GDG-YOLO：一种基于全局上下文增强的双分支轻量级网络用于喉部疾病检测

文献信息

中文摘要

目的： 喉部病变的早期准确检测对临床诊断和治疗至关重要。然而，传统喉镜评估具有主观性且易发生疏漏。尽管基于深度学习的方法已显示出前景，但其往往对微小病变敏感性差、易受光照变化和镜面反射影响，并依赖有限的标注数据。本研究旨在开发一种兼顾检测性能与计算复杂度的轻量级检测模型，用于喉部病变检测。
方法： 我们开发了GDG-YOLO，一种针对任务定制的轻量级检测框架，旨在应对喉镜成像中的三个常见挑战：微小病变的特征表征有限、异质光照下的特征不稳定，以及模糊或不规则病变边界相关的定位不确定性。该框架自适应地调整并整合已有架构组件。首先，将成熟的全局上下文模块嵌入C3k2结构以增强上下文特征表征。其次，设计任务特定的双路径颈部网络，将深层语义信息与高分辨率空间特征相结合以实现多尺度特征融合。第三，轻量级检测头结合分组卷积和基于分布焦点损失的回归，以在定位性能与计算复杂度之间取得平衡。此外，对训练集中的少数类应用AugMix数据增强以增加样本多样性并缓解类别不平衡。最后，在单中心临床喉镜图像数据集上对所提框架进行评估。
结论： GDG-YOLO取得了87.6%的mAP50，较基线YOLOv11n提升了5.6个百分点。消融研究和统计分析表明，完整的GDG-YOLO架构在所有评估指标上均显著优于YOLOv11n基线。此外，对比实验显示，在当前实验设置下，GDG-YOLO在精确率、F1分数、mAP50和mAP50:95上均高于所评估的代表性检测模型，包括YOLOv10n、RT-DETR和SSD。

英文摘要

OBJECTIVE: Early and accurate detection of laryngeal lesions is essential for clinical diagnosis and treatment. However, conventional laryngoscopic assessment is subjective and prone to oversight. Although deep learning-based methods have shown promise, they often suffer from poor sensitivity to small lesions, vulnerability to illumination variations and specular reflections, and dependence on limited annotated data. This study aims to develop a lightweight detection model that balances detection performance and computational complexity for laryngeal lesion detection.
METHODS: We developed GDG-YOLO, a task-tailored lightweight detection framework designed to address three prevalent challenges in laryngoscopic imaging: limited feature representation of small lesions, feature instability under heterogeneous illumination, and localization uncertainty associated with blurred or irregular lesion boundaries. This framework adaptively adjusts and integrates established architectural components. First, a mature Global Context Block is embedded into the C3k2 structure to enhance contextual feature representation. Second, a task-specific dual-path neck is designed to combine deep semantic information with high-resolution spatial features for multi-scale feature fusion. Third, a lightweight detection head incorporates grouped convolutions and distribution focal loss-based regression to strike a balance between localization performance and computational complexity. Furthermore, AugMix data augmentation is applied to minority classes in the training set to increase sample diversity and alleviate class imbalance. Finally, the proposed framework is evaluated on a single-center clinical laryngoscopic image dataset.
CONCLUSION: GDG-YOLO achieved an mAP50 of 87.6%, representing a 5.6-percentage-point improvement over the baseline YOLOv11n. Ablation studies and statistical analyses demonstrated that the complete GDG-YOLO architecture significantly outperformed the YOLOv11n baseline across all evaluated metrics. In addition, comparative experiments showed that GDG-YOLO achieved higher Precision, F1-score, mAP50, and mAP50:95 than the evaluated representative detection models, including YOLOv10n, RT-DETR, and SSD, under the current experimental setting.