Abstract
INTRODUCTION: Lung cancer remains the leading cause of cancer-related mortality. Although bronchoscopy allows direct visualization and tissue sampling, detecting subtle lesions is still challenging owing to limited resolution, variable imaging conditions, and the complex structure of the airway. Most existing approaches treat lesion segmentation and cancer diagnosis as separate tasks, which can reduce diagnostic coherence and limit clinical applicability. METHOD: We propose a novel Multi-Adapter-based Segment Any Bronchoscope Model (MASA), an end-to-end framework with an encoder that fuses spatial, frequency, and positional information and a dual decoder that performs simultaneous lesion segmentation and lung cancer diagnosis. MASA was trained/evaluated on the public BM-BronchoLC dataset. RESULTS: On BM-BronchoLC, MASA improved lesion segmentation over the strongest baseline (ESFPNet), raising mean Dice coefficient (mDice) by +3.01% and mean Intersection-over-Union (mIoU) by +1.24%. For diagnosis, MASA increased Macro-F1 by +8.1 points and area under the precision-recall curve (AUPRC) by +14.1%. CONCLUSION: MASA provides a unified and interpretable pipeline for automated bronchoscopic image analysis, generating pixel-level lesion maps alongside case-level diagnostic predictions. The framework shows strong promise for improving early lung cancer detection and enhancing the efficiency of bronchoscopic workflows in clinical practice.