1 Zibo Polytechnic University, Zibo, China
2 School of Computing and Digital Technology, University Malaysia of Computer Science & Engineering, Kuala Lumpur, Malaysia
Email: p09240011@student.unimy.edu.my and Salwani@unimy.edu.my
* Corresponding Author: Salwani@unimy.edu.my
Micro-foreign objects in infusion bags, such as tiny particles or impurities, present substantial risks to patient safety, potentially leading to infections, blockages, or other complications during intravenous therapy. Traditional inspection methods are often manual and inefficient, prompting the need for automated solutions. This paper proposes an optimized deep learning network based on YOLOv8, enhanced with a lightweight multi-scale feature fusion module incorporating attention mechanisms to improve detection accuracy for small-scale contaminants in limited datasets. The model addresses challenges like scale variance and computational overhead, making it suitable for real-time medical quality control. Evaluations on a dataset comprising 32 images with 80 annotations reveal significant performance gains: mean Average Precision (mAP) improves from 0.748 to 0.852, precision from 0.802 to 0.881, recall from 0.715 to 0.812, and F1-score from 0.757 to 0.845 over the baseline. Ablation studies confirm the fusion module's contribution, while comparisons with traditional models like YOLOv5 underscore the superiority. Visualizations, including loss curves, precision-recall curves, histograms, confusion matrices, and Grad-CAM, provide interpretability insights. Dataset statistics highlight an average image size of 720 pixels, emphasizing small object detection difficulties. This work advances AI-driven healthcare monitoring, facilitating integration into manufacturing lines or hospital protocols for enhanced safety. Future enhancements could involve multi-modal fusion for broader pollutant types.
Keywords: Multi-scale feature fusion, Deep learning, Micro-foreign object detection, Infusion bag monitoring, Small dataset optimization
How to cite this article: Zhang C, Daud SM. Multi-Scale Feature Fusion Optimization: A Deep Learning Network Based on YOLOv8 for Micro-Foreign Object Detection in Infusion Bags. Int J Drug Deliv Technol. 2026;16(5): 177-183. DOI: 10.25258/ijddt.16.5.20
Source of support: Nil.
Conflict of interest: None