Endoscopic Ultrasound of Pancreatic Tumors: A Dataset with Benchmarks for Convolutional Neural Network Classifiers

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Gastroenterology

Background:

• Pancreatic cancer diagnosis relies heavily on endoscopic ultrasound.
• Computer-assisted diagnosis (CAD) systems, particularly deep learning models, require substantial datasets for effective training.
• Limited research exists on CAD for endoscopic ultrasound in pancreatic cancer.

Purpose of the Study:

• To introduce a novel endoscopic ultrasound dataset for pancreatic tumor classification.
• To establish benchmark results for state-of-the-art models in classification and segmentation tasks.
• To evaluate the explainability of AI models using class activation maps.

Main Methods:

• A new dataset of 7825 endoscopic ultrasound images from 606 patients was created, including images with and without tumors, and segmentation masks for tumors.
• The dataset was manually divided into training and testing sets.
• State-of-the-art models like EfficientNetV2 for classification and U-Net for segmentation were benchmarked.

Main Results:

• EfficientNetV2 achieved 89.88% accuracy and 96.12% AUC for classification.
• U-Net obtained a 79.11% Dice score for segmentation.
• Explainable AI heatmaps showed limited overlap with segmentation masks (soft Dice scores 8.69-40.81%).

Conclusions:

• The novel dataset and promising classification/segmentation results highlight the potential of CAD for pancreatic tumor analysis.
• A gap exists between model classification performance and explainability, requiring further research.
• This work advances AI-driven diagnostic tools for pancreatic cancer detection via endoscopic ultrasound.