Human and Deep Learning Predictions of Peripheral Lung Cancer Using a 1.3 mm Video Endoscopic Probe.

Area of Science:

• Pulmonology
• Medical Imaging
• Artificial Intelligence

Background:

• Small peripheral pulmonary nodules (PPNs) pose diagnostic challenges.
• Iriscope, a 1.3 mm video endoscopic probe, enables direct visualization of PPNs via radial-probe endobronchial ultrasound (r-EBUS) catheters.
• Assessing malignancy risk in PPNs is crucial for patient management.

Purpose of the Study:

• To evaluate the diagnostic performance of physicians with varying experience levels in interpreting Iriscope images of PPNs.
• To assess the capability of an artificial intelligence (AI) deep learning (DL) model in predicting malignancy of PPNs visualized with Iriscope.
• To compare the diagnostic accuracy of human interpreters and AI against the final diagnosis of PPNs.

Main Methods:

• Analysis of video-recorded Iriscope sequences from patients undergoing bronchoscopy for PPNs (<20 mm).
• Independent interpretation of images by senior and junior physicians, classifying lesions as tumoral or non-tumoral.
• Training and testing a DL model on Iriscope images to predict malignancy, comparing its performance to human interpreters.

Main Results:

• Iriscope enabled direct visualization of PPNs in all 61 included patients.
• Senior physicians achieved a balanced accuracy of 85.4%, outperforming junior physicians (66.7%) in identifying malignant nodules.
• The DL model achieved 71.5% balanced accuracy, outperforming junior physicians but not senior physicians.

Conclusions:

• Iriscope is a valuable tool for managing PPNs, particularly for experienced bronchoscopists.
• AI, specifically DL models applied to Iriscope images, has the potential to augment the diagnostic capabilities of less experienced physicians.
• Further research may explore integrating AI into peripheral endoscopy workflows to improve PPN diagnosis.