Ke Yang, Yun Xue, Zhe Du, Shuchang Xu, Tian Tang, Zhifeng Qu
Abstract
Reliable detection of engraved surface codes on surgical instruments is essential for end-to-end traceability, yet remains challenging in practice because metallic reflection, motion blur, scale variation and weak textures often hinder stable localization. Here we present LA-DBNet, a lightweight detection framework built on DBNet for this task. The model uses MobileNetV4 with LiteFPN to reduce complexity while preserving multi-scale feature representations.
Introduction
End-to-end traceability of surgical instruments is fundamental to infection control and patient safety in modern healthcare systems [1]. Across decontamination, cleaning, packaging, sterilization, distribution and intraoperative use, each instrument must be uniquely identified and its handling history recorded to support quality assurance, error prevention and adverse-event management [2]. To enable management at the individual instrument level and full-process traceability in both regulatory and clinical practice, permanent direct part marking is typically applied to the instrument itself, most commonly in the form of engraved alphanumeric strings or two-dimensional codes linked to traceability records [3].
Materials and methods
Principle of DBNet
DBNet is a segmentation-based scene text detection framework whose central innovation lies in the introduction of a differentiable binarization module, which incorporates the conventional thresholding step into network training and enables adaptive learning of binarization thresholds [16]. This design reduces the model’s sensitivity to manually specified fixed thresholds, thereby improving the stability of text region segmentation [17]. As shown in Fig 1, DBNet consists of three main components: feature extraction, feature fusion, and prediction decoding [18].
Results
Datasets
Because no publicly available dataset currently exists for this task, we constructed our surgical instrument code dataset to evaluate the effectiveness and robustness of the proposed method in this specialized setting. To further evaluate the performance of the proposed method beyond the specialized surgical instrument dataset, we also conducted comparative experiments on the public ICDAR2015 benchmark.
Discussion
The results indicate that surgical instrument code detection requires not only a lightweight model but also effective preservation of fine structural information. The target regions in this task are typically elongated, weak in texture, and frequently affected by reflection, blur, and low contrast. In this setting, lightweight replacement alone can reduce computational cost but may also weaken the representation of local details needed for accurate localization.
Conclusion
This study proposed LA-DBNet for surgical instrument code detection. By combining a lightweight architecture with task-oriented feature enhancement and consistency learning, the proposed method improved the detection of engraved code regions under challenging imaging conditions. On the in-house dataset, LA-DBNet achieved an F1 of 95.8%, outperforming the baseline DBNet while substantially reducing the number of parameters and increasing inference speed.
Citation: Yang K, Xue Y, Du Z, Xu S, Tang T, Qu Z (2026) A lightweight alignment-aware DBNet for surgical instrument code detection. PLoS One 21(8): e0355611. https://doi.org/10.1371/journal.pone.0355611
Editor: Athanasios G. Pantelis, Athens Medical Group, Psychiko Clinic, GREECE
Received: April 16, 2026; Accepted: July 23, 2026; Published: August 7, 2026
Copyright: © 2026 Yang et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The image dataset underlying the findings of this study is publicly available in the Zenodo repository at https://doi.org/10.5281/zenodo.21452216. The annotation files, trained model weights, source code, and instructions for accessing and using these resources are publicly available in the GitHub repository at https://github.com/wy49497/LADBNet.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
