Development of deep learning-based narrow-band imaging endocytoscopic classification for predicting colorectal...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Endoscopic diagnosis of colorectal lesions benefits from data-driven methods.
• Endocytoscopy (EC), a high-magnification technique, has seen limited AI application, often using traditional machine learning.
• Existing AI models for EC predominantly use conventional methods like support vector machines.

Purpose of the Study:

• To develop and evaluate a novel computer-aided diagnosis (CAD) model for classifying colorectal lesions using narrow-band imaging endocytoscopy (EC-NBI).
• To leverage large-scale language model principles, specifically multi-stage pre-training and supervised deep clustering, for enhanced diagnostic performance.
• To compare the model's accuracy against state-of-the-art supervised methods and endoscopists' performance.

Main Methods:

• Development of a CAD model utilizing endocytoscopy narrow-band imaging (EC-NBI) data.
• Implementation of a multi-stage pre-training strategy inspired by large-scale language models.
• Integration of supervised deep clustering for improved lesion classification.
• Retrospective multi-center cohort validation and human-machine competition analysis.

Main Results:

• The developed EC-NBI CAD model demonstrated superior performance compared to state-of-the-art supervised methods in a multi-center retrospective cohort.
• The model surpassed the diagnostic accuracy of endoscopists in direct human-machine competitions.
• The CAD model significantly enhanced endoscopists' diagnostic performance when utilized as an assistive tool.

Conclusions:

• The EC-NBI CAD model substantially improves the accuracy and consistency of colorectal lesion diagnosis.
• This AI tool provides a strong foundation for future advancements in early cancer screening, particularly in differentiating superficial and deep submucosal invasive cancers.
• Further validation with expansive multi-center data is recommended to solidify clinical utility.