.png){kind=icon}
.png){kind=icon}
Return
LN-YOLO: Lung Nodule Detection Method Based on Improved YOLOv8
.png){kind=icon}
PRE
AI
X
W
X
DOI：10.3788/LOP252081.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Objective Lung cancer remains one of the leading causes of cancer-related mortality globally, making the early identification of lung nodules from computed tomography (CT) scans critical for improving patient survival. However, achieving precise and automated detection faces significant challenges: 1) the fine-grained features of tiny nodules are susceptible to loss during network downsampling; 2) insufficient fusion of multi-scale features leads to inadequate utilization of contextual information; and 3) the high similarity between nodules and anatomical background structures, such as blood vessels, often causes misdetections and missed detections. Although deep learning methods based on convolutional neural networks (CNNs) and two-stage detectors have made progress, they often struggle to strike an optimal balance between detection accuracy, computational complexity, and robustness. This study proposes LN-YOLO, an enhanced detection algorithm based on YOLOv8, designed to achieve high-precision, robust, and computationally efficient lung nodule detection suitable for potential clinical deployment. Methods Building upon the YOLOv8n baseline model, we systematically optimized the downsampling process, feature fusion network, detection head, and loss function. First, to mitigate the loss of subtle features from tiny nodules during spatial dimensionality reduction, we replaced the standard downsampling layers in the backbone network with a robust feature downsampling (RFD) module, designed to maximally preserve fine-grained spatial information. Second, to address inadequate multi-scale feature fusion, we designed a novel feature aggregation diffusion network (FADN). Its core component, the lung nodule feature aggregation module (LNFAM), facilitates iterative cross-level feature aggregation and multi-path information diffusion, constructing feature representations with richer contextual information for nodules of varying sizes. Furthermore, to resolve the feature conflict between classification and regression tasks in conventional shared detection heads, we constructed a decoupled head named SEAM-head, which integrates a separation-enhancement attention module (SEAM) into independent branches to provide task-specific feature enhancement. Finally, to overcome the vanishing gradient problem encountered by traditional IoU losses when processing small, isolated nodules, we adopted a hybrid bounding box regression loss, ShapeIoU-NWD, which combines shape-aware Shape-IoU with the normalized Wasserstein distance (NWD). The proposed LN-YOLO model and all comparative baselines were trained and comprehensively evaluated on the public LUNA16 benchmark dataset. Results and Discussions Ablation studies confirmed the independent contributions of each component. The model achieved a mean average precision at an IoU threshold of 0.5 (P-mAP@0.5) of 95.2%, a 4.1 percentage point improvement over the YOLOv8n baseline. Additionally, and reached and , representing increases of 5.9 and 1.7 percentage points, respectively. Notably, these significant performance gains were achieved while the model's parameter count was reduced to 2.95 million. In comparative experiments, LN-YOLO demonstrated the best overall performance, outperforming another improved YOLOv8 model by 1.5 percentage points in P-mAP@0.5. Visualization of the detection results showed high concordance between the model's predictions and ground-truth labels, further verifying its excellent detection accuracy and localization capabilities. Conclusions This study successfully proposes LN-YOLO, an improved YOLOv8-based framework for high-performance lung nodule detection in CT images. By systematically integrating the RFD module, the FADN architecture, the SEAM-head decoupled head, and the ShapeIoU-NWD loss function, the proposed method effectively addresses the critical challenges of feature loss in small nodules, inadequate multi-scale context fusion, and low localization accuracy under complex anatomical backgrounds. Extensive experimental results demonstrate that LN-YOLO not only achieves leading detection performance on the authoritative LUNA16 benchmark but also maintains a compact model size. This combination of high accuracy, strong robustness, and lightweight architecture positions LN-YOLO as a practical and efficient solution for intelligent computer-aided diagnosis (CAD) systems for lung cancer screening. Future work will focus on incorporating 3D spatial information, implementing hard-sample mining, and integrating the model into a complete end-to-end diagnostic workflow.
Keywords:
lung nodule detection
deep learning
YOLOv8
attention mechanism
multi-scale feature extraction