Development of convolutional neural network models that recognize normal anatomic structures during real-time...

Area of Science:

• Medical imaging
• Artificial intelligence in medicine
• Gastroenterology

Background:

• Endoscopic ultrasound (EUS) is a complex skill requiring extensive training.
• Limited global training facilities pose a challenge for EUS competency development.
• Convolutional neural network (CNN) models offer potential for automated object detection in medical imaging.

Purpose of the Study:

• To develop and evaluate EUS-based CNN models for recognizing normal anatomical structures.
• To assess the performance of CNN models in real-time linear- and radial-array EUS evaluations.
• To provide a tool for enhancing EUS training and skill acquisition.

Main Methods:

• Two CNN models (CNNv1 and CNNv2) were developed using recorded linear- and radial-array EUS videos.
• Expert endosonographers identified and labeled 20 normal anatomical structures for model training and validation.
• CNN model performance was compared against expert endosonographers using mean average precision (mAP) and total loss metrics.

Main Results:

• CNNv2 models demonstrated improved performance, with linear-array achieving 88.7% mAP (0.06 loss) and radial-array achieving 83.5% mAP (0.07 loss).
• CNNv2 models accurately detected all studied normal anatomical structures with over 98% agreement during clinical validation.
• Initial CNNv1 models showed lower performance (75.65% and 71.36% mAP).

Conclusions:

• The developed CNN models accurately recognize normal anatomical structures in both prerecorded and real-time EUS.
• These AI models hold promise for improving EUS training by providing real-time feedback.
• Further prospective trials are necessary to evaluate the impact on EUS trainee learning curves.