Page 84 - 2025S

UEC Int’l Mini-Conference No.54 77

A Novel Lightweight Pipeline for Real-Time Lesion Detection in

Bronchoscopic Imaging

2
3
Tawhid Ahmed Komol ∗1 , Norihiro Koizumi , Yu Nishiyama , and Peiji Chen 4
1 UEC Exchange Study Program (JUSST Program)
2,4 Department of Mechanical and Intelligent Systems Engineering
3 Department of Computer and Network Engineering
The University of Electro-Communications, Tokyo, Japan

Keywords: Bronchoscopy, Real-Time Detection, Lesion Localization, Deep Learning, Medical Image
Analysis, RT-DETR

Abstract
Bronchoscopy is a vital diagnostic tool for identifying lesions within the respiratory tract but most
computer-aided detection systems are based on static image classification models due to the absence of
lightweight, optimized architectures. In this study, we present a lightweight variant of the Real-Time
Detection Transformer (RT-DETR) optimized for bronchoscopic lesion detection. Using the BI2K
dataset, which contains 1550 annotated bronchoscopic images (600 benign and 950 malignant). We
evaluate and compare the performance of four real-time object detection models: YOLOv8-S, YOLO-
NAS, standard RT-DETR, and our proposed RT-DETR Lite. RT-DETR Lite achieves competitive
detection mean average precision of 0.87 while significantly reducing computational complexity to 32
GFLOPs and achieving 92 FPS on an RTX 3060 GPU. Architectural optimizations include substitut-
ing the ResNet50 backbone with MobileNetV3-Small, reducing encoder and decoder layers from 6 to
3, lowering the hidden dimension to 192, and decreasing the input resolution to 416×416. This bal-
ance of speed and accuracy demonstrates the suitability of RT-DETR Lite for real-time clinical use in
bronchoscopy.

∗
The author is supported by JASSO Scholarship.