Gallbladder Polyp Classification in Ultrasound Images Using an Ensemble Convolutional Neural Network Model

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Accurate diagnosis of true gallbladder polyps is clinically challenging.
• Distinguishing small polyps (<20 mm) is crucial for patient management.
• Current diagnostic methods require improvement for small gallbladder lesions.

Purpose of the Study:

• To develop and evaluate a deep learning model for differentiating true gallbladder polyps.
• To improve diagnostic accuracy for small gallbladder polyps (<20 mm) using ultrasound.
• To assess the model's performance with and without clinical data.

Main Methods:

• Utilized an ensemble model combining three Convolutional Neural Network (CNN) models.
• Analyzed abdominal ultrasound images from 501 patients with confirmed gallbladder polyps.
• Employed 5-fold cross-validation and incorporated patient age and polyp size data.

Main Results:

• The ensemble model achieved an AUC of 0.8960 and 83.63% accuracy using only ultrasound images.
• Incorporating age and size improved performance to AUC 0.9082 and 87.61% accuracy with 88.35% specificity.
• The model demonstrated superior performance compared to individual CNNs, especially for polyps >10 mm (AUC 0.9131).

Conclusions:

• The proposed ensemble deep learning model accurately classifies small gallbladder polyps (<20 mm) on ultrasound.
• This AI tool offers high specificity, potentially reducing unnecessary cholecystectomies.
• The model shows promise for improving differential diagnosis in gallbladder polyp management.