Recognition counting and size analysis method for neurons in rat ganglion

Abstract

A method for identifying, counting and size analysis of neurons in rat ganglions relates to the technical field of image analysis, and comprises the following steps: scanning rat ganglion HE staining sections; identifying tissues on the dyed section and converting the tissues into a first region of interest ROI1; ganglions in the tissue are recognized on the first region of interest ROI1, corresponding marks are made, the area of the ganglions is counted, and the recognized ganglions are output as a second region of interest ROI2; and identifying neuronal cell bodies in the ganglion on the second region of interest ROI2 and making corresponding marks, and counting the number and average size of the neuronal cell bodies. According to the invention, artificial intelligence of neuronal counting and ganglion and neuronal size analysis can be realized, and rapid and accurate batch operation is realized.

Description

Technical Field

[0001] The invention relates to the technical field of image analysis, and in particular to a method for identifying, counting and sizing neurons in a rat ganglion. Background Art

[0002] The current method for identifying neurons in rat ganglia is for pathologists to count the neurons in the ganglia and then use image analysis tools to perform quantitative statistics and size analysis on the manually circled neurons. This is a large workload and requires a high level of practical work experience from the diagnosticians, and is not suitable for batch operations. Existing image analysis tools have limited ability to identify neuronal cells in ganglia and cannot replace the manual circle workflow for accurate analysis. Summary of the invention

[0003] Based on the above problems, the purpose of the present invention is to provide a method for identifying, counting and sizing neurons in rat ganglia, so as to achieve the purpose of identifying, counting and sizing neurons in rat ganglia on HE stained sections.

[0004] The technical solution adopted by the present invention to achieve its invention object is a method for identifying, counting and analyzing the size of neurons in rat ganglia, comprising the following steps:

[0005] S1. Slice scanning: Use a digital slice scanner to scan rat ganglion HE-stained slices;

[0006] S2, tissue identification: identifying the tissue on the stained section and converting it into a first region of interest ROI1;

[0007] S3, ganglion identification: identifying ganglia in the tissue on the first region of interest ROI1 and making corresponding marks, counting the area of ​​the ganglia, and outputting the identified ganglia as the second region of interest ROI2;

[0008] S4. Identification of neuron cell bodies: Identify the neuron cell bodies in the ganglion in the second region of interest ROI2 and make corresponding marks, and count the number and average size of the neuron cell bodies.

[0009] Furthermore, the specific steps of the tissue identification in step S2 are:

[0010] S2.1. Select a certain number of rat ganglion HE-stained sections as a training set, draw a region of interest on the stained sections of the training set, including tissue areas and non-tissue areas, mark all tissue areas in the region of interest accordingly, for example, mark them in blue, and mark the non-tissue areas as background, for example, mark them in green, then apply a deep learning U-Net model, i.e., a Deep Learning-U-Net model, for training, and adjust it during training until the tissue recognition accuracy reaches more than 90%;

[0011] S2.2, import the stained slices that need to be formally identified into the software that has completed the model training as described in step S2.1, draw the initial region of interest ROI0, and perform tissue identification in the initial region of interest ROI0, automatically make corresponding marks on the identified tissue area, and automatically mark the non-tissue area as background;

[0012] S2.3, and then post-processing, the post-processing steps include:

[0013] S2.3.1. Completely fill in the blank areas within the tissue;

[0014] S2.3.2, replace the tissue with an area smaller than the preset value N1 with the background to improve the accuracy of tissue recognition;

[0015] S2.3.3. Smooth the edges of the tissue area;

[0016] S2.3.4. Remove the background, such as the green background, so that only the tissue remains in the stained section;

[0017] S2.3.5. Convert the retained tissue into the first region of interest ROI1.

[0018] Furthermore, the preset value N1 is equal to 13000 μm2.

[0019] Furthermore, the specific steps of the step S3 ganglion identification are:

[0020] S3.1. Select a certain number of rat ganglion HE-stained sections as a training set, draw a region of interest on the stained sections of the training set, run step S2 in the region of interest, mark all ganglia in the first region of interest ROI1 obtained in step S2 accordingly, for example, mark them in blue, mark non-ganglionic areas as background, for example, mark them in green, and after marking, apply a deep learning U-Net model, i.e., a Deep Learning-U-Net model, for training, and adjust it during training until the ganglion recognition accuracy reaches more than 90%;

[0021] S3.2, importing the first region of interest ROI1 obtained in step S2 for formal ganglion identification into the software that has completed the model training as described in step S3.1, and performing ganglion identification, the identified ganglia are automatically marked accordingly, and the non-ganglionic area is automatically marked as background; for example, the ganglia are marked in blue, and the non-ganglionic area is automatically marked as background, for example, in green;

[0022] S3.3, then post-processing is performed, and the post-processing step includes:

[0023] S3.3.1. Fill the blank area in the ganglion completely;

[0024] S3.3.2. Smooth the edges of the ganglion;

[0025] S3.3.3, remove the ganglia whose area is smaller than the preset value N2 to improve the accuracy of ganglion identification;

[0026] S3.3.4, removing the background, such as the green background, so that only the blue ganglia, for example, remain in the stained slice;

[0027] S3.3.5. Convert the retained ganglion into the second region of interest ROI2.

[0028] S3.4. Finally, the area of ​​the ganglion retained in the first region of interest ROI1 is obtained by statistics.

[0029] Furthermore, the preset value N2 is equal to 5000 μm2.

[0030] Furthermore, the specific steps of step S4 neuron cell body identification are:

[0031] S4.1. Select a certain number of rat ganglion HE-stained sections as a training set, draw a region of interest on the stained sections of the training set, run steps S2 and S3 in sequence in the region of interest, make corresponding marks on all neuronal cell bodies in the second region of interest ROI2 obtained in step S3, for example, mark them in blue, mark the non-neuronal cell body area as the background, for example, mark them in green, add boundaries to all neuronal cell bodies, for example, add red boundaries, and after adding boundaries, the neuronal cell bodies can be divided into independent individuals, which is convenient for subsequent counting and area calculation of each neuronal cell body. After marking, apply the deep learning U-Net model, i.e., the Deep Learning-U-Net model, for training, and adjust it during training until the neuronal cell body recognition accuracy reaches more than 90%;

[0032] S4.2, importing the second region of interest ROI2 obtained in step S3 for formal neuronal cell body identification into the software that has completed the model training as described in step S4.1, and performing neuronal cell body identification, the identified neuronal cell bodies are automatically marked accordingly, and the non-neuronal cell body area is automatically marked as the background; for example, the neuronal cell body is marked in blue, the non-neuronal cell body area is automatically marked as the background, for example, green, and the boundary is marked in red;

[0033] S4.3, and then post-processing, the post-processing step includes:

[0034] S4.3.1. Fill the blank space in the neuron cell body completely;

[0035] S4.3.2. Smooth the edges of the neuron cell bodies;

[0036] S4.3.3, clear the border of the neuron cell body, such as the red border;

[0037] S4.3.4, remove the neuron cell bodies whose area is smaller than the preset value N3 to improve the accuracy of neuron cell body recognition;

[0038] S4.3.5, separates adjacent neuronal cell bodies;

[0039] S4.4. Finally, count the number and average size of neuron cell bodies.

[0040] Furthermore, the preset value N3 is equal to 20 μm 2 .

[0041] The beneficial effects of the present invention are:

[0042] The present invention can realize artificial intelligence of neuron counting, ganglion and neuron size analysis, making large-scale batch operation possible. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention;

[0044] Figure 2 This is a diagram showing the effect of step S2 in the embodiment of the present invention;

[0045] Figure 3 for Figure 2 The corresponding uncolored original image;

[0046] Figure 4 This is a diagram showing the effect of step S3 in the embodiment of the present invention;

[0047] Figure 5 for Figure 4 The corresponding uncolored original image;

[0048] Figure 6 This is a diagram showing the effect of step S4 in the embodiment of the present invention;

[0049] Figure 7 for Figure 6 The corresponding uncolored original image. DETAILED DESCRIPTION

[0050] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Figures 1 to 7 A specific implementation of the method for identifying, counting and analyzing the size of neurons in rat ganglia of the present invention is shown, and the steps include:

[0052] S1. Slice scanning: Use a digital slice scanner to scan rat ganglion HE-stained slices;

[0053] S2, tissue identification: identifying the tissue on the stained section and converting it into a first region of interest ROI1;

[0054] S3, ganglion identification: identifying ganglia in the tissue on the first region of interest ROI1 and making corresponding marks, counting the area of ​​the ganglia, and outputting the identified ganglia as the second region of interest ROI2;

[0055] S4. Identification of neuron cell bodies: Identify the neuron cell bodies in the ganglion in the second region of interest ROI2 and make corresponding marks, and count the number and average size of the neuron cell bodies.

[0056] In this embodiment, the specific steps of the tissue identification in step S2 are:

[0057] S2.1. A certain number of rat ganglion HE-stained sections were selected as training sets. Regions of interest were drawn on the stained sections of the training sets, including tissue regions and non-tissue regions. All tissue regions within the region of interest were marked as blue (class2), and non-tissue regions were marked as green (class1). The identification of tissue regions and non-tissue regions was completed by pathologists. After that, a deep learning U-Net model, namely the Deep Learning-U-Net model, was applied for training, and it was adjusted during training until the tissue recognition accuracy reached more than 90%;

[0058] S2.2, import the stained slices that need to be formally identified into the software that has completed the model training as described in step S2.1, draw the initial region of interest ROI0, and perform tissue identification in the initial region of interest ROI0, the identified tissue area is automatically marked in blue (class2), and the non-tissue area is automatically marked in green (class1);

[0059] S2.3, run the post processing program again, the post processing steps include:

[0060] S2.3.1, set the parameter Fill Holes Lbl:class2, that is, fill the blank area in the blue tissue completely;

[0061] S2.3.2, set the parameter change by shape Lbl: replace class2 with class1 with an area less than 13000μm2, that is, replace the blue tissue with an area less than 13000μm2 with a green background to improve the accuracy of tissue recognition;

[0062] S2.3.3, set the parameter Close Lbl:class 2 to smooth the edge of the tissue area;

[0063] S2.3.4, set the parameter change Lbl: replace class1 with clear, that is, clear the green background so that only blue tissue is retained in the stained section;

[0064] S2.3.5. Set the parameter outline as ROI Lbl: output class2 as ROI1, that is, convert the retained blue tissue into the first region of interest ROI1.

[0065] All subsequent processing is performed within this ROI1 to narrow the scope of subsequent program processing and shorten the processing time.

[0066] In this embodiment, the specific steps of step S3 ganglion identification are:

[0067] S3.1. A certain number of rat ganglion HE-stained sections are selected as a training set, and a region of interest is drawn on the stained sections of the training set. Step S2 is performed in the region of interest. The first region of interest ROI1 obtained in step S2 includes ganglia, interstitium, blood vessels, blank areas, etc. The pathologist marks all ganglia as blue (class2) and non-ganglionic areas as green (class1). After marking, a deep learning U-Net model, i.e., a Deep Learning-U-Net model, is used for training, and the model is adjusted during training until the ganglion recognition accuracy reaches more than 90%;

[0068] S3.2, importing the first region of interest ROI1 obtained in step S2 for formal ganglion identification into the software that has completed the model training as described in step S3.1, and performing ganglion identification, the identified ganglia are automatically marked in blue (class2), and the non-ganglionic areas are automatically marked in green (class1);

[0069] S3.3, run the post processing program again, the post processing steps include:

[0070] S3.3.1. Set the parameter Fill Holes Lbl:class2 to fill the blank area in the blue ganglion completely.

[0071] S3.3.2, set the parameter Close Lbl:class 2 to smooth the edge of the ganglion;

[0072] S3.3.3, set the parameter change by shape Lbl: replace class2 with an area less than 5000μm2 with clear, that is, clear the ganglia with an area less than 5000μm2 to improve the accuracy of ganglion recognition;

[0073] S3.3.4, set the parameter change Lbl: replace class1 with clear, that is, clear the green background so that only the blue ganglia are retained in the stained slices;

[0074] S3.3.5. Set the parameter outline as ROI Lbl: Output class2 as ROI2, that is, convert the retained blue ganglion into the second region of interest ROI2.

[0075] All subsequent processing is performed within this ROI2 to narrow the scope of subsequent program processing and shorten the processing time.

[0076] S3.4. Finally, select class2 in the output variables. After running, you can get the blue ganglion area retained in the second region of interest ROI2.

[0077] In this embodiment, the specific steps of step S4 neuron cell body identification are:

[0078] S4.1. Select a certain number of rat ganglion HE-stained sections as training sets, draw regions of interest on the stained sections of the training sets, and run steps S2 and S3 in sequence within the regions of interest. The second region of interest ROI2 obtained in step S3 contains neuronal cell bodies, neuronal synapses, interstitium, blood vessels, blank areas, etc. The pathologist marks all neuronal cell bodies in the second region of interest ROI2 obtained in step S3 as blue (class2), and marks non-neuronal cell body areas as green (class1). Call the boundary algorithm dilate boundary to add red boundaries to all neuronal cell bodies (class3). After adding the red boundaries, the neuronal cell bodies can be divided into independent individuals, which is convenient for subsequent counting and area calculation of each neuronal cell body. After marking, the deep learning U-Net model, i.e., the Deep Learning-U-Net model, is used for training, and is adjusted during training until the accuracy of neuronal cell body recognition reaches more than 90%;

[0079] S4.2, importing the second region of interest ROI2 obtained in step S3, which needs to be formally identified as a neuron cell body, into the software that has completed the model training as described in step S4.1, to identify the neuron cell body, and the identified neuron cell body is automatically marked as blue (class2), the non-neuron cell body area is automatically marked as green (class1), and the boundary is automatically marked as red (class3);

[0080] S4.3, run the post processing program again, the post processing steps include:

[0081] S4.3.1. Set the parameter Fill Holes Lbl:class2 to fill the blank area in the blue neuron cell body.

[0082] S4.3.2, set the parameter Close Lbl:class 2 to smooth the edges of the neuron cell bodies;

[0083] S4.3.3, set the parameter change Lbl: replace class 3 with clear, that is, clear the red border;

[0084] S4.3.4, set the parameter change by shape Lbl: replace class2 with clear whose area is less than 20μm2, and clear the neuron cell bodies with an area less than 20μm2 to improve the accuracy of neuron cell body recognition;

[0085] S4.3.5, set the parameters Separate Objects Lbl: separate the adjacent neuron cell bodies;

[0086] By setting the above parameters, blue neuron cell bodies were selected for analysis;

[0087] S4.4. Finally, select class2 in the output variables. After running, you can get the number and average size of neuron cell bodies.

[0088] The embodiment of the present invention is implemented based on the existing Visiopharm software platform, which is an image analysis software based on AI technology. In the specific implementation, other software platforms with similar functions can also be used.

Claims

1. A method for identifying, counting and sizing neurons in rat ganglia, characterized in that: The steps include: S1. Slice scanning: Use a digital slice scanner to scan rat ganglion HE-stained slices; S2, tissue identification: identifying the tissue on the stained section and converting it into a first region of interest ROI1; S3, ganglion identification: identifying ganglia in the tissue on the first region of interest ROI1 and making corresponding marks, counting the area of ​​the ganglia, and outputting the identified ganglia as the second region of interest ROI2; S4. Identification of neuron cell bodies: Identify the neuron cell bodies in the ganglion in the second region of interest ROI2 and make corresponding marks, and count the number and average size of the neuron cell bodies.

2. A method for identifying, counting and analyzing the size of neurons in a rat ganglion according to claim 1, characterized in that: The specific steps of tissue identification in step S2 are: S2.

1. Select a certain number of rat ganglion HE-stained sections as training sets, draw regions of interest on the stained sections of the training sets, including tissue regions and non-tissue regions, mark all tissue regions within the regions of interest, mark non-tissue regions as background, and then apply a deep learning U-Net model, i.e., a Deep Learning-U-Net model, for training, and adjust it during training until the tissue recognition accuracy reaches more than 90%; S2.2, import the stained slices that need to be formally identified into the software that has completed the model training as described in step S2.1, draw the initial region of interest ROI0, and perform tissue identification in the initial region of interest ROI0, automatically make corresponding marks on the identified tissue area, and automatically mark the non-tissue area as background; S2.3, and then post-processing, the post-processing steps include: S2.3.

1. Completely fill in the blank areas within the tissue; S2.3.2, replace the tissue with an area smaller than the preset value N1 as the background; S2.3.

3. Smooth the edges of the tissue area; S2.3.

4. Clear the background so that only the tissue remains on the stained section; S2.3.

5. Convert the retained tissue into the first region of interest ROI1.

3. A method for identifying, counting and analyzing the size of neurons in a rat ganglion according to claim 2, characterized in that: The preset value N1 is equal to 13000 μm2.

4. A method for identifying, counting and analyzing the size of neurons in a rat ganglion according to claim 1, characterized in that: The specific steps of step S3 ganglion identification are: S3.

1. Select a certain number of rat ganglion HE-stained sections as a training set, draw a region of interest on the stained sections of the training set, run step S2 in the region of interest, make corresponding marks on all ganglia in the first region of interest ROI1 obtained in step S2, mark the non-ganglionic area as the background, and apply the deep learning U-Net model, i.e., Deep Learning-U-Net model, for training after marking, and adjust it during training until the ganglion recognition accuracy reaches more than 90%; S3.2, importing the first region of interest ROI1 obtained in step S2 for formal ganglion identification into the software that has completed the model training as described in step S3.1, and performing ganglion identification, automatically making corresponding marks on the identified ganglia, and automatically marking the non-ganglionic area as background; S3.3, then post-processing is performed, and the post-processing step includes: S3.3.

1. Fill the blank area in the ganglion completely; S3.3.

2. Smooth the edges of the ganglion; S3.3.

3. Remove the ganglia whose area is smaller than the preset value N2; S3.3.4, remove the background so that only the ganglia remain in the stained sections; S3.3.

5. Convert the retained ganglion into the second region of interest ROI2. S3.

4. Finally, the area of ​​the ganglion retained in the first region of interest ROI1 is obtained by statistics.

5. A method for identifying, counting and analyzing the size of neurons in a rat ganglion according to claim 4, characterized in that: The preset value N2 is equal to 5000 μm2.

6. A method for identifying, counting and analyzing the size of neurons in a rat ganglion according to claim 1, characterized in that: The specific steps of step S4 neuron cell body identification are: S4.

1. Select a certain number of rat ganglion HE-stained sections as a training set, draw a region of interest on the stained sections of the training set, run steps S2 and S3 in sequence in the region of interest, make corresponding marks on all neuronal cell bodies in the second region of interest ROI2 obtained in step S3, mark the non-neuronal cell body area as the background, add boundaries to all neuronal cell bodies, divide the neuronal cell bodies into independent individuals, apply a deep learning U-Net model, i.e., DeepLearning-U-Net model, for training after marking, and adjust it during training until the neuronal cell body recognition accuracy reaches more than 90%; S4.2, importing the second region of interest ROI2 obtained in step S3, which needs to be formally identified as a neuron cell body, into the software that has completed the model training as described in step S4.1, to identify the neuron cell body, automatically mark the identified neuron cell body accordingly, and automatically mark the non-neuron cell body area as the background; S4.3, and then post-processing, the post-processing step includes: S4.3.

1. Fill the blank space in the neuron cell body completely; S4.3.

2. Smooth the edges of the neuron cell bodies; S4.3.3, clear the boundaries of the neuron cell body; S4.3.4, remove the neuron cell bodies whose area is smaller than the preset value N3; S4.3.5, separates adjacent neuronal cell bodies; S4.

4. Finally, count the number and average size of neuron cell bodies.

7. A method for identifying, counting and analyzing the size of neurons in a rat ganglion according to claim 6, characterized in that: The preset value N3 is equal to 20 μm 2 .

Images