Self-Supervised Spatio-Temporal Network for Classifying Lung Tumor in EBUS Videos

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Endobronchial ultrasound-guided transbronchial biopsy (EBUS-TBB) is crucial for diagnosing peripheral pulmonary lesions (PPLs).
• Current computer-aided diagnostic (CAD) systems for EBUS often use static 2D frames, neglecting temporal video dynamics.
• Temporal dynamics in EBUS videos may offer vital clues for distinguishing benign from malignant lesions.

Purpose of the Study:

• To develop an artificial intelligence (AI) model for analyzing full-length EBUS videos.
• To incorporate temporal modeling into AI for improved classification of PPLs.
• To enhance the diagnostic accuracy of EBUS-TBB using AI.

Main Methods:

• Retrospective collection of EBUS videos from 465 patients (Nov 2019 - Jan 2022).
• Utilized a dual-path 3D convolutional network (SlowFast) for spatiotemporal feature extraction.
• Integrated contrastive learning (SwAV) to improve model generalizability on clinical EBUS data.

Main Results:

• The SlowFast + SwAV_Frame model achieved an AUC of 0.857 on the validation set, outperforming pulmonologists (AUC not provided).
• On the test set, the AI model achieved an AUC of 0.823, with 76.92% accuracy and 84.85% sensitivity.
• The AI model demonstrated superior performance compared to conventional 2D architectures and human expert assessment.

Conclusions:

• This study presents the first CAD framework for real-time malignancy classification using full-length EBUS videos.
• The AI framework reduces reliance on manual image selection and enhances diagnostic efficiency.
• The AI model shows significant potential for clinical application due to its high accuracy in classifying PPLs.