Construction method of pulmonary nodule target detection system based on improved YOLOv8

Abstract

The invention provides a construction method of a pulmonary nodule target detection system based on improved YOLOv8, and the method comprises the following steps: S1, selecting a self-made pulmonary nodule data set, and dividing the self-made pulmonary nodule data set into a training set and a verification set; s2, training an original YOLOv8 network model by using the training set, and constructing a migration model; s3, constructing an improved YOLOv8 network model; s4, constructing a detection head of a head network of the improved YOLOv8 network model; s5, constructing to obtain a loss function of the improved YOLOv8 network model; s6, training the models constructed in the steps S3-S5 by using a training set; and S7, forming a pulmonary nodule target detection system based on the improved YOLOv8 by the migration model obtained in the step S2 and a detection model formed by the models obtained in the steps S3 to S5, and respectively outputting corresponding detection results. According to the pulmonary nodule target detection system based on the improved YOLOv8, the low-quality harmful gradient is reduced, so that the detection system focuses on an anchor frame, and the overall accuracy of the detection system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a method for constructing a lung nodule target detection system based on improved YOLOv8. Background Art

[0002] In recent years, with the development of artificial intelligence, many methods have been developed for automatic detection of pulmonary nodules, which can be classified into two main modes: traditional detection mode and deep learning-based mode. Previous detection methods usually use threshold-based algorithms, morphological operations, energy optimization methods and clustering methods to detect nodules from CT images. With the advancement of technology, some scholars have begun to combine deep learning image processing algorithms with CT technology to achieve more accurate diagnosis and detection of lung diseases. Among them, deep learning-based target detection algorithms have shown great potential in medical image analysis. The YOLO (You Only Look Once) series of algorithms are widely used in various target detection tasks due to their efficient real-time detection capabilities. However, the standard version of YOLOv8 still faces challenges in dealing with complex backgrounds and multi-scale lesions. The current detection model for pulmonary nodule disease has too many parameters and lacks an effective lightweight strategy. By introducing a custom lightweight shared detail feature enhancement detection head (Detect_LWSDDH), the model can be effectively lightweighted and the number of parameters can be reduced. However, these methods may not be able to fully explore the deep features of the image, and cannot solve the missed detection problem of YOLOv8 and the detection accuracy and efficiency problems in complex scenes. Summary of the invention

[0003] In order to solve the problems of insufficient multi-scale feature extraction, high model complexity, large computing resource requirements, limited detail enhancement capabilities, insufficient bounding box regression accuracy, and low target classification accuracy in the above-mentioned prior art detection systems, the present invention proposes a method for constructing a lung nodule target detection system based on improved YOLOv8, which aims to reduce the competitiveness of high-quality anchor boxes while reducing low-quality harmful gradients, so that the detection system can focus on the anchor boxes and improve the overall accuracy of the detection system.

[0004] The present invention is achieved through the following technical solution: comprising the following steps:

[0005] S1: Select a self-made lung nodule dataset and divide it into a training set and a validation set;

[0006] S2: Use the training set in the self-made lung nodule dataset in step S1 to train the original YOLOv8 network model and build a migration model;

[0007] S3: In the backbone network and neck network of the original YOLOv8 network model, dual convolution and efficient dual-layer aggregation network fusion module (CSP-DCAN) are introduced to obtain an improved YOLOv8 network model;

[0008] S4: Introduce a lightweight shared detail enhancement detection head (Detect_LWSDDH) to process the multi-scale feature map to obtain a detection head of the head network of the improved YOLOv8 network model;

[0009] S5: Replace the CIoU loss function in the original YOLOv8 network model with the weighted IoU (WIoUV3) loss function to obtain the loss function of the improved YOLOv8 network model;

[0010] S6: Use the training set of the self-made lung nodule dataset in step S1 to train the improved YOLOv8 network model constructed in S3-S5; when training the detection model, first import the pre-trained weights of the original YOLOv8 network model to initialize the parameters of the network; then fine-tune the weights of other layers by learning the training set of the self-made lung nodule dataset in step S1;

[0011] S7: The migration model obtained in step S2, the improved YOLOv8 network model obtained in steps S3 to S5, the detection head of the head network of the improved YOLOv8 network model, and the loss function of the improved YOLOv8 network model are used together to form a lung nodule target detection system based on improved YOLOv8, and the lung nodule target detection system based on improved YOLOv8 outputs corresponding detection results respectively.

[0012] Furthermore, the specific steps of the CSP-DCAN module in step S3 are as follows:

[0013] S31: The input feature map enters this module and starts feature extraction through channel adjustment. The input feature map is processed by the Conv1 layer, and the number of output channels is 2*c, where c is a hidden channel number, which is adjusted by the expansion factor e;

[0014] S32: The input feature map is split into two parts, each of which is further processed by a DBC submodule;

[0015] S33: All processed features are concatenated together through the Concat operation to form an aggregated feature map;

[0016] S34: The number of channels is further adjusted through a convolutional layer, and the resulting feature map is output.

[0017] Furthermore, the Detect_LWSDDH module in step S4 includes the following steps:

[0018] S41: Reduce the number of channels of the input feature map through a 1x1 convolution layer;

[0019] S42: shared convolutional layer (ACL) performs detail enhancement on multi-layer features;

[0020] S43: Integrating attention mechanism in shared convolutional layer (ACL);

[0021] S44: By optimizing the prediction accuracy of the bounding box.

[0022] Furthermore, the self-made lung nodule data set in step S1 contains a total of 4000 images, and the images are subjected to data set cleaning, data conversion and other data preprocessing and normalization processing to obtain normalized images; the normalized images are divided into a training set and a validation set in a ratio of 8:2.

[0023] Furthermore, the learning rate in step S6 is 0.01, and the number of iterations is 300.

[0024] Furthermore, the outlier degree β of the IoU (WIoUV3) loss function in step S5 is given by the formula In the formula, the molecule L * IoU is the IoU loss of the anchor box, and the denominator L IoU is the exponential running average of the dynamic IoU loss.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The present invention significantly enhances the detection system's ability to detect multi-scale pulmonary nodule lesions, especially subtle lesions, by introducing a custom dual convolution module and a CSP-DCAN module fused with a CSP efficient double-layer aggregation network.

[0027] 2. The present invention adopts the WIoUv3 loss function of the dynamic non-monotonic focusing mechanism, which specifically optimizes the bounding box prediction of non-standard shaped lesions and improves the detection accuracy of the detection system.

[0028] 3. The present invention uses a customized lightweight shared detail feature enhancement detection head (Detect_LWSDDH), which not only improves the detection accuracy of the detection system for key areas, but also maintains the lightweight and high efficiency of the detection system.

[0029] 4. The present invention adopts the CSP-DCAN module to replace the C2f module, and combines it with a custom dual convolution module to enable the detection system to better handle complex scenes, especially multi-scale and subtle lesions.

[0030] 5. The present invention uses the lightweight design of the Detect_LWSDDH detection head and CSP-DCAN module to reduce the computational complexity of the detection system while improving the detection capability of features of different scales.

[0031] 6. In the design of the loss function, the present invention adopts the weighted WIoUv3 loss function, which is specially optimized for lesions with non-standard shapes, thereby improving the accuracy of bounding box regression of the detection system.

[0032] 7. The present invention significantly improves the detection accuracy, efficiency and robustness of the detection system by introducing a custom dual convolution module, a CSP efficient two-layer aggregation network, a lightweight shared detail enhancement detection head and a WIoUv3 loss function. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of the construction method of the present invention.

[0034] Figure 2 It is a computing network architecture diagram of the YOLOv8 network of the present invention.

[0035] Figure 3 It is a CSP-DCAN network structure diagram of the YOLOv8 network of the present invention.

[0036] Figure 4 for Figure 3 DBC network structure diagram in.

[0037] Figure 5 for Figure 4 Diagram of the GDC network structure.

[0038] Figure 6 It is a structural diagram of the lightweight shared detail enhancement detection head (Detect-LWSDDH) of the YOLOv8 network of the present invention.

[0039] Figure 7 for Figure 6 ACL network structure diagram in .

[0040] Figure 8 It is a schematic diagram of the output results of the detection system of the present invention. DETAILED DESCRIPTION

[0041] Advantages and features of the present invention will be illustrated and explained by the following non-limiting description of preferred embodiments thereof, which are given by way of example only with reference to the accompanying drawings.

[0042] like Figures 1 to 7As shown, the present invention provides a method for constructing a lung nodule target detection system based on improved YOLOv8. The present invention first prepares a data set, then adopts a transfer learning method, uses a pre-trained YOLOv8 model as a basis, and constructs a transfer model by performing it on a homemade lung nodule data set. Then, a detection model is constructed based on YOLOv8. The detection model is a module (CSP-DCAN) that introduces a custom double convolution module and a CSP efficient double-layer aggregation network fusion into the backbone network and a module (CSP-DCAN) that introduces a custom double convolution module and a CSP efficient double-layer aggregation network fusion into the neck network. It can effectively extract features from different scales and capture different details and contextual information of the target, thereby improving detection accuracy. The head network introduces a custom lightweight shared detail enhancement detection head (Detect_LWSDDH) to improve the sharing of detail information between different detection layers. Finally, the transfer model and the detection model together constitute a deep learning system to detect the state of lung nodules. The specific steps of the present invention are as follows:

[0043] S1: Select a self-made lung nodule dataset and divide it into a training set and a validation set.

[0044] Specific data set preparation: A self-made lung nodule data set is used. The self-made lung nodule data set contains a total of 4,000 images. All images in the data set are subjected to data set cleaning, data conversion and other data preprocessing and normalization to obtain normalized images. The processed images are divided into a training set and a validation set in a ratio of 8:2. The training set is used for model training, and the validation set is used to verify the performance of the model.

[0045] S2: Use the training set in the homemade lung nodule dataset in step S1 to train the original YOLOv8 network model and build a migration model.

[0046] Migration model building:

[0047] Select the original YOLOv8 network model that has been pre-trained on a large-scale dataset. Use the pre-trained weights of the original YOLOv8 network model to initialize the model directly, thereby speeding up training and leveraging existing knowledge. Train the model on the self-made pulmonary nodule dataset, keep the model architecture unchanged, adapt to new categories, and retrain on the training set in the self-made pulmonary nodule dataset in step S1, so as to achieve effective migration and application of features.

[0048] S3: In the backbone network and neck network of the original YOLOv8 network model, double convolution and efficient double-layer aggregation network fusion module (CSP-DCAN) are introduced to obtain an improved YOLOv8 network model.

[0049] The specific steps of the CSP-DCAN module are as follows:

[0050] S31: The input feature map enters this module and starts feature extraction through channel adjustment. The input feature map is processed by the Conv1 layer, and the number of output channels is 2*c, where c is the number of hidden channels, which is adjusted by the expansion factor e.

[0051] S32: The input feature map is split into two parts, each of which is further processed by a DBC sub-module to extract efficient feature representation.

[0052] S33: All processed features are concatenated together through the Concat operation to form an aggregated feature map.

[0053] S34: Finally, a convolutional layer is used to further adjust the number of channels and output the resulting feature map.

[0054] like Figures 2 to 5 As shown, based on the original YOLOv8 network model, a dual convolution and efficient double-layer aggregation network fusion module (CSP-DCAN) is introduced in the backbone network of the original YOLOv8 network model, and a dual convolution and efficient double-layer aggregation network fusion module (CSP-DCAN) is introduced in the neck network. The module is fused by a dual convolution module and a CSP efficient double-layer aggregation network. The present invention introduces the CSP-DCAN module, which can effectively extract features from different scales through a combination of multiple GDC modules and convolutional layers. Multi-scale feature extraction is crucial for detecting targets of different sizes and scales. Multi-scale feature extraction can capture different details and contextual information of the target, thereby improving detection accuracy. The improved YOLOv8 network model is shown in FIG. Figure 2 As shown, Figure 2 Input in the figure represents the input layer, Conv represents the convolution layer, CSP-DCAN represents the CSP-DCAN module, SPPF represents the SPPF module in the YOLOv8 operation network, Concat represents the Concat function, and Upsample represents the Upsample function. The DBC module is the specific architecture of the core part of the entire backbone network improvement, in which the GDC module represents the addition of grouped convolution and point-by-point convolution. The DBC module uses two GDC modules in parallel, and the DBC module can fuse features from different convolution paths. This efficient feature fusion method not only enriches the feature representation, but also improves the model's ability to capture details in complex scenes, and effectively extracts multiple features from the input feature map. This parallel structure allows the model to capture feature information at different levels at the same time, improving the ability to detect targets of different scales and complexities.

[0055] like Figure 3As shown in the figure, the CSP-DCAN module is able to effectively extract features from different scales through the combination of multiple GDC modules and convolutional layers. Multi-scale feature extraction is crucial for detecting objects of different sizes and scales because it can capture different details and contextual information of the object, thereby improving detection accuracy. By using the grouped convolution and point convolution operations in the GDC module, the CSP-DCAN module is able to extract features more meticulously and effectively capture information at different scales. This enhanced feature expression capability helps improve the performance of the model, especially when dealing with complex and subtle objects.

[0056] S4: Introduce a lightweight shared detail enhancement detection head (Detect_LWSDDH) to process the multi-scale feature map to obtain a detection head of the head network of the improved YOLOv8 network model;

[0057] like Figure 6 As shown, in order to improve the detection ability of the YOLOv8 model under complex backgrounds, the present invention introduces a lightweight shared detail enhancement convolutional detection head (Detect_LWSDDH) at the head of the original YOLOv8 network model, which is represented by the Detect_LWSDDH module below. The Detect_LWSDDH module is implemented by the following specific steps:

[0058] S41: Reduce the number of channels of the input feature map through a 1x1 convolution layer, thereby reducing the computational complexity.

[0059] S42: The shared convolutional layer (ACL) performs detail enhancement on multi-layer features to maintain the detailed nature of the features.

[0060] S43: Integrate the attention mechanism in the shared convolutional layer (ACL) to enhance the focus on important regions.

[0061] S44: Keep the detection head lightweight and efficient by optimizing the prediction accuracy of the bounding box.

[0062] The Detect_LWSDDH module has the following advantages:

[0063] 1. Reduction of the number of channels: First, the number of channels of the input feature map is reduced through a 1x1 convolutional layer, thereby reducing the computational complexity.

[0064] 2. Detail enhancement and feature sharing: Through convolutional layers and shared convolutional layers (ACL), the details of multi-layer features are enhanced, and the feature maps are enhanced to ensure that subtle features can be effectively captured even in complex backgrounds. The network structure diagram of the shared convolutional layer (ACL) is shown in the figure. Figure 7In addition, the attention mechanism is integrated to enhance the model's attention to important areas, so that detailed information is shared between different detection layers, thereby improving the overall detection accuracy.

[0065] 3. Prediction output: Use a separate convolutional layer to perform bounding box prediction and category prediction on the processed feature map. Through this series of processing, the detection head of the head network of the improved YOLOv8 network model can more stably detect lung nodule lesions in complex backgrounds.

[0066] S5: Replace the CIoU loss function in the original YOLOv8 network model with the weighted IoU (WIoUV3) loss function to obtain the loss function of the improved YOLOv8 network model;

[0067] The present invention adopts the weighted IoU (WIoUv3) loss function to replace the traditional CIoU loss function to optimize the accuracy of bounding box regression.

[0068] The WIoUv3 loss function can more accurately handle non-standard shaped lung nodule lesions by weighting the center point error, aspect ratio error, and angle error. According to the importance of the errors in each dimension in the detection task, the WIoUv3 loss function can dynamically adjust the weighting coefficient. This adaptive mechanism enables the model to handle lesions of different shapes and sizes more robustly.

[0069] After using the WIoUv3 loss function, the model has significantly improved the accuracy of bounding box prediction, especially when dealing with lesions with non-standard shapes.

[0070] In the object detection task, outlier can be understood as the degree of deviation between the predicted bounding box and the true bounding box. If the IoU value of a predicted box is much lower than that of other predicted boxes, then the predicted box may be considered an abnormal box because it has a low degree of match with the true bounding box. The outlier β is defined to describe the quality of the anchor box, and the formula is as follows:

[0071]

[0072] In the formula, the molecule L * IoU is the IoU loss of the anchor box, and the denominator L IoU is the exponential running average of the dynamic IoU loss. The Wise-IoU v3 formula is as follows:

[0073] L WIOU =R WIOU L WIOU

[0074]

[0075] In the formula, W g With H g Indicates the width and height of the minimum bounding box. In order to prevent R WIoU Produces gradients that hinder convergence, W g With H g Separate it from the computation graph (superscript * indicates this operation), thus effectively eliminating the factors that hinder convergence.

[0076] S6: Use the training set of the self-made lung nodule dataset in step S1 to train the improved YOLOv8 network model constructed in S3-S5; when training the detection model, first import the pre-trained weights of the original YOLOv8 network model to initialize the parameters of the network; then fine-tune the weights of other layers by learning the training set of the self-made lung nodule dataset in step S1. Preferably, the learning rate is set to 0.01 and the number of iterations is set to 300.

[0077] S7: The migration model obtained in step S2, the improved YOLOv8 network model obtained in steps S3 to S5, the detection head of the head network of the improved YOLOv8 network model, and the loss function of the improved YOLOv8 network model are used together to form a lung nodule target detection system based on improved YOLOv8, and the lung nodule target detection system based on improved YOLOv8 outputs corresponding detection results respectively.

[0078] The improved YOLOv8 network model was trained using the training set of the self-made lung nodule dataset. The original YOLOv8 network model was compared with the improved YOLOv8 network model in a running experiment. The experimental process was as follows: the original YOLOv8 network model and the improved YOLOv8 network model were trained on the training set of the self-made lung nodule dataset for 300 iterations. The operating environment: CUDA 11.6; Python 3.8.18; GPU NVIDIA GeForce RTX 3090. Figure 8 , which is an output result diagram of the lung nodule target detection system based on improved YOLOv8 of the present invention.

[0079] Output the corresponding test results and evaluate:

[0080] In order to verify the effectiveness of each improved module in the detection system of the present invention, the original YOLOv8 network model is used as a benchmark, and precision, recall, average precision, and parameter quantity are used as evaluation indicators. Through different combinations of multiple improved modules, an ablation experiment is designed. The experimental results are shown in Table 1:

[0081] Table 1. Comparison of ablation experiment results

[0082]

[0083] According to Table 1, the improved YOLOv8 network model has a 9.1% improvement in accuracy P, and a 4.4% and 2.5% improvement in mAP50 and mAP50:95% respectively compared to the original YOLOv8 network model. As can be seen from the table, the number of parameters of the improved YOLOv8 network model has increased, and its comprehensive performance is better than the original YOLOv8 network model and other models in the ablation experiment. The CSP-DCAN module is added, which aggregates the features of multiple dual convolution modules and further processes these features to achieve efficient feature extraction. It can also effectively extract features from different scales. Multi-scale feature extraction is crucial for detecting targets of different sizes and scales because it can capture different details and contextual information of the target, thereby improving detection accuracy. Compared with the original YOLOv8 network model, the average accuracy mAP50 of the network after adding the CSP-DCAN module has increased by 2.3 percentage points. On the basis of adding the CSP-DCAN module to the original YOLOv8 network model, adding the WIOUv3 loss function increases mAP50 by 3.5 percentage points. On the basis of adding CSP-DCAN module and loss function WIOUv3 to the original YOLOv8 network model, the network after replacing the Detect module in the head with the customized lightweight shared detail feature enhancement detection head (Detect_LWSDDH) can effectively avoid extracting redundant features, which improves mAP50 by 4.4 percentage points. Improving the loss function to WIOUv3 can improve the regression accuracy of BBox.

[0084] In order to more fully evaluate the pulmonary nodule target detection system based on the improved YOLOv8 of the present invention, this detection system is compared with several other models under the same configuration and data set. The models are the single-stage model YOLOv3-tiny, YOLOv5s, YOLOv7-tiny, and the basic original YOLOv8 network model. The experimental results are shown in Table 2:

[0085] Table 2

[0086]

[0087] According to the analysis in Table 2, it can be obtained that the improved YOLOv8 lung nodule target detection system has superior performance in the lung nodule target detection task. Under the same GFLOPS, the average precision (map) increases by 4.4 points. When the number of parameters of YOLOv8n is smaller than that of YOLOv3-tiny and YOLOv5s, the recall rate and the number of parameters are better than these models. Compared with the original YOLOv8 network model, the lung nodule target detection system improved by the YOLOv8 of the present invention has improved average accuracy and average precision (map).

[0088] In summary, the effectiveness of the improved method proposed in the present invention is proved. The present invention forms a lung nodule target detection system based on improved YOLOv8 through the migration model and the detection model, and the output detection result keeps GFLOPS unchanged while improving the average precision (map).

[0089] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solutions formed by equivalent replacement or equivalent transformation shall fall within the protection scope required by the present invention.

Claims

1. A method for constructing a lung nodule target detection system based on improved YOLOv8, characterized in that: The following steps are included: S1: Select a self-made lung nodule dataset and divide it into a training set and a validation set; S2: Use the training set in the self-made lung nodule dataset in step S1 to train the original YOLOv8 network model and build a migration model; S3: In the backbone network and neck network of the original YOLOv8 network model, dual convolution and efficient dual-layer aggregation network fusion module (CSP-DCAN) are introduced to obtain an improved YOLOv8 network model; S4: Introduce a lightweight shared detail enhancement detection head (Detect_LWSDDH) to process the multi-scale feature map to obtain a detection head of the head network of the improved YOLOv8 network model; S5: Replace the CIoU loss function in the original YOLOv8 network model with the weighted IoU (WIoUV3) loss function to obtain the loss function of the improved YOLOv8 network model; S6: Use the training set of the self-made lung nodule data set in step S1 to train the improved YOLOv8 network model constructed in S3-S5; when training the detection model, first import the pre-trained weights of the original YOLOv8 network model to initialize the parameters of the network; Then, the weights of other layers are fine-tuned by learning the training set of the self-made lung nodule dataset in step S1; S7: The migration model obtained in step S2, the improved YOLOv8 network model obtained in steps S3 to S5, the detection head of the head network of the improved YOLOv8 network model, and the loss function of the improved YOLOv8 network model are used together to form a lung nodule target detection system based on improved YOLOv8, and the lung nodule target detection system based on improved YOLOv8 outputs corresponding detection results respectively.

2. The method for constructing a lung nodule target detection system based on improved YOLOv8 according to claim 1, characterized in that: The specific steps of the CSP-DCAN module in step S3 are as follows: S31: The input feature map enters this module and starts feature extraction through channel adjustment. The input feature map is processed by the Conv1 layer, and the number of output channels is 2*c, where c is a hidden channel number, which is adjusted by the expansion factor e; S32: The input feature map is split into two parts, each of which is further processed by a DBC submodule; S33: All processed features are concatenated together through the Concat operation to form an aggregated feature map; S34: The number of channels is further adjusted through a convolutional layer, and the resulting feature map is output.

3. The method for constructing a lung nodule target detection system based on improved YOLOv8 according to claim 2, characterized in that: The Detect_LWSDDH module in step S4 includes the following steps: S41: Reduce the number of channels of the input feature map through a 1x1 convolution layer; S42: shared convolutional layer (ACL) performs detail enhancement on multi-layer features; S43: Integrating attention mechanism in shared convolutional layer (ACL); S44: By optimizing the prediction accuracy of the bounding box.

4. The method for constructing a lung nodule target detection system based on improved YOLOv8 according to claim 3, characterized in that: The self-made lung nodule data set in step S1 contains a total of 4000 images, and the images are subjected to data set cleaning, data conversion and other data preprocessing and normalization processing to obtain normalized images; the normalized images are divided into a training set and a validation set in a ratio of 8:

2.

5. The method for constructing a lung nodule target detection system based on improved YOLOv8 according to claim 3, characterized in that: The learning rate in step S6 is 0.01, and the number of iterations is 300.

6. The method for constructing a lung nodule target detection system based on improved YOLOv8 according to claim 3, characterized in that: The outlier degree β of the IoU (WIoUV3) loss function in step S5 is given by the formula In the formula, the molecule L * IoU is the IoU loss of the anchor box, and the denominator L IoU is the exponential running average of the dynamic IoU loss.

Images