Multi-adapter SAM-inspired bronchoscopic image segmentation for lung cancer diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Lung cancer is a leading cause of cancer mortality.
• Bronchoscopy aids lung cancer detection but struggles with subtle lesions due to imaging limitations.
• Current methods often separate lesion segmentation and diagnosis, hindering clinical application.

Purpose of the Study:

• To develop a unified framework for simultaneous bronchoscopic lesion segmentation and lung cancer diagnosis.
• To improve the accuracy and efficiency of automated analysis of bronchoscopic images.

Main Methods:

• Introduced the Multi-Adapter-based Segment Any Bronchoscope Model (MASA), an end-to-end framework.
• MASA utilizes an encoder fusing spatial, frequency, and positional information.
• A dual decoder performs simultaneous lesion segmentation and lung cancer diagnosis.

Main Results:

• MASA enhanced lesion segmentation, improving mean Dice coefficient (mDice) by +3.01% and mean Intersection-over-Union (mIoU) by +1.24% on the BM-BronchoLC dataset.
• For diagnosis, MASA increased Macro-F1 by +8.1 points and area under the precision-recall curve (AUPRC) by +14.1%.
• Outperformed the strongest baseline (ESFPNet) in both segmentation and diagnosis tasks.

Conclusions:

• MASA offers a unified, interpretable pipeline for automated bronchoscopic image analysis.
• The model generates pixel-level lesion maps and case-level diagnostic predictions.
• MASA shows potential for improving early lung cancer detection and clinical bronchoscopy workflows.