Animal micronucleus auxiliary detection method and system based on deep learning

Abstract

The invention relates to an animal genetic toxicity experiment technology, and discloses an animal micronucleus cell auxiliary detection method and system based on deep learning, and the method comprises the steps: making an animal micronucleus image data set through a provided animal micronucleus test slide; constructing a network model of an animal micronucleus cell assisted detection algorithm based on deep learning for the manufactured animal micronucleus data set; enriching target information extracted by a backbone network through dynamic convolution for a network model constructed according to the animal micronucleus data set; carrying out convolution operation through a wavelet convolution layer so as to optimize the network model; and after the training is finished, testing the network model through the test set, and evaluating the performance of the network model. The target detection technology based on deep learning is gradually applied to the medical field, targets needed by medical staff can be rapidly screened out, the efficiency is improved, meanwhile, manpower waste is reduced, and a novel cell micronucleus visual analysis and detection strategy shows a good application prospect.

Description

Technical Field

[0001] The present invention relates to the technical field of animal genotoxicity experiments, and particularly to a method and system for assisting in the detection of micronucleated cells in animals based on deep learning. Background Art

[0002] By applying the micronucleus test, it is possible to obtain the effects on the target under conditions such as the administration dose, the number of administrations, and the observation time. The traditional micronucleus detection technology uses the method of manual microscopic examination, and uses the human eye to identify and count polychromatic erythrocytes, orthochromatic erythrocytes, and polychromatic micronucleated cells on the glass slide under the microscope. This requires a large amount of time for cell counting work; and it is impossible to ensure that the counting personnel can perform cell detection work efficiently and at a high level for a long time. Summary of the Invention

[0003] Aiming at the problems of cumbersome and time-consuming in the animal micronucleus test in the prior art, the present invention provides a method and system for assisting in the detection of micronucleated cells in animals based on deep learning.

[0004] In order to solve the above technical problems, the present invention is solved by the following technical solutions:

[0005] A method for assisting in the detection of micronucleated cells in animals based on deep learning, the method comprising:

[0006] Making an animal micronucleus data set, and making an animal micronucleus image data set through the provided animal micronucleus test glass slides;

[0007] Building a network model of an algorithm for assisting in the detection of micronucleated cells in animals based on deep learning for the made animal micronucleus data set;

[0008] The first adjustment of network model testing, enriching the target features extracted by the backbone network through dynamic convolution for the network model constructed based on the animal micronucleus data set;

[0009] The second adjustment of network model testing, optimizing the network model by performing convolution operations through a wavelet convolution layer;

[0010] After the training is completed, the network model is tested by a test set, and the performance of the network model is evaluated.

[0011] Preferably, after optimizing the network model by performing convolution operations through a wavelet convolution layer, it further includes the third adjustment of network model testing, and adjusting the network model by introducing an improved SPPF model.

[0012] Preferably, for the production of the animal micronucleus dataset, the customer provides the animal micronucleus test slides. Target images are collected from the provided animal micronucleus test slides, and the collected target images are labeled using the Labelme annotation tool according to the detection target categories. The labeled dataset is then divided to produce the animal micronucleus dataset.

[0013] Preferably, the detection target categories are divided into polychromatic cells, orthochromatic cells, and polychromatic micronucleated cells.

[0014] Preferably, the training of the network model includes

[0015] Dataset preparation: The animal micronucleus dataset produced is divided into a training set, a validation set, and a test set in a ratio of 8:1:1;

[0016] Building the network model: Dynamic convolution and wavelet convolution are introduced into the backbone network;

[0017] Model hyperparameter configuration: Modify the hyperparameters in the configuration file in the source code of the used model, and select appropriate hyperparameters by referring to experimental data and combining with one's own experiments.

[0018] Preferably, for the first-step adjustment of the network model test, for the network model constructed based on the animal micronucleus dataset, the information of the intermediate feature map is increased through dynamic convolution;

[0019]

[0020] where X is the input feature, W1, W2…W i are the convolution kernels, and α i is the contribution degree of each expert; * represents the convolution operation.

[0021] Preferably, for the second-step adjustment of the network model test, the network model is optimized by performing convolution operations through the wavelet convolution layer; The implementation method is to perform convolution in the wavelet domain, generate information maps of multiple frequency bands from the input features, and then perform feature fusion through inverse transformation,

[0022] Y = IWT(Conv(W, WT(X)))

[0023] where X is the input feature, W is the weight of a k×k depth convolution kernel, WT(*) is the wavelet transform, IWT(*) is the inverse wavelet transform, Conv(*) is the convolution, and Y is the output feature. Preferably, after the training is completed, the network model is tested using the test set, and the performance of the network model is evaluated. The metric used for model evaluation is the mean average precision mAP to evaluate the performance of the model,

[0024]

[0025] where N is the number of categories, and AP i is the average precision of the i-th category; AP is the average precision of the model at different recall rates;

[0026]

[0027] where P is precision, R is recall, TP is true positive, FP is false positive, and FN is false negative.

[0028] Preferably, in the third step of adjusting the network model test, the network model is adjusted by introducing an improved SPPF model, and the effective information in the feature map is extracted and fused through multiple branches by SENEtV2.

[0029] SENetV2 = x + F(x · Ex(∑Sq(x)))

[0030] where x is the input feature map, F(*) is a non-linear transformation function; Ex(*) is a channel-wise feature extraction operation; Sq(*) is a Squeeze operation that compresses the spatial information of the feature map into a channel descriptor through global average pooling to generate a channel feature vector.

[0031] To solve the above technical problems, the present invention also provides an animal micronucleus cell assisted detection system based on deep learning, which includes:

[0032] An animal micronucleus dataset production module for producing an animal micronucleus image dataset by providing animal micronucleus test slides.

[0033] A network model construction module for constructing a network model of an animal micronucleus cell assisted detection algorithm based on deep learning for the produced animal micronucleus dataset.

[0034] A first-step adjustment module for the network model test, which enriches the target features extracted by the backbone network through dynamic convolution for the network model constructed based on the animal micronucleus dataset.

[0035] A second-step adjustment module for the network model test, which optimizes the network model by performing convolution operations through wavelet convolutional layers.

[0036] A network model test module for testing the network model with a test set after training and evaluating the performance of the network model.

[0037] Due to the adoption of the above technical solutions, the present invention has significant technical effects:

[0038] The object detection technology based on deep learning in the present invention has gradually been applied in the medical field. It can quickly screen out the targets needed by medical staff, improve efficiency and reduce waste of manpower. The new cell micronucleus visualization analysis and detection strategy shows good application prospects.

[0039] The present invention designs an auxiliary micronucleus test detection system based on deep learning, which greatly improves the review efficiency of experimenters during the micronucleus experiment and greatly alleviates the work burden of experimenters.

[0040] The present invention adds dynamic convolution, wavelet convolution, and SENetV2 structure on the basis of YOLOv11. On the basis of ensuring the detection performance of polychromatic cells and orthochromatic cells, the detection performance of the model in polychromatic micronuclei is improved. Brief Description of the Drawings

[0041] Figure 1 It is the optimization training flow chart of the animal test detection model in the embodiment of the present invention;

[0042] Figure 2 It is the structure diagram of the animal test detection model in the embodiment of the present invention;

[0043] Figure 3-1 It is the schematic diagram of all intercepted results of the detection results of the improved model in the embodiment of the present invention; Figure 3-2 Schematic diagram of all intercepted results of the detection results of the improved model in the embodiment of the present invention

[0044] Figure 4-1 It is the confusion matrix diagram of the original model of the present invention; Figure 4-2 It is the confusion matrix diagram of the improved model of the present invention;

[0045] Figure 5-1 It is the P-R curve diagram of the original model of the present invention, Figure 5-2 It is the P-R curve diagram of the improved model of the present invention.

[0046] Among them, SPPF (Simplified Spatial Pyramid Pooling), spatial pyramid pooling;

[0047] SENetV2 (Squeeze-and-ExcitationV2) the second generation squeeze-and-excitation network;

[0048] DynamicConv dynamic convolution;

[0049] C3K2 is the C3 module with 2x2 convolution;

[0050] WTConv (Wavelet Transform Convolution) wavelet transform convolution. Detailed implementation mode

[0051] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

[0052] Embodiment 1

[0053] An animal micronucleus cell assisted detection method based on deep learning, the method includes:

[0054] Production of an animal micronucleus data set, making an animal micronucleus image data set through the provided animal micronucleus test slides; obtaining corresponding animal micronucleus test slides from customers, and collecting more than 5,000 target images of each field of view under a microscope with a magnification of more than 100 times; formulating the detection target categories as: polychromatic erythrocytes (PCE), orthochromatic erythrocytes (NCE), and polychromatic micronucleus cells (PCE-MN); then annotating the collected animal micronucleus images through the annotation software Labelme, and conducting expert review after annotation; according to the model used, making the animal micronucleus data set into a YOLO format data set, and the ratio of the training set, validation set, and test set is 8:1:1.

[0055] Build a network model for an animal micronucleus cell assisted detection algorithm based on deep learning for the produced animal micronucleus data set;

[0056] Use YOLOv11 as the basic architecture to conduct the detection work of the animal micronucleus test: YOLOv11 introduces the C3k2 mechanism and the cross-channel partial self-attention (C2PSA) mechanism, which can achieve more accurate detection tasks; YOLOv11 improves the architecture design, realizes a faster processing speed, and well balances accuracy and performance, meeting the requirements of this task.

[0057] The first step of network model test adjustment is to enrich the target information extracted by the backbone network through dynamic convolution for the network model constructed based on the animal micronucleus data set; the introduction of dynamic convolution can significantly increase the number of model parameters without increasing additional floating-point operations; it can enable the model to better adapt to different input features, thereby improving the generalization ability of the model.

[0058] The second step of network model test adjustment is to optimize the network model by performing convolution operations through wavelet convolutional layers; further improve the detection performance of YOLOv11, and use wavelet convolutional layers to perform convolution operations in the C3k2 structure to better capture local and global features. The first step of network model test adjustment is to increase the information of the intermediate feature map through dynamic convolution for the network model constructed based on the animal micronucleus data set;

[0059]

[0060] Among them, X is the input feature, W1, W2...W i is the convolution kernel, and α i is the contribution degree of each expert; * represents the convolution operation, and Y is the output feature.

[0061] The second step of network model testing is to adjust by performing convolution operations through the wavelet convolution layer to optimize the network model; the implementation method is to perform convolution in the wavelet domain, generate information maps of multiple frequency bands from the input features, and then perform feature fusion through inverse transformation,

[0062] Y = IWT(Conv(W, WT(X)))

[0063] Among them, X is the input feature, W is the weight of a k×k depth convolution kernel, WT(*) is the wavelet transform, IWT(*) is the inverse wavelet transform, Conv(*) is the convolution, and Y is the output feature.

[0064] After the training is completed, the network model is tested through the test set, and the performance of the network model is evaluated.

[0065] After the training is completed, the network model is tested through the test set, and the performance of the network model is evaluated. The metric used for model evaluation is the mean average precision mAP to evaluate the performance of the model,

[0066]

[0067] Among them, N is the number of categories, and AP i is the average precision of the i-th category; AP is the average precision of the model at different recall rates;

[0068]

[0069] Among them, P is the precision, R is the recall rate, TP is the true positive, FP is the false positive, and FN is the false negative.

[0070] Example 2

[0071] Based on Example 1, after this example optimizes the network model by performing convolution operations through the wavelet convolution layer, it further includes the third step of network model testing to adjust the network model by using the SPPF model adopted. The main purpose of the adopted SPPF model is to fuse information at a larger scale to improve the performance of object detection and further enhance the representation ability of the model; the present invention uses the SPPFSENetV2 structure to achieve the goal of enriching the semantic features of the feature map

[0072] The third step of network model testing and adjustment: The network model is adjusted by adopting the SPPF model. The SENEtV2 is used to extract and fuse the effective information in the feature map of multiple branches.

[0073] SENetV2 = x + F(x · Ex(∑Sq(x)))

[0074] Among them, x is the input feature map, F(*) is a non-linear transformation function; Ex(*) is a per-channel feature extraction operation; Sq(*) is a Squeeze operation that compresses the spatial information of the feature map into a channel descriptor through global average pooling to generate a channel feature vector.

[0075] In Figure 2 In the work of improving the network model, the modification is mainly carried out in the feature extraction network and the feature fusion network of the model. First, in the feature extraction network - backbone network, the ordinary convolution in the original model is replaced by dynamic convolution to realize the extraction of the input feature map through multiple convolution kernels and enrich the extracted feature information; wavelet convolution is introduced into the C3K2 structure to add a branch path to the input feature map to obtain the feature information after the wavelet domain, and then it is fused with the input feature map for the next operation to increase the expression ability of effective features; the SENetV2 attention mechanism is added to the SPPF to enrich the features and improve the generalization ability of the model. The above improvements are to better extract the effective features of the input image, so as to achieve the goal of improving the detection accuracy of the model. Adding SENetV2 after feature fusion can help the model better focus on the key feature channels, reduce the interference of redundant information, and in the feature map of small targets, it will also strengthen the channels that capture fine-grained features and increase the detection performance of small targets.

[0076] In Figure 3, compared with the detection results of the original model, the modified model improves the detection accuracy of the target to be detected, and can correctly identify impurities or other non-detection targets in the test image. From the confusion matrix obtained by testing the test set in Figure 4, it can be seen that the recall rates of all categories of the improved model are higher than those of the original model; Figure 5 shows the AP data of all categories, and the AP values of all categories of the improved model are higher than those of the original model, especially in the small target category of PCE-MN.

[0077] Example 3

[0078] Based on the above embodiments, this embodiment is an animal micronucleus cell assisted detection system based on deep learning, which includes:

[0079] An animal micronucleus dataset production module, which produces an animal micronucleus image dataset by providing animal micronucleus test slides.

[0080] The network model construction module constructs a network model of an animal micronucleus cell auxiliary detection algorithm based on deep learning for the made animal micronucleus dataset;

[0081] The first-step adjustment module of network model testing enriches the target features extracted by the backbone network through dynamic convolution for the network model constructed based on the animal micronucleus dataset;

[0082] The second-step adjustment module of network model testing optimizes the network model by performing convolution operations through wavelet convolutional layers;

[0083] The network model testing module tests the network model with the test set after training and evaluates the performance of the network model.

Claims

1. Animal micronucleus cell-assisted detection method based on deep learning, the method comprising: Animal micronucleus dataset preparation: animal micronucleus image dataset is prepared using the provided animal micronucleus test slides; For the prepared animal micronucleus dataset, a network model of the animal micronucleus cell auxiliary detection algorithm based on deep learning was built; The first step of network model testing is to adjust the network model built based on the animal micronucleus dataset by using dynamic convolution to enrich the target features extracted by the backbone network; The second step of network model testing is to optimize the network model by performing convolution operations through the wavelet convolution layer; After the training is completed, the network model is tested through the test set, and the performance of the network model is evaluated.

2. The animal micronucleus cell-assisted detection method based on deep learning according to claim 1, characterized in that: After optimizing the network model by performing convolution operation through wavelet convolution layer, the third step of network model testing is also included, in which the network model is adjusted through the improved SPPF model.

3. The animal micronucleus cell-assisted detection method based on deep learning according to claim 1, characterized in that: Animal micronucleus dataset production: the customer provides animal micronucleus test slides, and the target images are collected from the provided animal micronucleus test slides. The collected target images are annotated using the Labelme annotation tool according to the detection target category, and the annotated dataset is divided to produce an animal micronucleus dataset.

4. The animal micronucleus cell-assisted detection method based on deep learning according to claim 3, characterized in that: The detection target categories are divided into polychromatic cells, orthochromatic cells and polychromatic micronucleated cells.

5. The animal micronucleus cell-assisted detection method based on deep learning according to claim 1, characterized in that: The training of the network model includes Dataset preparation: To make the animal micronucleus dataset, the dataset was divided into training set, validation set, and test set with a ratio of 8:1:1; Build a network model and introduce dynamic convolution and wavelet convolution into the backbone network; Model hyperparameter configuration: modify the hyperparameters in the configuration file in the source code of the model used, and select appropriate hyperparameters by referring to experimental data and combining your own experiments.

6. The animal micronucleus cell-assisted detection method based on deep learning according to claim 1, characterized in that: The first step of the network model test is to adjust the network model built based on the animal micronucleus dataset by adding information to the intermediate feature map through dynamic convolution; Among them, X is the input feature, W1, W2…W i is the convolution kernel, α i is the contribution of each expert; * represents the convolution operation, and Y is the output feature.

7. The animal micronucleus cell-assisted detection method based on deep learning according to claim 1, characterized in that: The second step of network model testing is to optimize the network model by performing convolution operations through wavelet convolution layers. This is achieved by performing convolution in the wavelet domain, generating information graphs of multiple frequency bands from the input features, and then performing feature fusion through inverse transformation. Y = IWT(Conv(W, WT(X))) Among them, X is the input feature, W is the weight of a k×k deep convolution kernel, WT(*) is the wavelet transform, IWT(*) is the inverse wavelet transform, Conv(*) is the convolution, and Y is the output feature.

8. The animal micronucleus cell-assisted detection method based on deep learning according to claim 1, characterized in that: After the training is completed, the network model is tested through the test set, and the performance of the network model is evaluated. The indicator used for model evaluation is the mean average accuracy (mAP) to evaluate the performance of the model. Where N is the number of categories, AP i is the average precision of the i-th category; AP is the average precision of the model under different recall rates; Among them, P is precision, R is recall, TP is true positive, FP is false positive, and FN is false negative.

9. The animal micronucleus cell-assisted detection method based on deep learning according to claim 2, characterized in that: The third step of network model testing is to adjust the network model through the improved SPPF model. SENEtV2 implements multiple branches to extract and fuse effective information in the feature graph. SENetV2=x+F(x·Ex(∑Sq(x))) Among them, x is the input feature map, F(*) is the nonlinear transformation function; Ex(*) is the channel-by-channel feature extraction operation; Sq(*) is the Squeeze operation that compresses the spatial information of the feature map into a channel descriptor through global average pooling to generate a channel feature vector.

10. Animal micronucleus cell auxiliary detection system based on deep learning, characterized in that: include: Animal micronucleus dataset production module, which produces animal micronucleus image datasets using the provided animal micronucleus test slides; The network model building module builds a network model of the animal micronucleus cell auxiliary detection algorithm based on deep learning for the prepared animal micronucleus dataset; The first step of network model testing is to adjust the module. For the network model built based on the animal micronucleus dataset, dynamic volume is used to enrich the target features extracted by the backbone network. The second step of network model testing is to adjust the module, which performs convolution operation through the wavelet convolution layer to optimize the network model; The network model testing module tests the network model through the test set after training and evaluates the performance of the network model.

Images