@article{QIMS155082,
author = {Linfang Li and Xuan Wen and Mengfei Li and Xinfa Wang and Xiangrong Feng and Xiangpeng Lv},
title = {Downsampling attention fusion network (DAFNet): a You Only Look Once network for lung nodule detection},
journal = {Quantitative Imaging in Medicine and Surgery},
volume = {16},
number = {7},
year = {2026},
keywords = {},
abstract = {Background: Lung nodule detection in computed tomography (CT) imaging is crucial for the early diagnosis and treatment of lung diseases. However, accurate detection remains challenging due to the varying sizes, shapes, and densities of nodules, as well as interference from surrounding tissues. This study aimed to develop a novel deep learning framework to improve the accuracy and efficiency of lung nodule detection.Methods: We developed the downsampling attention fusion network (DAFNet). Specifically, we incorporated a dual-branch downsampling module that employs a parallel learning strategy to effectively extract multiscale features while reducing computational complexity as compared to traditional convolutional modules. Additionally, we designed a global attention module (GAM), which cascades the channel and spatial attention mechanisms to enhance feature representation across different dimensions. To evaluate DAFNet, experiments were conducted on a lung CT image dataset (LCTD) and the public dataset Lung Nodule Analysis 2016 (LUNA16). LCTD comprises 2,172 CT images from 1,060 individuals. On the LUNA16, common object detection metrics such as mean average precision (mAP) and recall were adopted for evaluation, rather than the official specified free-response receiver operating characteristic (FROC) evaluation protocol.Results: On the augmented LCTD, DAFNet achieved 93.2% precision and 90.7% recall. Meanwhile, on LUNA16, DAFNet outperformed state-of-the-art methods in terms of mAP. Furthermore, DAFNet achieved GPU-accelerated near-real-time inference with a processing speed of 1.5 ms per image, and the model parameter size was only 2.5 M, enabling efficient and lightweight inference.Conclusions: Our study examined a novel method for detecting lung nodules, offering new insights and technical references for relevant research fields. The proposed modules can be flexibly and seamlessly integrated into various detection frameworks as plug-and-play components, effectively enhancing the flexibility and scalability of lung nodule detection. As our DAFNet does not incorporate false-positive reduction or malignancy grading in its nodule detection, the clinical translation and practical application of the model remain to be further validated and optimized in future work.},
issn = {2223-4306}, url = {https://qims.amegroups.org/article/view/155082}
}