Algorithm combining virtual chromoendoscopy features for colorectal polyp classification.

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Colonoscopy is crucial for colorectal cancer prevention, but accurate optical diagnosis of colorectal polyps (CRPs) remains challenging.
• Variability in endoscopist accuracy highlights the need for advanced diagnostic tools.
• Computer-aided diagnosis (CAD) systems offer potential to enhance CRP characterization.

Purpose of the Study:

• To evaluate a novel CAD algorithm for classifying CRP malignancy.
• To leverage complementary information from three imaging modalities: high-definition white light (HDWL), Blue light imaging (BLI), and linked color imaging (LCI).

Main Methods:

• A CAD algorithm was developed using HDWL, BLI, and LCI still images from prospective colonoscopy exams.
• A triplet network architecture was employed for algorithm training, with histopathology serving as the gold standard for polyp classification.
• The training dataset underwent validation using threefold cross-validation.

Main Results:

• The study analyzed 609 images from 203 CRPs; 174 were premalignant and 29 benign.
• The combined algorithm achieved 90.6% accuracy, 89.7% sensitivity, and 96.6% specificity for malignancy classification.
• The CAD system demonstrated a rapid classification time of approximately 90 milliseconds per image.

Conclusions:

• The developed CAD algorithm accurately differentiates benign from premalignant colorectal polyps using non-magnified endoscopic images.
• This represents the first algorithm to integrate HDWL, BLI, and LCI modalities with a triplet network approach for polyp classification.