EUVIMEDEUVIMED entdecken
Crossref

MEDA-YOLO: a lightweight multi-scale edge fusion framework for pulmonary nodule detection

Kai Bian, Junfeng Ding, Haohao Wang, Jing Xu, Lipeng Gao

Measurement Science and Technology · 20.03.2026 · 37 (11) · 115402

DOI: 10.1088/1361-6501/ae4f09 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 23.03.2026. Verfasst von Kai Bian und 4 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 56 Literaturangaben. Gefördert von: Scientific Research Foundation for High-level Talents of Anhui University of Science and Technology, Medical Special Cultivation Project of Anhui University of Science and Technology, Natural Science Research Project of Anhui Educational Committee.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Early screening of pulmonary nodules is essential for preventing the progression of lung diseases and improving patient prognosis. Traditional detection methods struggle to address the features of pulmonary nodules, including irregular morphology, blurred boundaries, and high heterogeneity. In clinical practice, challenges such as difficulty in identifying peripheral pulmonary nodules, inconsistent sizing, blurred or lost feature information, and the need for mobile medical device deployment are encountered. This study proposes MEDA-YOLO, a lightweight deep learning framework for pulmonary nodule detection using the public LUNA16 dataset. To enhance boundary-blurred nodule detection, a multi-scale edge fusion module integrates atrous convolution with edge-aware mechanisms for effective multi-scale edge feature extraction. An efficient shared detection head addresses scale inconsistency using a convolutional kernel sharing and a learnable feature scaling strategy, enabling adaptive fusion across feature levels while reducing parameters. To mitigate downsampling information loss, a DysampleADown feature reconstruction module combines lightweight upsampling, depthwise separable convolutions, and channel attention to preserve critical details. A hybrid knowledge distillation strategy balances model size and accuracy by combining bridging cross-task protocol inconsistency and channel-wise knowledge distillation methods. A sigmoid-based equal partitioning approach resolves task protocol conflicts, while channel-wise probability normalization guides the student model toward foreground features. Experimental results demonstrate MEDA-YOLO achieves 88.12% precision, 90.55% mAP0.5, 63.29% mAP50/95, and 211 frames processed per second inference speed, with 2.5 MB model size and 2.7 GFLOPs computational cost. Compared to baseline, small nodule detection improves by 3.5% while computational complexity reduces by 57.1%, providing an efficient clinical deployment solution.

Förderung

  • Scientific Research Foundation for High-level Talents of Anhui University of Science and Technology
  • Medical Special Cultivation Project of Anhui University of Science and Technology
  • Natural Science Research Project of Anhui Educational Committee
  • Research Funds of Joint Research Center for Occupational Medicine and Health of IHM
  • Postgraduate Innovation Fund Project of Anhui University of science and technology
  • Innovation Project
  • Major Science and Technology Project in Anhui Province
  • of Anhui University of science and technology

Zitieren

APA

Bian, K., Ding, J., Wang, H., Xu, J., & Gao, L. (2026). MEDA-YOLO: a lightweight multi-scale edge fusion framework for pulmonary nodule detection. Measurement Science and Technology, 37(11), 115402. https://doi.org/10.1088/1361-6501/ae4f09

Vancouver

Bian K, Ding J, Wang H, Xu J, Gao L. MEDA-YOLO: a lightweight multi-scale edge fusion framework for pulmonary nodule detection. Measurement Science and Technology. 2026;37(11):115402. doi:10.1088/1361-6501/ae4f09

BibTeX
@article{Bian2026,
  author = {Bian, Kai and Ding, Junfeng and Wang, Haohao and Xu, Jing and Gao, Lipeng},
  title = {MEDA-YOLO: a lightweight multi-scale edge fusion framework for pulmonary nodule detection},
  journal = {Measurement Science and Technology},
  year = {2026},
  volume = {37},
  number = {11},
  pages = {115402},
  doi = {10.1088/1361-6501/ae4f09}
}
Kostenlos starten →