Artificial intelligence for pancreatic cancer detection: Recent development and future direction

Table 1 Summary of studies assessing computed tomography and magnetic resonance using artificial intelligence-based approach for pancreatic cancer

Ref.	Overall dataset	Testing data	Model	Model performance on testing data
				Accuracy (%)	AUC	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)
CT
Zhu et al[43], 2019 (United States)	439 cases	23 cases	CNN	NA	NA	94.1	98.5	NA	NA
Liu et al[42], 2019 (China)	338 patients	100 patients	CNN	NA	0.9632	NA	NA	NA	NA
Chu et al[44], 2019 (China)	380 patients	125 patients	ML	99.2%	0.999	100	98.5	NA	NA
Li et al[75], 2018 (China)	206 patients	No separate testing data (10-fold CV)	CNN	72.8%1	NA	NA	NA	NA	NA
Wei et al[37], 2018 (China)	260 patients	60 patients	SVM	NA	0.837	66.7	81.8	NA	NA
MR
Kaissis et al[49], 2020 (Germany)	207 patients	26 patients	ML	NA	0.93	84	92	NA	NA
Corral et al[50], 2019 (United States)	139 cases	No separate testing data (10-fold CV)	DL	NA	0.781	921	52%1	NA	NA
Gao et al[51], 2019 (China)	96 patients	No separate testing data (5-fold CV	DL	85.131	0.91171	NA	NA	NA	NA

¹The performance was based on n-fold cross-validation on training data.

AUC: Area under the curve; CNN: Convolutional neural network; CT: Computed tomography; CV: Cross-validation; DL: Deep learning; IPMN: Intraductal papillary mucinous neoplasm; MR: Magnetic resonance; NA: Not available; NN: Neural network; NPV: Negative predictive value; PCA: Principal component analysis; PPV: Positive predictive value; SVM: Support vector machine.

Table 2 Summary of endoscopic ultrasound using artificial intelligence-based approach studies pancreatic cancer and malignant pancreatic cyst detection

Ref.	Overall dataset	Testing data	Model	Model performance on testing data
				Accuracy (%)	AUC	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)
Marya et al[62], 2020 (United States)	583 patients (1174461 images)	123 patients	CNN	NA	0.976	95	91	87	97
Tonozuka et al[33], 2020 (Japan)	139 patients (920 images)	47 patients (470 images)	CNN	NA	0.94	92.4	84.1	86.8	90.7
Ozkan et al[54], 2016 (Turkey)	332 images	72 images	NN	87.5	NA	83.3	93.33	NA	NA
Saftoiu et al[59], 2015 (Multicenter in Europe)	167 cases	15% of cases	NN	NA	NA	94.64	94.44	97.24	89.47
Zhu et al[56], 2013 (China)	388 images	50% of all data (200 trials)	SVM	93.86	NA	92.52	93.03	91.75	94.39
Zhang et al[55], 2010 (China)	216 patients	50% of all data (50 trials)	SVM	97.98	NA	94.32	99.45	98.65	97.77
Das et al[57], 2008 (United States)	319 images	50% of all data	NN	NA	0.93	93	92	87	96
Norton et al[53], 2001 (United States)	21 patients	4 patients	ML	80	NA	100	50	NA	NA
Elastography
Saftoiu et al[61], 2012 (Multicenter in Europe)	258 cases	No separate testing data (10-fold CV)	NN	84.272	0.942	87.592	82.942	96.252	57.222
Saftoiu et al[60], 2008 (Denmark and Romania)	68 cases	No separate testing data (10-fold CV)	NN	NA	0.9572	NA	NA	NA	NA
IPMN
Machicado et al[64], 2021 (United States)1	35 cases of EUS-nCLE (15027 frames)	No separate testing data (5-fold CV)	(1) CNN (segmentation); and (2) CNN (holistic)	(1) 82.92; and (2) 85.72	NA	(1) 83.32; and (2) 83.32	(1) 82.42; and (2) 88.22	(1) 83.32; and (2) 88.22	(1) 82.42; and (2) 83.32
Kuwahara et al[63], 2019 (Japan)	50 cases	No separate testing data (10-fold CV)	CNN	942	NA	95.72	92.62	91.72	96.22

¹Presented two designs of CNN algorithms: segmentation based model and holistic based model.

²The performance was based on n-fold cross-validation on training data.

AUC: Area under the receiver operating characteristic curve; CE-EUS: Contrast enhanced endoscopic ultrasound; CNN: Convolutional neural network; CV: Cross-validation; EUS-nCLE: Endoscopic ultrasound-guided needle based confocal laser endomicroscopy; IPMN: Intraductal papillary mucinous neoplasm; NA: Not available; NN: Neural network; NPV: Negative predictive value; PCA: Principal component analysis; PPV: Positive predictive value; SVM: Support vector machine.