Automatic Pancreatic Cyst Lesion Segmentation on EUS Images Using a Deep-Learning Approach.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Accurate segmentation of pancreatic cyst lesions (PCL) is crucial for automated diagnosis using endoscopic ultrasonography (EUS).
• Existing segmentation methods require improvement for clinical application.

Purpose of the Study:

• To propose and evaluate a deep learning approach for automatic PCL segmentation on EUS images.
• To compare the performance of Attention U-Net against other U-Net variants for PCL segmentation.

Main Methods:

• Implementation of the Attention U-Net deep learning model for PCL segmentation.
• Comparative analysis with Basic U-Net, Residual U-Net, and U-Net++ models on internal and external datasets.
• Evaluation using Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) metrics.

Main Results:

• Attention U-Net achieved superior DSC and IoU scores on the internal test set compared to other models.
• On the external test set, Attention U-Net showed comparable performance to Basic U-Net, with no statistically significant difference.
• Cross-over study results indicated Attention U-Net's strong performance on internal data, with no significant differences observed against Residual U-Net or U-Net++ on external data.

Conclusions:

• Deep learning, specifically the Attention U-Net model, is effective for segmenting PCL in EUS images.
• This study represents the first implementation of deep learning for PCL segmentation on EUS data.
• The findings support the potential of automated deep learning approaches for PCL diagnosis.