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

Pancreatic cancer is diagnosed mainly during endoscopic ultrasound, but limited few works explore computer-assisted diagnosis (CAD) for this modality. At the moment, most CAD systems are based on deep learning models. However, those models need large datasets to be trained effectively. In this work, to advance the research for pancreatic cancer CAD, we introduce a novel endoscopic ultrasound dataset for the classification task of pancreatic tumors. We present a dataset of 7825 endoscopic images corresponding to 606 patients. The dataset is divided into two classes: ultrasound images containing tumors (175 examinations), with corresponding segmentation masks, and ultrasound images without tumors (431 examinations). The clinical conditions of all patients was verified through histopathological examinations. This dataset has been manually divided into training and testing sets with respect to examinations. Benchmark results for classification, segmentation and class activation maps using state-of-the-art models are also presented. For classification, the best results were obtained with the model EfficientNetV2 with an accuracy of 89.88% and an Area Under Curve (AUC) of 96.12%. For segmentation, the model U-Net achieved a Dice score of 79.11%. For explainable AI, heatmaps are compared with segmentation masks using a soft Dice score; the soft DICE scores ranged between 8.69 and 40.81%. Classification and segmentation results are results are promising, highlighting the potential of CAD for pancreatic tumor analysis. However, there is a need to bridge the gap between the classification performance and explainability of models.