Table 4.

Deep learning studies in pancreatic cancer detection and diagnosis.

Reference	Image	Software	Endpoints	Sample Size (Training + Validation)	Results
Chu et al. [69]	CT	Deeply supervised nets with encoder-decoder architecture	PDAC detection	456	Sensitivity = 94.1%, specificity = 98.5%
Liu et al. [70]	CT	CNN	Differentiating pancreatic cancer vs. normal pancreas	295 + 691 (local test 1 + local test 2 + external test)	AUC = 0.997 (local test 1) AUC = 0.999 (local test 2) AUC = 0.920 (external test)
Ozkan et al. [71]	EUS	ANN with Relief-F feature reduction method	Pancreatic cancer diagnosis for different age groups	260 + 72	Age groups in years: <40, 40–60, >60: accuracy = 92%, 88.5%, 91.7%, respectively all age groups: accuracy = 87.5%
Săftoiu et al. [72]	EUS	ANN (MLP)	Differential diagnosis of chronic pancreatitis and pancreatic cancer	68 (10-fold cross validation)	Benign vs. malignant pancreatic lesions: AUC = 0.957 Chronic pancreatitis vs. pancreatic cancer: AUC = 0.965
Săftoiu et al. [73]	EUS	ANN (MLP)	Diagnosis of focal pancreatic masses	258 (10-fold cross validation)	Average AUC = 0.94 over 100 runs of a complete cross-validation cycle
Si et al. [74]	CT	CNN ResNet18 (pancreas location), U-Net32 (pancreas segmentation), ResNet34 (pancreatic tumor diagnosis)	Fully automated diagnosis of pancreatic tumors	319 + 347	AUC = 0.871 on testing for detection of all tumor types
Tonozuka et al. [75]	EUS	CNN	PDAC detection	92 (10-fold cross validation) + 47	AUC = 0.924 (cross validation) AUC = 940 (test)
Zhang et al. [76]	CT	Faster R-CNN combined with Feature Pyramid Network for feature extraction	Pancreatic tumor detection	2650 + 240 (images)	AUC = 0.946

Abbreviations used in this table: Artificial Neural Network (ANN), Area Under Curve (AUC), Convolutional Neural Network (CNN), Multilayer Perceptron (MLP), Pancreatic Adenocarcinoma (PDAC).