Deep Learning and Automatic Differentiation of Pancreatic Lesions in Endoscopic Ultrasound: A Transatlantic Study

Abstract

Introduction: Endoscopic ultrasound (EUS) allows for characterization and biopsy of pancreatic lesions. Pancreatic cystic neoplasms (PCN) include mucinous (M-PCN) and nonmucinous lesions (NM-PCN). Pancreatic ductal adenocarcinoma (P-DAC) is the commonest pancreatic solid lesion (PSL), followed by pancreatic neuroendocrine tumor (P-NET). Although EUS is preferred for pancreatic lesion evaluation, its diagnostic accuracy is suboptimal. This multicentric study aims to develop a convolutional neural network (CNN) for detecting and distinguishing PCN (namely M-PCN and NM-PCN) and PSL (particularly P-DAC and P-NET).

Methods: A CNN was developed with 378 EUS examinations from 4 international reference centers (Centro Hospitalar Universitário São João, Hospital Universitario Puerta de Hierro Majadahonda, New York University Hospitals, Hospital das Clínicas Faculdade de Medicina da Universidade de São Paulo). About 126.000 images were obtained-19.528 M-PCN, 8.175 NM-PCN, 64.286 P-DAC, 29.153 P-NET, and 4.858 normal pancreas images. A trinary CNN differentiated normal pancreas tissue from M-PCN and NM-PCN. A binary CNN distinguished P-DAC from P-NET. The total data set was divided into a training and testing data set (used for model's evaluation) in a 90/10% ratio. The model was evaluated through its sensitivity, specificity, positive and negative predictive values, and accuracy.

Results: The CNN had 99.1% accuracy for identifying normal pancreatic tissue, 99.0% and 99.8% for M-PCN and NM-PCN, respectively. P-DAC and P-NET were distinguished with 94.0% accuracy.

Discussion: Our group developed the first worldwide CNN capable of detecting and differentiating the commonest PCN and PSL in EUS images, using examinations from 4 centers in 2 continents, minimizing the impact of the demographic bias. Larger multicentric studies are needed for technology implementation.

Figure 1.

Study design for the development of the convolutional neural network. ADC, adenocarcinoma; M, mucinous; m-PCN, mucinous pancreatic cystic neoplasm; N, normal; NET, neuroendocrine tumor; NM, nonmucinous; NM-PCN, nonmucinous pancreatic cystic neoplasm; P-DAC, pancreatic adenocarcinoma; P-NET, pancreatic neuroendocrine tumor.

Figure 2.

Output obtained from the convolutional neural network. The bars are a representation of the estimated probability by the CNN. The model output was given by the finding with the highest probability. ADC, pancreatic adenocarcinoma; M, mucinous pancreatic cystic neoplasm; N, normal pancreas; NET, pancreatic neuroendocrine tumor; NM, nonmucinous pancreatic cystic neoplasm.