in particular to a rapid identification system and method for cerebellar Purkinje neurons based on multimodal deep learning. Background Technology

[0002] Purkinje cells are the only efferent neurons in the cerebellar cortex. Their morphological development and functional activity are closely related to the regulation of cerebellar physiological functions. They are also associated with the occurrence and development of various neurological diseases such as ataxia and spinocerebellar degeneration. Purkinje cell research in in vitro primary culture systems is one of the core models in the field of neurobiology.

[0003] Currently, in vitro Purkinje cell identification research heavily relies on high-purity purification and culture systems. The mainstream technique involves purifying cells using methods such as Percoll density gradient centrifugation and immunomagnetic bead sorting, followed by manual microscopic identification using Calbindin D-28k specific immunofluorescence staining. While improved rat / mouse cerebellar Purkinje cell purification and culture methods were published between 2021 and 2023, improving cell viability, core issues such as long culture cycles and cumbersome procedures remain unresolved, and manual microscopic identification has many limitations. With the development of computer vision and deep learning technologies, multimodal deep learning techniques are gradually being applied to Purkinje neuron identification. In 2022, a method for identifying Purkinje neurons in brain tissue slices based on convolutional neural networks (CNNs) was published. Between 2023 and 2024, the YOLO algorithm and Transformer architecture were also applied to neuron morphology recognition. However, these deep learning technologies have not yet developed mature application solutions, relying solely on cell morphology features and failing to achieve fully automated identification.

[0004] Existing technologies for identifying Purkinje cells have several core shortcomings. First, in vitro identification of Purkinje neurons heavily relies on high-purity purification and culture, which involves cumbersome and lengthy culture steps, resulting in low cell survival rates and high technical barriers. Second, manual microscopic identification is inefficient, prone to human error, and cannot meet the needs of high-throughput experiments. Third, existing deep learning-based identification technologies have low accuracy and high false-positive rates in identifying Purkinje neurons in mixed culture systems, and can only be applied to samples from single species and single culture stages, exhibiting poor generalization ability. Fourth, existing technologies have not achieved multimodal fusion deep learning recognition of specific biomarkers and morphological features, failing to overcome the technical bottlenecks in mixed culture systems, and most only achieve semi-automation of the identification process, without forming a fully automated solution from image acquisition to result output. Summary of the Invention

[0005] The purpose of this application is to address the core problems in the existing technology of identifying cerebellar Purkinje neurons, which relies on high-purity purification and culture, involves cumbersome procedures, long culture cycles, and low cell survival rates.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] A rapid identification system for cerebellar Purkinje neurons based on multimodal deep learning includes a hardware system and a software system. The software system includes an image acquisition and preprocessing module, a standard database construction and model training module, an automated identification module, and a result output and verification module. The standard database construction and model training module builds a multimodal deep learning identification model based on a few-shot learning architecture, which is adapted to the training requirements of biological samples with limited labeled data. The multimodal deep learning identification model can extract the fluorescence features, morphological features, and texture features of Purkinje neurons to complete feature matching and neuron identification.

[0008] Preferably, the core of the hardware system is a fully automated cerebellar cell microscopy imaging unit, which consists of an upright fluorescence microscope, a high-resolution image acquisition system with more than 5 million pixels, a fully automated motorized stage, and an automatic focusing control system; it can automatically identify the cell-covered areas in the sample, avoid overlapping cell fields of view, and simulate the logic of manual microscopy to complete automatic focusing, field of view traversal, and image acquisition.

[0009] Preferably, the image acquisition and preprocessing module is used to interface with the hardware system to complete fully automatic image acquisition, simultaneously acquiring bright-field cell morphology images and corresponding fluorescently labeled images in the field of view to form paired multimodal images; preprocessing the acquired images, and inputting the preprocessed images into the multimodal deep learning recognition model.

[0010] Preferably, the multimodal deep learning recognition model uses Calbindin D-28k / L7 / Pcp2 / IP3R1 specific fluorescent staining as the gold standard, and collects Purkinje neuron pairing images from different species, different culture cycles, and different culture systems as a standard database. The feature inputs of the multimodal deep learning recognition model include fluorescence features, morphological features, and texture features.

[0011] Preferably, the automated identification module is used to automatically generate an identification report, which includes the number and proportion of target cells, morphological parameter statistics, and a visual image of the identification marker.

[0012] This application also provides a rapid identification method for cerebellar Purkinje neurons based on multimodal deep learning, which is based on the identification system described above.

[0013] Preferably, the identification method steps are as follows:

[0014] S1: Pretreatment of cell culture vector;

[0015] S2: Mixed culture of cerebellar neurons;

[0016] S3: Purkinje neuron-specific marker;

[0017] S4: Multimodal deep learning workflow: The sample is placed in a fully automated microscopic imaging system, and bright-field cell morphology images and corresponding field-of-view fluorescent label images are acquired simultaneously to form paired multimodal images. After preprocessing the images such as background denoising and contrast enhancement, they are input into the trained multimodal deep learning recognition model. The model automatically completes feature extraction and matching, locates and identifies Purkinje neurons, and finally outputs an identification report, which supports manual review and correction.

[0018] Preferably, the specific steps of S1 are as follows: the coverslips for inoculating cells are coated with 400-500 μg / mL poly-L-lysine solution, placed in a 35 mm culture dish, and incubated overnight at 37°C in a 5% CO2 incubator; the coverslips are washed twice with autoclaved PBS buffer, and dried in a fume hood under sterile conditions for later use, providing a stable solid-phase carrier for cell adhesion and reducing cell overlap.

[0019] Preferably, in step S3, specific labeling is performed using immunofluorescence staining.

[0020] Preferably, in step S3, the Purkinje neuron-specific adenovirus transfection fluorescent labeling method is used.

[0021] Compared with the prior art, this application has at least the following technical effects:

[0022] 1. The identification system constructed in this application relies on the feature fusion advantages of the multimodal deep learning recognition model to achieve an accuracy of ≥98%, sensitivity of ≥97%, specificity of ≥99%, and false positive rate of ≤1% for identifying cerebellar Purkinje neurons in a mixed culture system. This is far superior to the approximately 85% accuracy of existing technologies in mixed systems, and solves the core pain point of target cell identification in mixed systems.

[0023] 2. The system constructed in this application has a single-field sample identification time of ≤2s and can complete the identification of a whole plate of 96-well samples within 30 minutes. Compared with traditional manual microscopic examination, the identification efficiency is improved by ≥90%, and compared with existing semi-automatic deep learning recognition technology, the efficiency is improved by ≥60%, making it perfectly suitable for high-throughput drug screening and large-scale sample analysis scenarios.

[0024] 3. The identification method provided in this application does not require 2-4 weeks of high-purity purification culture, and can stably identify samples cultured in vitro for more than 24 hours, shortening the experimental cycle by ≥80%; it avoids mechanical damage to cells during the purification process, and the survival rate of Purkinje neurons is increased by ≥50% compared with traditional purification culture;

[0025] 4. The multimodal deep learning recognition model in this application can be adapted to multiple experimental species such as SD rats and C57BL / 6 mice, and can be adapted to samples at different differentiation stages from 1 day to 30 days of in vitro culture. It can also be adapted to multiple systems such as purification culture and mixed culture, breaking the limitation of the single application scenario of the existing technology.

[0026] 5. The identification system does not require professional neuromorphological identification experience. Ordinary experimental personnel can complete the entire process after 1 hour of training. The automated identification of the multimodal deep learning recognition model reduces human error and improves the batch-to-batch stability of experimental results. Attached Figure Description

[0027] Figure 1 This is a diagram illustrating the overall architecture of the Purkinje neuron automatic recognition system of the present invention.

[0028] Figure 2 This is a flowchart illustrating the rapid identification method for cerebellar Purkinje neurons according to the present invention.

[0029] Figure 3 This is a diagram showing the identification results of Purkinje neurons purified and cultured from SD rats for 7 days in Example 1;

[0030] Figure 4 This is a diagram showing the identification results of Purkinje neurons from mice cultured together for 4 days in Example 2.

[0031] Figure 5 A bar chart comparing the technical effects of the embodiments and comparative examples;

[0032] Figure 6 This is a reference diagram for a fully automated microscopic imaging and analysis system. Detailed Implementation

[0033] This application provides a rapid identification system for cerebellar Purkinje neurons based on multimodal deep learning, comprising a hardware system and a software system. The core of the hardware system is a fully automated cerebellar cell microscopy imaging unit, consisting of an upright fluorescence microscope (Leica DML type), a high-resolution image acquisition system with over 5 megapixels (Leica DFC310FX), a fully automated motorized stage, and an automatic focusing control system. It can automatically identify cell-covered areas in the sample, avoid overlapping cell fields of view, and simulate manual microscopy logic to complete automatic focusing, field of view traversal, and image acquisition without manual intervention, providing high-quality paired image input for the multimodal deep learning recognition model.

[0034] Please see Figure 1 The software system of the identification system includes an image acquisition and preprocessing module, a standard database construction and model training module, an automated identification module, and a result output and verification module.

[0035] In one embodiment, the image acquisition and preprocessing module is used to interface with the hardware system to complete fully automatic image acquisition, simultaneously acquiring bright-field cell morphology images and corresponding fluorescently labeled images to form paired multimodal images; preprocessing the acquired images, including background denoising, contrast enhancement, cell edge sharpening, and overlapping region segmentation, and inputting the preprocessed images into a multimodal deep learning recognition model.

[0036] The image acquisition and preprocessing module eliminates noise interference during image acquisition, accurately segments single-cell regions, provides high-quality input for feature extraction of multimodal deep learning recognition models, and further improves the model's recognition accuracy.

[0037] The standard database construction and model training module is used for database construction and model training. The database construction uses Calbindin D-28k / L7 / Pcp2 / IP3R1 specific fluorescent staining as the gold standard. It collects Purkinje neuron pairing images from different species (rats, mice), different culture periods (1d, 3d, 7d, 14d, 30d), and different culture systems (mixed culture, purified culture) and incorporates them into the standard database to provide sufficient and diverse samples for training multimodal deep learning recognition models.

[0038] In one embodiment, feature extraction involves extracting three core categories of features: cell fluorescence features (fluorescent positive areas, intensity, and distribution patterns), morphological features (area, perimeter, number and length of protrusions, etc.), and texture features (cell boundary texture, average cytoplasmic gray value, etc.).

[0039] The model training is based on a few-shot learning architecture, constructing a multimodal deep learning recognition model that integrates CNN and Transformer. The AdaBoost algorithm is used to optimize feature weights, completing the model training and validation. The model achieves a recognition accuracy of ≥99% on the validation set samples, providing reliable model support for subsequent automated recognition.

[0040] In one implementation, mainstream object detection architectures such as YOLOv8 and Swing Transformer can be used to replace the CNN+Transformer fusion model, adapting to the rapid screening scenario of ultra-high throughput samples and optimizing recognition efficiency.

[0041] The multimodal deep learning recognition model in this application, compared with the traditional deep learning recognition based on a single morphological feature, integrates specific fluorescence features and multidimensional morphological features to construct a dedicated multimodal deep learning recognition model. This solves the problem of interference from other neurons in the mixed culture system and significantly reduces the false positive rate. The small sample learning architecture addresses the industry pain point of limited biological sample annotation data, improves the model's generalization ability, and adapts to sample identification in multiple scenarios.

[0042] The automated recognition module is used to input the preprocessed multimodal image of the test sample into the trained multimodal deep learning recognition model. The model automatically completes feature extraction and matching, accurately locates Purkinje neurons in the field of view, and completes cell labeling, counting, and quantitative analysis of morphological parameters. It supports switching between multiple classification algorithms such as linear classification, support vector machine, Bayesian classification, and neural network to adapt to different sample scenarios.

[0043] The result output and verification module is used to automatically generate identification reports, including the number and proportion of target cells, morphological parameter statistics, and visualization images of identification markers; it supports manual verification and correction, and the verified samples can be included in the database for incremental training of multimodal deep learning recognition models to continuously improve the model's recognition capabilities.

[0044] Furthermore, based on the aforementioned identification system, this application provides a rapid identification method for cerebellar Purkinje neurons based on multimodal deep learning, such as... Figure 2 As shown, this method does not require complex cell purification steps. It achieves rapid identification of Purkinje neurons in a mixed culture system by combining cerebellar neuron co-culture, specific labeling, and multimodal deep learning recognition. The core steps include cell culture vector pretreatment, cerebellar neuron co-culture, Purkinje neuron specific labeling, and the corresponding image acquisition and multimodal deep learning recognition process.

[0045] The identification method steps are as follows:

[0046] S1: Cell culture vector pretreatment

[0047] The coverslips inoculated with cells were coated with 400-500 μg / mL poly-L-lysine solution, placed in 35 mm culture dishes, and incubated overnight at 37°C in a 5% CO2 incubator. The coverslips were then washed twice with autoclaved PBS buffer and air-dried in a fume hood under sterile conditions to provide a stable solid-phase carrier for cell adhesion and reduce cell overlap.

[0048] S2: Cerebellar neuron co-culture

[0049] (1) Take experimental animals (SD rats / C57BL / 6 mice) that are 1 day old (P0), disinfect their bodies with iodine and 75% medical alcohol, decapitate them under sterile conditions and remove their brains, and place the brain tissue in pre-cooled D-Hanks balanced salt solution.

[0050] (2) The cerebellar tissue was separated under a stereomicroscope, the meninges and blood vessels were removed, and the tissue was transferred to a sterile culture dish containing DMEM / F12 basal medium and cut into 1 mm pieces. 3 homogenous paste;

[0051] (3) Add 0.125% trypsin solution and shake in a constant temperature water bath at 37℃ for 15 min, then add DMEM / F12 medium containing 10% fetal bovine serum (FBS) to terminate digestion;

[0052] (4) After pipetting to form a single-cell suspension, centrifuge at 1500 rpm for 7 min, discard the supernatant, and resuspend in DMEM / F12 complete medium containing 10% FBS;

[0053] (5) Let the cell suspension stand at room temperature for 20 min, and then seed it onto the pretreated coverslip at the corresponding density;

[0054] (6) Incubate at 37℃ and 5% CO2. Replace with serum-free DMEM / F12 medium after 24 h. Replace half the medium every 2 days thereafter. It can be cultured for 1-30 days.

[0055] S3: Purkinje neuron-specific marker

[0056] In one embodiment, specific labeling is achieved using immunofluorescence staining, and the specific steps are as follows:

[0057] (1) Take a coverslip with cells, wash it once with PBS buffer, permeate it with 0.5% Triton X-100 at room temperature for 5 min, and then wash it three times with PBS;

[0058] (2) Block with 5% normal goat serum at room temperature for 1 h, discard the blocking solution and add Purkinje neuron-specific primary antibody, incubate in a humidified box at 4°C for 36-48 h;

[0059] (3) After rinsing with PBS 3 times, add the corresponding fluorescently labeled secondary antibody and incubate in a humidified chamber at 4°C in the dark for 4 h;

[0060] (4) After washing with PBS three times, the slide is mounted with glycerol for microscopic examination and image acquisition. The core markers are Calbindin D-28k, L7 / Pcp2, and IP3R1, which can be used alone or in combination.

[0061] In other embodiments, a Purkinje neuron-specific adenovirus transfection fluorescent labeling method is used to adapt to scenarios of live cell dynamic identification and long-term dynamic tracking and identification.

[0062] S4: Multimodal Deep Learning Process

[0063] The sealed sample is placed in a fully automated microscopic imaging system, and bright-field cell morphology images and corresponding field-of-view fluorescent labeled images are acquired simultaneously to form paired multimodal images. After preprocessing the images such as background denoising and contrast enhancement, they are input into the trained multimodal deep learning recognition model. The model automatically completes feature extraction and matching, locates and identifies Purkinje neurons, and finally outputs an identification report, which supports manual review and correction.

[0064] The above content will be explained in conjunction with specific verification experiments:

[0065] I. Experimental Materials and Their Sources:

[0066] 1. Laboratory animals

[0067]

[0068] 2. Core Reagents and Consumables

[0069]

[0070] 3. Main instruments and equipment

[0071]

[0072] II. Verification Experiment

[0073] Example 1: Identification and characterization of purified cultured Purkinje neurons from 1-day-old SD rats

[0074] 1. Database and Model Pre-training: Complete the construction of a standard database of Purkinje neurons from SD rats at different culture cycles, and complete the training of a multimodal deep learning recognition model fused with CNN+Transformer, with a model validation set accuracy of ≥99%.

[0075] 2. Cell purification and culture

[0076] 3. Take 8 SD rats that are 1 day old, and take cerebellar tissue under sterile conditions. Remove the meninges and blood vessels, digest with trypsin to prepare a single-cell suspension, centrifuge at 1500 rpm for 7 min and discard the supernatant.

[0077] 4. Add 2 mL of 35% Percoll separation solution and perform density gradient centrifugation to collect the interface cells. Resuspend the cells in DMEM / F12 complete medium containing 10% FBS.

[0078] 5. Calculate 2.5 × 10 6 pcs / cm 2 The culture medium was inoculated onto pretreated coverslips at a density of 100 μL and incubated at 37°C in a 5% CO2 incubator. After 24 h, the medium was replaced with serum-free medium, and half of the medium was replaced every 2 days for 7 days.

[0079] Specific labeling and image acquisition: Immunofluorescence staining was performed using Calbindin D-28k antibody, and image acquisition was completed using the fully automated microscopic imaging system of this invention after mounting.

[0080] Automated identification: The acquired images are input into the multimodal deep learning recognition system of this invention to complete automated identification, counting and parameter analysis.

[0081] Test results: Please refer to Figure 3 In this embodiment, the system achieved an accuracy rate of 98.7% in identifying purified and cultured Purkinje neurons, with a single field of view identification time of 1.2 seconds, and a consistency rate of 99.2% with the results of manual gold standard microscopic examination.

[0082] Example 2: Purkinje cell identification of cerebellar neurons from C57BL / 6 mouse co-culture

[0083] 1. Database and Model Pre-training: Completed the construction of a standard database of Purkinje neurons for C57BL / 6 mice at different culture cycles, completed incremental training of the multimodal deep learning fusion model, and achieved a model validation set accuracy of ≥99%.

[0084] 2. Cell co-culture

[0085] 3. Take 6 C57BL / 6 mice that are 1 day old (P0), take cerebellar tissue under sterile conditions, remove the meninges and blood vessels, digest with trypsin to prepare a single cell suspension, centrifuge at 1500 rpm for 7 min and discard the supernatant;

[0086] 4. Resuspend the cells in DMEM / F12 complete medium containing 10% FBS, at a ratio of 5 × 10⁻⁶. 6 pcs / cm 2 Inoculate the pretreated coverslip at a density of 10 ...

[0087] Incubate at 5.37℃ in a 5% CO2 incubator. Replace with serum-free medium after 24 hours and incubate for 4 days.

[0088] Specific labeling and image acquisition: Immunofluorescence staining was performed using L7 / Pcp2 antibody, and after mounting, fully automated image acquisition was completed using the system of this invention.

[0089] Automated identification: The acquired images are input into a multimodal deep learning recognition system to complete automated identification.

[0090] Test results: Please refer to Figure 4 In this embodiment, the system achieved an accuracy of 98.2% in identifying Purkinje neurons in the mixed culture system, with a single field of view identification time of 1.5 s, a sensitivity of 97.6%, a specificity of 99.1%, and a consistency of 98.8% with the results of manual gold standard microscopy.

[0091] Example 3: Dynamic identification of Purkinje neurons in a rat mixed culture system at different culture cycles

[0092] 1. Cerebellar neuronal samples from SD rats were prepared according to the mixed culture method in Example 2. Samples were taken at 1 day, 7 days, 14 days, and 30 days of culture, and stained with Calbindin D-28k antibody. The multimodal deep learning recognition system of the present invention was used for automated identification.

[0093] 2. Identification Results: The sample recognition accuracy rates at the four culture time points were 97.8%, 98.5%, 98.6%, and 98.3%, respectively, all maintaining a high accuracy rate of over 97.5%. This demonstrates that the multimodal deep learning recognition model of this method has stable recognition ability for Purkinje neurons at different differentiation stages and excellent generalization ability.

[0094] Comparative Example 1: Traditional purification culture + manual microscopic examination and identification (existing conventional techniques)

[0095] 1. Rat Purkinje cells were purified and cultured according to the method in Example 1. After 7 days of culture, Calbindin D-28k immunofluorescence staining was performed, and manual microscopic counting and identification were carried out by two laboratory personnel with more than 5 years of experience in neurobiology.

[0096] 2. Results: The average identification time for a single field of view under manual microscopy was 35 seconds, and the identification time for a whole plate of samples in a 96-well plate was about 12 hours; the consistency of the results of the two experimenters was 87.3%, with large batch-to-batch errors; the cell culture period was 7 days, and the cell survival rate during the purification process was 42%, which was far lower than the 78% cell survival rate under the mixed culture system of this invention.

[0097] Comparative Example 2: AdaBoost algorithm recognition based on a single morphological feature (existing technology)

[0098] 1. Mouse mixed culture samples were prepared according to the method in Example 2. Only bright-field morphological images were collected, and the AdaBoost algorithm (non-multimodal deep learning) with single morphological features in the original technology was used for identification.

[0099] 2. Results: The accuracy of this method in identifying Purkinje neurons in the mixed culture system was only 82.7%, with a false positive rate of 12.3%, which is far lower than the recognition accuracy of over 98% of the multimodal deep learning recognition model of this invention. It cannot effectively distinguish Purkinje neurons in the mixed system from other cerebellar neurons.

[0100] Comparative Example 3: Existing CNN Neuron Recognition Methods

[0101] 1. Samples were prepared according to the method in Example 2 and identified using a single CNN brain slice neuron recognition method (non-multimodal deep learning) disclosed in a patent published in 2022.

[0102] 2. Results: The method achieved an accuracy of 84.2% and a sensitivity of 80.5% for mixed culture samples. However, it could not adapt to the complex background of in vitro mixed culture and had a high false negative rate for low-abundance Purkinje neurons. It could not meet the experimental requirements and showed a significant gap compared with the multimodal deep learning recognition of the present invention.

[0103] Based on the above comparison (see above) Figure 5 The identification system constructed in this application integrates Purkinje neuron-specific fluorescent labeling features with cell morphological features to build a multimodal deep learning recognition model based on few-shot learning, and is equipped with, for example... Figure 6 The fully automated microscopic imaging and analysis system shown enables rapid and highly accurate identification of Purkinje neurons in a mixed culture system, eliminating the need for complex cell purification steps, significantly shortening the experimental cycle, and lowering the technical threshold.

Claims

1. A rapid identification system for cerebellar Purkinje neurons based on multimodal deep learning, comprising a hardware system and a software system, characterized in that: The software system includes an image acquisition and preprocessing module, a standard database construction and model training module, an automated recognition module, and a result output and verification module. The standard database construction and model training module builds a multimodal deep learning recognition model, which is based on a few-shot learning architecture. This few-shot learning architecture is adapted to the training needs of biological samples with limited labeled data. The multimodal deep learning recognition model can extract the fluorescence features, morphological features, and texture features of Purkinje neurons to complete feature matching and neuron recognition.

2. The rapid identification system for cerebellar Purkinje neurons based on multimodal deep learning according to claim 1, characterized in that: The core of the hardware system is a fully automated cerebellar cell microscopy imaging unit, which consists of an upright fluorescence microscope, a high-resolution image acquisition system with over 5 million pixels, a fully automated motorized stage, and an automatic focusing control system. It can automatically identify areas where cells are laid out in a sample, avoid overlapping cell fields of view, and simulate the logic of manual microscopy to complete automatic focusing, field of view traversal, and image acquisition.

3. The rapid identification system for cerebellar Purkinje neurons based on multimodal deep learning according to claim 1, characterized in that: The image acquisition and preprocessing module is used to interface with the hardware system to complete fully automatic image acquisition, simultaneously acquiring bright-field cell morphology images and corresponding fluorescently labeled images to form paired multimodal images; preprocessing the acquired images, and inputting the preprocessed images into the multimodal deep learning recognition model.

4. The rapid identification system for cerebellar Purkinje neurons based on multimodal deep learning according to claim 1, characterized in that: The multimodal deep learning recognition model uses Calbindin D-28k / L7 / Pcp2 / IP3R1 specific fluorescent staining as the gold standard, and collects Purkinje neuron pairing images from different species, different culture cycles, and different culture systems as a standard database. The feature inputs of the multimodal deep learning recognition model include fluorescence features, morphological features, and texture features.

5. The rapid identification system for cerebellar Purkinje neurons based on multimodal deep learning according to claim 1, characterized in that: The automated identification module is used to automatically generate an identification report, which includes the number and proportion of target cells, morphological parameter statistics, and a visual image of the identification marker.

6. A rapid identification method for cerebellar Purkinje neurons based on multimodal deep learning, characterized in that: Identification is performed using the identification system described in any one of claims 1-5.

7. The method for rapid identification of cerebellar Purkinje neurons based on multimodal deep learning according to claim 6, characterized in that: The identification method steps are as follows: S1: Pretreatment of cell culture vector; S2: Mixed culture of cerebellar neurons; S3: Purkinje neuron-specific marker; S4: Multimodal deep learning workflow: The sample is placed in a fully automated microscopic imaging system, and bright-field cell morphology images and corresponding field-of-view fluorescent label images are acquired simultaneously to form paired multimodal images. After preprocessing the images such as background denoising and contrast enhancement, they are input into the trained multimodal deep learning recognition model. The model automatically completes feature extraction and matching, locates and identifies Purkinje neurons, and finally outputs an identification report, which supports manual review and correction.

8. The method for rapid identification of cerebellar Purkinje neurons based on multimodal deep learning according to claim 7, characterized in that: The specific steps of S1 are as follows: the coverslips for inoculating cells are coated with 400-500 μg / mL poly-L-lysine solution, placed in a 35 mm culture dish, and incubated overnight at 37°C in a 5% CO2 incubator; the coverslips are washed twice with autoclaved PBS buffer, and then air-dried in a fume hood under sterile conditions for later use, providing a stable solid-phase carrier for cell adhesion and reducing cell overlap.

9. The method for rapid identification of cerebellar Purkinje neurons based on multimodal deep learning according to claim 7, characterized in that: In S3, specific labeling is achieved using immunofluorescence staining.

10. The method for rapid identification of cerebellar Purkinje neurons based on multimodal deep learning according to claim 7, characterized in that: The S3 method employs Purkinje neuron-specific adenovirus transfection with fluorescent labeling.

Images