Page 85 - 2025S

78 UEC Int’l Mini-Conference No.54

A Novel Lightweight Pipeline for Real-Time Lesion Detection in
Bronchoscopic Imaging
Tawhid Ahmed Komol*, Norihiro Koizumi, Yu Nishiyama, Chen Peiji
UEC Exchange Study Program (JUSST Program)*
Department of Mechanical and Intelligent Systems Engineering
The University of Electro-Communications, Tokyo, Japan
Introduction RT-DETR Model Optimization
• Bronchoscopy is a vital diagnostic tool for identifying To achieve real-time performance & reduce GFLOPs to customized
lesions within the respiratory tract but most computer- the original RT-DETR as follows:
aided detection systems are based on static image Table 2: Model Optimization for Lightweight RT-DETR
classification models due to the absence of lightweight, Component Original Modified for Lightweight
optimized architectures [1][2]. Backbone: ResNet-50 MobileNetV3-Small
Encoder Layers: 6 3
• This research explores the implementation of Real-Time Decoder Layers: 6 3
Detection Transformer based with MobileNetV3-S Hidden Dim: 256 192
backbone detection pipelines that not only classify but also Object Queries: 300 100
localize lesions within bronchoscopic imaging. Resolution: 640×640 416×416
Background Study Results

Comparative analysis between existing medical imaging and Our best model RT-DETR result based on the evaluation metrics
our proposed work in bronchoscopic lesion detection: and showing the real time detection of malignant and benign
with confidence rate and also with the confusion matrix below:
Table 1: Comparative analysis between existing and proposed work
Table 3: Benchmarking Lightweight Model
Study Year Methodology Model mAP FPS (GPU)
Static classification Model mAP Precision F1-Score G-Flops FPS(@3060)
[2] X. Zhang et 2023 using handcrafted and ResNet50 0.84 Static
al. (Offline) YOLOv8-S 0.83 0.85 0.81 67 86
CNN features
Real-time detection in 52 FPS (RTX YOLO-NAS 0.85 0.87 0.84 40 90
[3] Y. Wang et al. 2022 YOLOv4-Tiny 0.78
medical video 2080Ti)
RT-DETR 0.87 0.87 0.86 75 60
Real-time airway
[4] K. Kobayashi 58 FPS (RTX RT-DETR Lite 0.87 0.88 0.87 32 92
2022 lesion detection in YOLOv5s 0.82
et al. 3060)
bronchoscopy
Our Proposed Real-time detection Optimized
92 FPS (RTX
Work (This 2025 and localization Lightweight 0.87
3060)
Study) pipeline RT-DETR
Dataset
The research utilized the BI2K Bronchoscopic Dataset, which
includes 1550 annotated lesion images sourced from
bronchoscopic procedures. For 2 categories: benign (600), Detected Malignant Detected Benign
malignant (950)
Fig 3: Results of Real Time Detection with RT-DETR
460 40
Normal Benign Malignant
Fig 1: Bronchoscopic Categories Normal, Benign & Malignant
Methodology 30 920

The clear outline of the proposed real-time detection pipeline is
presented here, from data preprocessing and model selection
to training and inference.
Fig 4: Confusion Matrix of RT-DETR
Future Work

• Extend real-time detection to continuous bronchoscopic
video streams.
• Develop a GUI diagnostic assistant tool for clinicians.
• Evaluate generalizability across multi-center datasets.
Real-Time Inference
References
with RT-DETR
1. Bochkovskiy, A., Wang, C. Y., & Liao, H. Y. M. (2020). YOLOv4: Optimal Speed and Accuracy of Object
Detection. arXiv preprint arXiv:2004.10934.
2. X. Zhang, Y. Zhao, Y. Liu, and H. Chen, "Classification of bronchoscopic images for pulmonary lesion
diagnosis using DenseNet and transfer learning," Computers in Biology and Medicine, vol. 154, p. 106315,
2023.
3. Y. Wang, J. Liu, M. Chen, and T. Lin, "Lightweight Real-Time Detection of Polyps and Airway Abnormalities
Using YOLOv4-Tiny in Medical Videos," IEEE Access, vol. 10, pp. 42261–42272, 2022.
Fig 2: Methodology flowchart of the proposed work 4. K. Kobayashi, S. Nakamura, and H. Matsuda, "Real-Time Detection of Abnormal Findings in Bronchoscopic
Video Using YOLOv5s," Sensors, vol. 22, no. 3, p. 891, 2022.