Deep Learning in Endoscopic Ultrasound: A Breakthrough in Detecting Distal Cholangiocarcinoma

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Cholangiocarcinoma (CCA) is a lethal bile duct cancer often diagnosed late.
• Distinguishing benign from malignant biliary tumors is diagnostically challenging.
• Advanced imaging techniques are crucial for early and accurate diagnosis.

Purpose of the Study:

• To enhance diagnostic accuracy for distal cholangiocarcinoma (dCCA) using endoscopic ultrasound (EUS).
• To develop and evaluate advanced convolutional neural networks (CNNs) for classifying and segmenting EUS images.
• To specifically target dCCA, pancreas, and bile duct structures within EUS imagery.

Main Methods:

• A retrospective study utilizing EUS images from dCCA patients.
• Development of a custom CNN for image classification, trained on 156 images and augmented to 1248.
• Implementation of DeepLabv3+ with ResNet50 for organ and tumor segmentation, using Tversky loss.

Main Results:

• The classification CNN achieved 97.82% accuracy, 100% precision and specificity, and 94.44% sensitivity.
• Segmentation models showed 84% (pancreas) and 90% (bile duct) accuracy with strong Intersection over Union (IoU) scores.
• The AI application outperformed the UNet model in generalization and boundary delineation.

Conclusions:

• AI holds significant potential for improving diagnostic accuracy and efficiency in EUS for dCCA.
• The developed MATLAB application serves as a valuable tool for medical professionals.
• This AI-driven approach can facilitate informed decision-making and improve patient outcomes for cholangiocarcinoma.