Full-view digital slice identification method combining dynamic detection and region optimization

Abstract

The invention belongs to the technical field of medical image recognition, and particularly discloses a full-view digital slice recognition method combining dynamic detection and region optimization, which comprises the following steps of: performing layered extraction and labeling on an original full-view data slice; carrying out initial training on the Yolov8 model by using the low-level data slice image and the annotation information; the initially trained Yolov8 model dynamically monitors the remaining data slice images of each level to obtain slices containing a target area; carrying out region optimization training on the initially trained Yolov8 model by utilizing a slice containing a target region; and inputting the full-view data slice to be identified into the full-view data slice identification model to complete full-view digital slice identification. According to the method, dynamic detection and region optimization are combined, the problems that an existing full-view data slice processing technology is high in investment cost and complex in data processing are solved, and full-view digital slices can be efficiently processed and accurately recognized.

Description

Technical Field

[0001] The invention belongs to the technical field of medical image recognition, and in particular relates to a full-view digital slice recognition method combining dynamic detection and region optimization. Background Art

[0002] Breast cancer is one of the most common malignant tumors in Chinese women, and its incidence is on the rise. Early detection and treatment of breast cancer are crucial to improving the survival rate of patients, and the whole-view digital slide (WSI) technology plays an important role in the diagnosis of breast cancer. It facilitates the long-term preservation of pathological images, remote pathological analysis, and improves the consistency and accuracy of pathological evaluation through high-resolution scanning of pathological slides.

[0003] The application of computer vision technology to full-field digital slides is an active research area, especially in the automatic recognition of pathology images. WSI technology can convert traditional pathology slides into high-resolution digital images, providing computer vision algorithms with a large amount of data for analysis and pattern recognition. Existing technologies have explored different global structural information fusion methods and found that the information captured at the local small area (patch) level is the most important for WSI classification.

[0004] However, the high resolution and huge amount of data of WSI images pose a huge challenge to computing resources. These algorithms and models require a lot of hardware resources, and when the algorithms are put into use, they also impose high computing cost requirements on edge computing devices. They can explain specific problems in terms of principle and feasibility, but they cannot achieve efficient and economical training and lightweight deployment. How to efficiently extract WSI image features, how to save resources for training and how to deploy at low cost is an urgent problem to be solved.

[0005] Currently, full-view data slices have high requirements for hardware computing resources due to the huge amount of data and high image resolution. This is not only reflected in the high cost investment in the training stage, but also in the test and use stage. The operation of the model requires complex data processing to achieve effective detection. Multiple instance learning (MIL) technology also faces challenges when processing WSI, including complex relationships between instances, possible information loss during instance aggregation, and imbalanced data categories. Summary of the invention

[0006] The purpose of the present invention is to solve the problems of high investment cost and complex data processing in the existing full-view data slice processing technology, and proposes a full-view digital slice recognition method combining dynamic detection and regional optimization.

[0007] The technical solution of the present invention is: a full-view digital slice recognition method combining dynamic detection and regional optimization, comprising the following steps:

[0008] S1. Obtaining original full-view data slices, performing layered extraction on the original full-view data slices according to the ratio of resolution, obtaining data slice images of different levels, and obtaining annotation information of each level for the data slice images of different levels;

[0009] S2. Select a low-level data slice image and annotation information that can be loaded into the Yolov8 model, input the Yolov8 model for preliminary training, and obtain a preliminarily trained Yolov8 model;

[0010] S3. Use the initially trained Yolov8 model to dynamically monitor the data slice images of the remaining layers to obtain slices containing the target area;

[0011] S4. Using the slices containing the target area to perform regional optimization training on the initially trained Yolov8 model, a full-view data slice recognition model is obtained;

[0012] S5. Input the full-view data slice to be identified into the full-view data slice identification model for classification and identification, output the target area identification result, and complete the full-view digital slice identification.

[0013] The beneficial effects of the present invention are:

[0014] 1. The present invention combines dynamic detection and regional optimization, so that full-field digital slices can be efficiently processed and accurately identified.

[0015] 2. The two-stage training of dynamic detection and Yolov8 model not only processes large-size WSI images quickly and efficiently, but also effectively enables the Yolov8 model to assist the WSI segmentation process while learning the characteristics of the target area;

[0016] 3. Due to the design of the regional optimization method, the training accuracy of the Yolov8 model in different levels of resolution is effectively guaranteed, and Yolov8 can effectively be compatible with different resolutions when performing target area recognition in WSI.

[0017] 4.Yolov8 is an edge computing-friendly model. The trained model can be directly deployed on mobile devices or personal hosts, achieving low-cost application and making it easier for individual users to deploy and use.

[0018] Preferably, the step S1 specifically includes the following sub-steps:

[0019] S11. Obtaining the original full-view data slices, performing hierarchical extraction on the large-size full-view data slices according to the ratio of the resolution, and obtaining data slice images of different levels;

[0020] S12. Select a low-level data slice image that can be loaded into the Yolov8 model for annotation, and obtain annotation information of the low-level data slice image, wherein the annotation box in the annotation information is marked with the target area position;

[0021] S13. According to the ratio of the resolution, the annotation frame of the annotation information of the low-level data slice image is enlarged or reduced to obtain the annotation information of each level.

[0022] Preferably, the Yolov8 model in step S2 includes a backbone network, a connection network and a head network connected in sequence;

[0023] The backbone network is used to receive the data slice image and perform feature extraction, and output a feature map;

[0024] The connection network is used to receive the feature map and perform feature fusion on the feature map to output a fused feature;

[0025] The head network is used to receive the fused features, perform target detection and recognition, and output a target area recognition result.

[0026] Preferably, the backbone network comprises a first convolutional layer, a second convolutional layer, a first C2F layer, a third convolutional layer, a second C2F layer, a fourth convolutional layer, a third C2F layer, a fifth convolutional layer, a fourth C2F layer and an SPPF layer connected in sequence;

[0027] The input end of the first convolutional layer is the input end of the entire backbone network, and the output end of the SPPF layer, the output end of the second C2F layer and the output end of the third C2F layer are all connected to the connection network;

[0028] The output ends of the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer and the fifth convolutional layer are all provided with a regional optimization mechanism.

[0029] Preferably, the regional optimization mechanism is specifically as follows: copying the feature maps output by each convolutional layer into feature maps I and feature maps J, where i represents the position number of the element on feature map I, and j represents the position number of the element on feature map J; using the element with position number i to calculate the cosine similarity of all elements on feature map J, and dividing the sum of all cosine similarities by the number of elements of the feature map output by the convolutional layer to obtain a new feature map after regional optimization; repeating the above steps until all elements on feature map I have completed the similarity calculation of all elements of feature map J, and the calculation process ends;

[0030] The calculation formula of the regional optimization mechanism is:

[0031]

[0032] Among them, y i Represents the new feature map after region optimization, x i represents the feature map output by each convolutional layer in the backbone network corresponding to the position number i of the element on the feature map I, x j represents the feature map output by each convolutional layer in the backbone network corresponding to the position number j of the element on the feature map J, f(·,·) represents the function for calculating the similarity, g(·) represents the linear change function, and M represents the number of elements in the feature map output by the convolutional layer.

[0033] The beneficial effects of the above preferred solution are:

[0034] By adding a regional optimization mechanism to the backbone network, the training accuracy of the Yolov8 model at different levels of resolution is effectively guaranteed, and Yolov8 can effectively be compatible with different resolutions when performing target area recognition in WSI.

[0035] Preferably, the connection network comprises a first feature fusion module, a second feature fusion module, a third feature fusion module and a fourth feature fusion module which are connected in sequence;

[0036] The input end of the first feature fusion module is also connected to the output end of the third C2F layer;

[0037] The input end of the second feature fusion module is also connected to the output end of the second C2F layer.

[0038] Preferably, the head network adopts a decoupling head structure, and the decoupling head structure includes a classification branch and a regression branch.

[0039] Preferably, the classification loss function of the classification branch during the training process includes a binary cross entropy loss function and a Varifocal Loss loss function, and the calculation formula thereof is:

[0040]

[0041] Among them, BCE represents the binary cross entropy loss function; N represents the number of samples; y i represents the true label; p i represents the probability predicted by the model; VFL represents the Varifocal Loss function; q i It represents the overlap between the predicted box and the true box; log represents the logarithmic function with the natural logarithm e as the base.

[0042] Preferably, the regression loss function of the regression branch during the training process includes a CIoU loss function and a Distribution Focal Loss loss function, and the calculation formula thereof is:

[0043]

[0044] Among them, CIoU represents the CIoU loss function, IoU represents the intersection-over-union ratio, ρ represents the distance between the center point of the predicted box and the real box, c represents the diagonal length of the minimum closed area, w represents the width of the predicted box and the real box, h represents the height of the predicted box and the real box, Area(A∪B) represents the intersection area of ​​the detection box A and the annotation box B, Area(A∩B) represents the union area of ​​the detection box A and the annotation box B, DFL represents the distribution focal loss function, P(yi) represents the label y is located at the real label y i The probability of S i Indicates that the predicted box is located at the true label y i The predicted probability at , arctan represents the inverse tangent function, log represents the logarithmic function with the natural logarithm e as the base, π represents pi, and N represents the number of samples.

[0045] The beneficial effects of the above preferred solution are:

[0046] By combining the above loss functions, YOLOv8 can significantly improve detection accuracy while maintaining real-time performance. Through these optimizations, YOLOv8 performs well in target detection tasks, with significant improvements in both speed and accuracy.

[0047] Preferably, the dynamic monitoring in step S3 is specifically as follows: a detection frame of a fixed size is set, each digital slice image of the remaining levels is detected from top to bottom and from left to right, and all slices whose proportion of the target area to the detection frame exceeds a threshold are screened out to obtain slices containing the target area.

[0048] The beneficial effects of the above preferred solution are:

[0049] Through dynamic detection of fixed size, when processing large-scale WSI images, both low resource requirements for processing and efficient and accurate feature learning are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Shown is a flow chart of a full-field digital slice recognition method combining dynamic detection and region optimization.

[0051] Figure 2 The figure shows a schematic diagram of the backbone network structure provided in Example 1 of the present invention.

[0052] Figure 3 Shown are the result graphs of mAP50 and mAP50-90 of the dynamic detection and region optimization method in Example 2 of the present invention.

[0053] Figure 4 It is a curve chart showing the change of mAP50 of the dynamic detection and region optimization method in Example 2 of the present invention.

[0054] Figure 5 It is a curve chart showing the change of mAP50-90 of the dynamic detection and region optimization method in Example 2 of the present invention. DETAILED DESCRIPTION

[0055] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the accompanying drawings are only exemplary and are intended to explain the principles and spirit of the present invention, rather than to limit the scope of the present invention.

[0056] Embodiment 1, as Figure 1 As shown, a full-view digital slice recognition method combining dynamic detection and region optimization includes the following steps:

[0057] S1. Obtaining the original full-view data slices, performing layered extraction on the original full-view data slices according to the ratio of resolution, obtaining data slice images of different levels, and obtaining annotation information of each level for the data slice images of different levels;

[0058] S2. Select a low-level data slice image and annotation information that can be loaded into the Yolov8 model, input the Yolov8 model for preliminary training, and obtain a preliminarily trained Yolov8 model;

[0059] S3. Use the initially trained Yolov8 model to dynamically monitor the data slice images of the remaining layers to obtain slices containing the target area;

[0060] S4. Using the slices containing the target area to perform regional optimization training on the initially trained Yolov8 model, a full-view data slice recognition model is obtained;

[0061] S5. Input the full-view data slice to be identified into the full-view data slice identification model for classification and identification, output the target area identification result, and complete the full-view digital slice identification.

[0062] In this embodiment, step S1 specifically includes the following sub-steps:

[0063] S11. Obtaining the original full-view data slices, performing hierarchical extraction on the large-size full-view data slices according to the ratio of resolution, and obtaining 9 levels of data slice images, wherein the size of the original full-view data slices is usually 4-5G, and the size of the data slice images after hierarchical extraction is about 5M;

[0064] S12. Select a low-level data slice image that can be loaded into the Yolov8 model for annotation, and obtain annotation information of the low-level data slice image, wherein the annotation box in the annotation information is marked with the target area position;

[0065] S13. According to the ratio of the resolution, the annotation frame of the annotation information of the low-level data slice image is enlarged or reduced to obtain the annotation information of each level.

[0066] In this embodiment, the Yolov8 model in step S2 includes a backbone network, a connection network and a head network connected in sequence;

[0067] The backbone network is used to receive the data slice image and perform feature extraction, and output a feature map;

[0068] The connection network is used to receive the feature map and perform feature fusion on the feature map to output a fused feature;

[0069] The head network is used to receive the fused features, perform target detection and recognition, and output a target area recognition result.

[0070] In this embodiment, if Figure 2 As shown, the backbone network includes a first convolutional layer, a second convolutional layer, a first C2F layer, a third convolutional layer, a second C2F layer, a fourth convolutional layer, a third C2F layer, a fifth convolutional layer, a fourth C2F layer and an SPPF layer connected in sequence;

[0071] The input end of the first convolutional layer is the input end of the entire backbone network, and the output end of the SPPF layer, the output end of the second C2F layer and the output end of the third C2F layer are all connected to the connection network;

[0072] The second C2F layer outputs a feature map of size 80*80*256 to the connection network;

[0073] The third C2F layer outputs a feature map of size 40*40*512 to the connection network;

[0074] The SPPF layer outputs a feature map of size 20*20*512 to the connection network.

[0075] The output ends of the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer and the fifth convolutional layer are all provided with a regional optimization mechanism.

[0076] In this embodiment, the region optimization mechanism is specifically as follows: copying the feature maps output by each convolution layer into feature maps I and feature maps J, where i represents the position number of the element on feature map I, and j represents the position number of the element on feature map J; using the element with position number i to calculate the cosine similarity of all elements on feature map J, and dividing the sum of all cosine similarities by the number of elements of the feature map output by the convolution layer to obtain a new feature map after region optimization; repeating the above steps until all elements on feature map I have completed the similarity calculation of all elements of feature map J, and the calculation process ends;

[0077] The calculation formula of the regional optimization mechanism is:

[0078]

[0079] Among them, y i Represents the new feature map after region optimization; x i represents the feature map output by each convolutional layer in the backbone network corresponding to the position number i of the element on the feature map I; j represents the feature map output by each convolutional layer in the backbone network corresponding to the position number j of the element on the feature map J; f(·,·) represents the function for calculating the similarity, and the dot product operation is usually selected to calculate the similarity between elements; g(·) represents the linear change function, which is designed as a 1×1 convolution in the regional optimization mechanism to adjust the output dimension after regional calculation; M represents the number of elements in the feature map output by the convolutional layer.

[0080] In this embodiment, the connection network includes a first feature fusion module, a second feature fusion module, a third feature fusion module and a fourth feature fusion module connected in sequence; the input end of the first feature fusion module is also connected to the output end of the third C2F layer; the input end of the second feature fusion module is also connected to the output end of the second C2F layer; the idea of ​​each feature fusion module is to first process the feature map into a fixed size, and then add the channel dimension of the feature maps with the same size, specifically:

[0081] The first feature fusion module is used to receive a feature map of size 20*20*512 and a feature map of size 40*40*512, and upsample the feature map of size 20*20*512 to a size of 40*40*512, and then add the feature map of size 40*40*512 received in the connection network and the feature map obtained by the upsampling in channel dimension to obtain a feature map of size 40*40*1024, and compress the feature map to a feature map of size 40*40*256;

[0082] The second feature fusion module is used to receive the feature map of 40*40*256 size output by the first feature fusion module, and upsample it to a feature map of 80*80*256 size, and then add the feature map of 80*80*256 size received in the connection network and the feature map obtained by the upsampling in terms of channel dimension to obtain a feature map of 80*80*512 size, and compress the feature map of 80*80*512 size into a feature map of 80*80*256 size for regression and classification calculation;

[0083] The third feature fusion module is used to receive the feature map of 80*80*256 size output by the second feature fusion module, compress it into a feature map of 40*40*256 size, and then add it to the feature map of 40*40*256 size output by the first feature fusion module to obtain a feature map of 40*40*512 size for regression and classification calculations.

[0084] The fourth feature fusion module is used to receive the feature map of 40*40*512 size output by the third feature fusion module, compress it into a feature map of 20*20*512 size, and then add the feature map of 20*20*512 size received from the backbone network and the feature map in channel dimension to obtain a feature map of 20*20*1024 size, and compress the 20*20*1024 feature map into a feature map of 20*20*512 size for regression and classification calculations.

[0085] In this embodiment, the head network adopts a decoupling head structure, and the decoupling head structure includes a classification branch and a regression branch.

[0086] In this embodiment, the classification loss function of the classification branch during the training process includes a binary cross entropy loss function and a Varifocal Loss loss function, and the calculation formula is:

[0087]

[0088] Among them, BCE represents the binary cross entropy loss function; N represents the number of samples; y i represents the true label; p i represents the probability predicted by the model; VFL represents the Varifocal Loss function; q i It represents the overlap between the predicted box and the true box; log represents the logarithmic function with the natural logarithm e as the base.

[0089] In this embodiment, the regression loss function of the regression branch during the training process includes a CIoU loss function and a Distribution Focal Loss loss function, and the calculation formula thereof is:

[0090]

[0091] Among them, CIoU represents the CIoU loss function, IoU represents the intersection-over-union ratio, ρ represents the distance between the center point of the predicted box and the real box, c represents the diagonal length of the minimum closed area, w represents the width of the predicted box and the real box, h represents the height of the predicted box and the real box, Area(A∪B) represents the intersection area of ​​the detection box A and the annotation box B, Area(A∩B) represents the union area of ​​the detection box A and the annotation box B, DFL represents the distribution focal loss function, P(yi) represents the label y is located at the real label y i The probability of S i Indicates that the predicted box is located at the true label y i The predicted probability at , arctan represents the inverse tangent function, log represents the logarithmic function with the natural logarithm e as the base, π represents pi, and N represents the number of samples.

[0092] In this embodiment, the dynamic monitoring described in step S3 is specifically as follows: a detection frame of size 640*640 is set, and each digital slice image of the remaining levels is detected from top to bottom and from left to right, and all slices whose overlap rate between the detection frame coordinates and the annotation frame coordinates reaches 60% are screened out, and slices containing target areas at different levels are obtained, and the size of the slices containing the target area is 640*640.

[0093] Example 2, based on Example 1, uses the Mean Average Precision (mAP) indicator to demonstrate the performance of the Yolov8 model. mAP is a commonly used evaluation indicator used to measure the performance of the target detection model. mAP comprehensively considers the detection accuracy of all categories and can fully reflect the detection capability of the model. Specifically, it includes the following formula:

[0094]

[0095] Among them, Precision represents accuracy, that is, the proportion of samples predicted to be positive that are actually positive; TP represents the number of true positives, that is, the number of positive samples correctly predicted by the model as positive samples; FP represents the number of false positives, that is, the number of negative samples incorrectly predicted by the model as positive samples.

[0096]

[0097] Among them, Recall represents the recall rate, that is, the proportion of samples that are actually positive that are predicted to be positive; FN represents the number of false negatives, that is, the number of positive samples that the model mistakenly predicts to be negative samples.

[0098] According to the calculated Precision and Recall, draw the Precision-Recall curve, and calculate the area under the Precision-Recall curve to get AP (Average Precision), which is calculated as follows:

[0099] AP=Σ n (R n -R n-1 ) n

[0100] Among them, R n Indicates the accuracy of the nth point in the Precision-Recall curve; P n Represents the recall rate of the nth point in the Precision-Recall curve; R n-1 Indicates the accuracy of the n-1th point in the Precision-Recall curve.

[0101] According to AP, the mean of AP is calculated for all categories (cancer areas and non-cancer areas) to obtain mAP, which is calculated as follows:

[0102]

[0103] Where L represents the number of categories and represents the AP of the lth category.

[0104] mAP50 and mAP50-90 are selected as the evaluation of the model performance during the training process. mAP50 is the average result of the average accuracy of the model on all categories when the IoU threshold is 50%. This indicator provides the average performance of the model under medium overlap. mAP50-90 is an interval average accuracy, which calculates the average accuracy of the model over the entire range of IoU from 50% to 90%. This indicator can show the model's ability to detect targets with different degrees of overlap and is usually used to more comprehensively evaluate model performance. After actual testing, it was found that Figure 3 , Figure 4 and Figure 5 As shown in the figure, the dynamic detection and region optimization methods can reach 0.91 and 0.85 in mAP50 and mAP50-90 respectively, which shows that the recognition effect of the model has excellent performance.

[0105] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A full-view digital slice recognition method combining dynamic detection and regional optimization, characterized in that: The method comprises the following steps: S1. Obtaining the original full-view data slices, performing hierarchical extraction and annotation on the original full-view data slices, and obtaining data slice images of different levels and annotation information of each level; S2. Select a low-level data slice image and annotation information that can be loaded into the Yolov8 model, input the Yolov8 model for preliminary training, and obtain a preliminarily trained Yolov8 model; S3. Use the initially trained Yolov8 model to dynamically monitor the data slice images of the remaining layers to obtain slices containing the target area; S4. Using the slices containing the target area to perform regional optimization training on the initially trained Yolov8 model, a full-view data slice recognition model is obtained; S5. Input the full-view data slice to be identified into the full-view data slice identification model for classification and identification, output the target area identification result, and complete the full-view digital slice identification.

2. The full-view digital slice recognition method combining dynamic detection and regional optimization according to claim 1, characterized in that: The step S1 specifically includes the following sub-steps: S11. Obtaining the original full-view data slices, performing hierarchical extraction on the large-size full-view data slices according to the ratio of the resolution, and obtaining data slice images of different levels; S12. Select a low-level data slice image that can be loaded into the Yolov8 model for annotation, and obtain annotation information of the low-level data slice image, wherein the annotation box in the annotation information is marked with the target area position; S13. According to the ratio of the resolution, the annotation frame of the annotation information of the low-level data slice image is enlarged or reduced to obtain the annotation information of each level.

3. The full-view digital slice recognition method combining dynamic detection and regional optimization according to claim 1, characterized in that: The Yolov8 model in step S2 includes a backbone network, a connection network and a head network connected in sequence; The backbone network is used to receive the data slice image and perform feature extraction, and output a feature map; The connection network is used to receive the feature map and perform feature fusion on the feature map to output a fused feature; The head network is used to receive the fused features, perform target detection and recognition, and output a target area recognition result.

4. The full-view digital slice recognition method combining dynamic detection and regional optimization according to claim 3 is characterized in that: The backbone network includes a first convolutional layer, a second convolutional layer, a first C2F layer, a third convolutional layer, a second C2F layer, a fourth convolutional layer, a third C2F layer, a fifth convolutional layer, a fourth C2F layer and an SPPF layer connected in sequence; The input end of the first convolutional layer is the input end of the entire backbone network, and the output end of the SPPF layer, the output end of the second C2F layer and the output end of the third C2F layer are all connected to the connection network; The output ends of the first convolutional layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer and the fifth convolutional layer are all provided with a regional optimization mechanism.

5. The full-view digital slice recognition method combining dynamic detection and regional optimization according to claim 4, characterized in that: The region optimization mechanism is specifically as follows: the feature maps output by each convolution layer are copied into feature maps I and feature maps J, where i represents the position number of the element on feature map I, and j represents the position number of the element on feature map J; the cosine similarity of all elements on feature map J is calculated using the element with position number i, and the sum of all cosine similarities is divided by the number of elements of the feature map output by the convolution layer to obtain a new feature map after region optimization; Repeat the above steps until all elements on feature map I have completed the similarity calculation of all elements on feature map J, and the calculation process ends; The calculation formula of the regional optimization mechanism is: Among them, y i Represents the new feature map after region optimization, x i represents the feature map output by each convolutional layer in the backbone network corresponding to the position number i of the element on the feature map I, x j represents the feature map output by each convolutional layer in the backbone network corresponding to the position number j of the element on the feature map J, f(·,·) represents the function for calculating the similarity, g(·) represents the linear change function, and M represents the number of elements in the feature map output by the convolutional layer.

6. The full-view digital slice recognition method combining dynamic detection and regional optimization according to claim 4, characterized in that: The connection network includes a first feature fusion module, a second feature fusion module, a third feature fusion module and a fourth feature fusion module connected in sequence; The input end of the first feature fusion module is also connected to the output end of the third C2F layer; The input end of the second feature fusion module is also connected to the output end of the second C2F layer.

7. The full-view digital slice recognition method combining dynamic detection and regional optimization according to claim 3 is characterized in that: The head network adopts a decoupling head structure, and the decoupling head structure includes a classification branch and a regression branch.

8. The full-view digital slice recognition method combining dynamic detection and regional optimization according to claim 7, characterized in that: The classification loss function of the classification branch during the training process includes a binary cross entropy loss function and a Varifocal Loss loss function, and the calculation formula is: Among them, BCE represents the binary cross entropy loss function; N represents the number of samples; y i represents the true label; p i represents the probability predicted by the model; VFL represents the Varifocal Loss function; q i It represents the overlap between the predicted box and the true box; log represents the logarithmic function with the natural logarithm e as the base.

9. The full-view digital slice recognition method combining dynamic detection and regional optimization according to claim 7, characterized in that: The regression loss function of the regression branch during the training process includes the CIoU loss function and the DistributionFocal Loss loss function, and the calculation formula is: Among them, CIoU represents the CIoU loss function, IoU represents the intersection-over-union ratio, ρ represents the distance between the center point of the predicted box and the real box, c represents the diagonal length of the minimum closed area, w represents the width of the predicted box and the real box, h represents the height of the predicted box and the real box, Area(A∪B) represents the intersection area of ​​the detection box A and the annotation box B, Area(A∩B) represents the union area of ​​the detection box A and the annotation box B, DFL represents the distribution focal loss function, P(yi) represents the label y is located at the real label y i The probability of S i Indicates that the predicted box is located at the true label y i The predicted probability at , arctan represents the inverse tangent function, log represents the logarithmic function with the natural logarithm e as the base, π represents pi, and N represents the number of samples.

10. The full-view digital slice recognition method combining dynamic detection and regional optimization according to claim 1, characterized in that: The dynamic monitoring described in step S3 is specifically as follows: a detection frame of a fixed size is set, each digital slice image of the remaining levels is detected from top to bottom and from left to right, and all slices whose proportion of the target area to the detection frame exceeds a threshold are screened out to obtain slices containing the target area.

Images