Enhanced YOLO11-based pulmonary nodule intelligent identification method

Abstract

The invention discloses an enhanced YOLO11-based pulmonary nodule intelligent identification method, and belongs to the technical field of computer-aided diagnosis. On the basis of a YOLO11 backbone network, a multi-scale network MSN module is designed to replace an original C3K2 module, and through multi-scale fusion and weight distribution, the detection performance of pulmonary nodules is remarkably improved. Secondly, for pulmonary nodule segmentation, a spatial pyramid pooling (SPP) module is integrated in an improved U-Net model, the receptive field of the network is enhanced, and therefore the segmentation effect is improved; according to the overall framework, target detection of YOLO11 is combined with fine segmentation of U-Net, so that the problem of pulmonary nodule detection under a complex background is solved, and the detection precision and robustness are improved. Experimental results show that the kit has excellent performance on an LUNA16 data set, has high sensitivity and accuracy, can provide effective technical support for diagnosis of clinical early-stage lung cancer, and has wide application prospects and popularization value.

Description

Technical Field

[0001] The present invention relates to a novel deep learning-based lung nodule image detection method, which belongs to the technical field of computer-aided diagnosis, and in particular to a data processing method for intelligent lung nodule recognition based on an enhanced YOLO11 model. Background Art

[0002] Pulmonary nodules are small, round, abnormal lesions within lung tissue. Their early detection is crucial for the prevention and treatment of lung cancer. Traditionally, pulmonary nodule detection relies primarily on manual analysis of chest CT images by radiologists. This method is time-consuming and labor-intensive, and is susceptible to subjective factors, leading to high rates of missed and misdiagnoses. To improve the accuracy of pulmonary nodule detection, deep learning, particularly convolutional neural networks (CNNs), has been widely used in medical image analysis in recent years and has demonstrated significant potential in pulmonary nodule detection.

[0003] Most deep learning-based lung nodule detection methods rely on target detection models, which can automatically learn to extract features from images and automatically identify and locate nodules based on these features. Currently, the most commonly used deep learning models include Faster R-CNN, the YOLO (You Only Look Once) series, and RetinaNet. Compared with traditional manual feature extraction methods, these methods not only improve detection efficiency but also reduce reliance on physician experience. However, deep learning-based lung nodule detection methods still face many challenges and difficulties in practical applications: 1. Diversity of lung nodules and interference from complex background: Lung nodules have diverse morphologies, with large variations in size and density. They are easily confused with surrounding tissues such as pulmonary blood vessels, which increases the difficulty of detection and requires the model to have high generalization ability and robustness.

[0004] 2. Data imbalance and labeling difficulties: In lung nodule datasets, nodule samples are far fewer than normal samples, resulting in data imbalance and affecting model detection performance. Furthermore, high-quality labeling of medical images relies on expert participation, and the scarcity of data labeling further constrains the performance of deep learning models.

[0005] 3. The contradiction between high false positive rate and inference efficiency: Existing models generate a large number of false positives while ensuring high sensitivity, which increases the burden on doctors. In addition, the models are usually complex and have slow inference speed, making it difficult to achieve efficient and real-time lung nodule detection in clinical environments with limited resources. Summary of the Invention

[0006] This paper proposes an intelligent lung nodule recognition method based on enhanced YOLO11. This method addresses the challenges of lung nodule diversity and complex background interference, providing an efficient detection solution. Pulmonary nodules vary in shape, size, and density, and are easily affected by surrounding pulmonary blood vessels and other tissues, making accurate nodule detection in complex backgrounds extremely challenging.

[0007] The present invention adopts the following technical solutions: S1. Obtain an image of the lung nodule to be detected; S2. Use the enhanced YOLO11 model to perform target detection on the acquired lung nodule image to be detected, thereby identifying and extracting the lung nodule area image; S3. Crop the selected lung nodule image and use the improved U-Net network to segment its edges to obtain a refined lung nodule edge image.

[0008] S4. Restore the lung nodule image output by the improved U-Net network (S3) to its original size before input, and then paste the lung nodule image back to the corresponding position in the output image of the enhanced YOLO11 model (S2) to obtain the prediction result of the lung nodule.

[0009] Furthermore, the implementation process of the enhanced Yolo11 model specifically includes the following: S1. Design the MSN (Multi-Scale Network) module architecture and replace the C3K2 modules in the backbone network and feature fusion network of YOLO11 with the MSN module. This module integrates the output features of three different sizes of lung nodules (from large to small) through multi-scale fusion to improve detection performance. S2. The C2PSA module is used to weight the multi-scale fused lung nodule features. The fused features are first divided into two parts, and the channel attention mechanism is applied to each part to adjust the weight. Then, all feature blocks are reassembled to form the final multi-scale lung nodule fusion feature.

[0010] Furthermore, the improved U-net specifically includes: S1. A spatial pyramid pooling (SPP) module is added to the final downsampling step. The SPP module processes features using max pooling of three different sizes and stacks the pooled outputs in the channel dimension to form a fused lung nodule feature. Subsequently, a 1×1 convolution is used to restore the number of channels to the original feature number, further improving the model's ability to capture complex structures.

[0011] Compared with the existing technology, the present invention proposes an intelligent lung nodule recognition method based on enhanced YOLO11, providing an innovative and efficient solution to the diversity and complex background interference problems in lung nodule detection. By introducing the multi-scale network (MSN) module into YOLO11, the effective fusion and weight distribution of the multi-scale features of lung nodules are achieved, significantly improving the detection performance of the model in complex backgrounds. In addition, the present invention integrates the spatial pyramid pooling (SPP) module into the improved U-Net network, and enhances the ability to capture the edges and detailed features of lung nodules through multi-scale pooling and channel dimension fusion. The entire system combines the preliminary detection of YOLO11 with the refined segmentation of U-Net to achieve higher detection accuracy and edge segmentation effect, solving the problems of high false positive rate and blurred edges commonly found in traditional methods, and significantly improving the recognition effect and diagnostic reliability of lung nodules. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described below with reference to the accompanying drawings, in which: Figure 1 This is a working diagram of the MSN-YOLO11 structure of the enhanced YOLO11-based intelligent lung nodule recognition method of the present invention; Figure 2 This is a working diagram of the improved U-net structure of the pulmonary nodule intelligent recognition method based on enhanced YOLO11 of the present invention; Figure 3 This is the MSN structure working diagram of the enhanced YOLO11-based pulmonary nodule intelligent recognition method of the present invention; Figure 4 Schematic diagram of the C3K2 structure of the intelligent lung nodule recognition method based on enhanced YOLO11 of the present invention; Figure 5 Schematic diagram of the C2PSA structure of the intelligent lung nodule recognition method based on enhanced YOLO11 of the present invention; Figure 6 Schematic diagram of the PSABlock structure of the intelligent lung nodule recognition method based on enhanced YOLO11 of the present invention; Figure 7 This is a flow chart of the lung nodule detection system of the present invention's lung nodule intelligent recognition method based on enhanced YOLO11. DETAILED DESCRIPTION

[0013] The experimental verification results of the present invention will be described below with reference to the accompanying drawings in the specification to make its technical content clearer and easier to understand.

[0014] like Figures 1 to 7As shown, a lung nodule intelligent recognition method based on enhanced YOLO11 includes the following steps: S1. Pulmonary Nodule Detection: Using YOLO11 as the backbone, the original network was rationally improved at multiple levels by introducing the Multi-Scale Network (MSN) module, the Spatial Pyramid Pooling (SPP) module, and the C2PSA module. These improvements significantly enhanced the model's ability to detect complex lung backgrounds and diverse nodule features, making MSN-YOLO11 more advantageous in the task of pulmonary nodule detection. Specifically, multiple MSN modules were incorporated into the backbone network to improve feature extraction capabilities. The feature fusion network effectively integrated features at different scales, and the upsampling layer further fused multi-scale features, thereby enriching the perception of nodules of different sizes.

[0015] The original YOLO11 is more suitable for target detection tasks with multiple targets and complex scenarios. The C3K2 module is used in the backbone network and feature fusion network, which can effectively extract target features. However, it becomes redundant when facing lung CT images and consumes a lot of time to predict target features. Therefore, the MSN module is designed in this paper to replace the C3K2 module as the main structure of target detection. Compared with the original YOLO11, MSN-YOLO11 demonstrates higher accuracy and adaptability in lung nodule target detection, and can better meet clinical needs.

[0016] S2. Edge segmentation of lung nodules: The lung nodule image is cropped within the detection frame obtained by MSN-YOLO11, which minimizes the proportion of irrelevant background in the entire image, making the image mainly contain lung nodules and a small amount of surrounding background, thereby effectively solving the problem of imbalance between positive and negative samples in the segmentation process and providing more favorable conditions for subsequent segmentation tasks. In order to further improve the segmentation effect and enhance the network's receptive field, such as Figure 2 As shown in the figure, in the improved U-Net model, a spatial pyramid pooling (SPP) structure is introduced in the final layer of the downsampling path. This structure performs max pooling operations on feature maps at various scales, such as 3×3, 5×5, and 7×7, stacking all pooled features with the original features in the channel dimension. 3×3 convolutions are then used to interact with the channel features, ensuring that the number of channels and size of features meet the network structure requirements. This multi-scale pooling process enables the network to obtain a larger receptive field, significantly improving feature extraction and segmentation performance, ultimately forming the complete structural diagram shown in the figure. Example

[0017] The enhanced YOLO11 proposed in this paper performs well on the Luna16 pulmonary nodule public dataset, reaching the highest level in terms of sensitivity and parameter quantity.

[0018] Table 1 Comparison of scores of various pulmonary nodule detection systems on the Luna16 dataset

[0019] In the overall lung nodule detection and segmentation experiment, the combination of MSN-YOLO11 and SPP-Unet achieved the best results, and its accuracy ranked first among all networks.

[0020] Table 2 Comparison of lung nodule segmentation performance results

[0021] This paper proposes an intelligent lung nodule recognition method based on an enhanced YOLO11. By incorporating innovative modules such as the Multi-Scale Network (MSN), the Segmentation Attention (PSA), and the Spatial Pyramid Pooling (SPP), the model significantly improves lung nodule detection accuracy and segmentation performance. The MSN-YOLO11 constructed in this paper not only possesses stronger feature extraction capabilities and higher computational efficiency, but also effectively processes complex lung CT images, minimizing the impact of positive and negative sample imbalance on model performance. This provides accurate and reliable technical support for the clinical diagnosis of early lung cancer. The overall method is simple and effective, suitable for the field of medical image analysis, and has broad application prospects and promotional value.

[0022] While embodiments of the invention have been shown and described, it will be readily apparent to those skilled in the art that It is understood that many changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention. The scope of the invention is defined by the following claims and their equivalents.

Claims

1. A lung nodule intelligent recognition method based on enhanced YOLO11, characterized in that: The following steps are involved: S1. Obtaining a lung CT image to be detected as a lung nodule image to be detected; S2. Use the enhanced YOLO11 model to perform target detection on the acquired lung nodule image to be detected, thereby identifying and extracting the lung nodule area image; the enhanced YOLO11 model includes a backbone network and a feature fusion network; S3. Cropping the lung nodule image selected in step S2 and performing edge segmentation on the image using an improved U-Net network, thereby obtaining a refined lung nodule edge image; the improved U-Net network includes an encoder and a decoder structure; S4. Restore the lung nodule image output by the improved U-Net network to its original size before input, and then paste the lung nodule image back to the corresponding position in the output image of the enhanced YOLO11 model to obtain the final lung nodule prediction result.

2. The method for intelligently identifying pulmonary nodules based on enhanced YOLO11 according to claim 1, wherein: The implementation process of the enhanced YOLO11 model specifically includes: S1. Design the MSN module architecture and replace the C3K2 modules of the backbone network and feature fusion network in YOLO11 with the MSN module. The MSN module performs multi-scale fusion integration on the output features of large, medium, and small pulmonary nodules to achieve effective detection of pulmonary nodules of different sizes. S2. Use the C2PSA module to assign weights to the multi-scale fused lung nodule features, where: a) First, the fused features are divided into two independent parts; b) Apply channel attention mechanism to each part to adjust the weight; c) All weight-adjusted feature blocks are reassembled to form the final multi-scale lung nodule fusion feature.

3. The method for intelligent recognition of pulmonary nodules based on enhanced YOLO11 according to claim 1, characterized in that: The improved U-Net network specifically includes: S1. Add a spatial pyramid pooling SPP module to the last layer of the downsampling path. The specific processing steps of the SPP module are as follows: a) Multi-scale feature processing is performed through maximum pooling of three different sizes: 3×3, 5×5, and 7×7; b) The output results of the pooling at each size are stacked in the channel dimension to form the fused lung nodule features; c) Use 1×1 convolution operation to restore the number of channels to the original number of feature channels; S2. Through the processing of the above-mentioned SPP module, the receptive field range of feature extraction is increased and the ability to capture complex structures is improved.

4. The method for intelligently identifying pulmonary nodules based on enhanced YOLO11 according to claim 2, wherein: The C2PSA module specifically includes: S1. First, the input feature map is grouped into channels, and the features are divided into two groups for parallel processing; S2. Apply channel attention mechanism to each set of features, including: a) Global average pooling operation to obtain global information at the channel level; b) Generate channel weights through nonlinear transformation; c) Multiply the weights by the original features to achieve adaptive adjustment of feature importance; S3. Concatenate and fuse the two sets of adjusted features and output the enhanced feature map.

5. The method for intelligently identifying pulmonary nodules based on enhanced YOLO11 according to claim 2, wherein: The specific structure of the MSN module includes: S1, the input features first undergo 1×1 convolution to perform channel dimensionality reduction; S2. Divide the reduced-dimensional features into multiple branches for parallel processing: a) The first branch uses 3×3 depth-wise separable convolution; b) The second branch uses 5×5 depth-wise separable convolution; c) The third branch uses 7×7 depth-wise separable convolution; S3, concatenate and fuse the features processed by each branch, and adjust the number of channels through 1×1 convolution; S4. Use residual connection to add the input features and the processed features to obtain the final output.

6. The method for intelligent recognition of pulmonary nodules based on enhanced YOLO11 according to claim 1, characterized in that: The method further comprises the following steps: S1. Preprocess the input CT image, including: a) Image normalization processing; b) Data augmentation processing, including random flipping, rotation, and scaling operations; S2. Post-process the model prediction results, including: a) Non-maximum suppression (NMS) processing to eliminate duplicate detection frames; b) Confidence threshold screening to filter out low-confidence detection results.

7. The method for intelligent recognition of pulmonary nodules based on enhanced YOLO11 according to claim 1, characterized in that: In S4, the process of restoring the segmentation result to the original size includes: S1, record the position coordinates of the cropped image in the original image; S2, adjusting the segmented result image to the original size by bilinear interpolation method; S3. Paste the adjusted segmentation result back to the original position accurately according to the recorded position coordinates.

Images