Construction of a convolutional neural network classifier developed by computed tomography images for pancreatic cancer diagnosis

Table 1 Statistics of our datasets

	CT phase		Number of patients					Number of images
		Without cancer cancer	With cancer			Total	Without cancer	With cancer			Total
			At tail1	At head	Total			At tail	At head	Total
Sets	Plain Scan	182	91	123	214	396	1182	416	496	912	2094
	Arterial phase	179	91	129	220	399	1282	575	735	1310	2592
	Venous phase	178	93	129	222	400	1287	573	699	1272	2559
Total		539	275	381	656	1195	3751	1564	1930	3494	7245

¹“At tail” means at tail or body of pancreas, while “At head” means at head or neck of pancreas. CT: Computed tomography.

Table 2 Characteristics of study participants

Variables	With pancreatic cancer (n = 222)	Without cancer (n = 190)	Z value	P value
Age (yr)	63.8 (8.7)	61.0 (12.3)	3.391	0.66
Gender (male/female)	124/98	98/92	0.3852	0.54
Diagnosis method surgery/biopsy3	161/61	-	-	-
Tumor location; head and neck/tail and body	129/63	-	-	-
Tumor size4 [median (quartile 1, quartile 3)], cm	3.5 (2.7-4.3)	-	-	-
≤ 2	29	-	-	-
2-4	134	-	-	-
> 4	59	-	-	-

Data are expressed as the mean ± SD.

¹t value.

²χ² value.

³“Biopsy” includes patients who underwent biopsy only, patients who underwent biopsy before surgery were calculated in “Surgery”, and tumor size was evaluated in greatest dimension.

⁴Tumor size was mainly evaluated by the gross surgical specimen, for patient who only underwent biopsy, the tumor size was calculated by computed tomography image.

Table 3 Performance of the binary classifiers

	Plain scan	Arterial phase	Venous phase	χ2 value	P value
Accuracy	0.954747	0.957580	0.951549	0.346	0.841
Specificity	0.982710	0.975695	0.978692	0.149	0.928
Sensitivity	0.915758	0.940808	0.922756	0.914	0.633

Table 4 Diagnostic accuracy of the binary classifiers in plain scan: Convolutional neural network vs gastroenterologists and trainees

	CNN		Doctors
		Gastroenterologists	Trainees	Total
No. of doctors		10	15	25
Accuracy	0.954747	0.922	0.736	0.815
Specificity	0.982710	0.923	0.725	0.847
Sensitivity	0.915758	0.921	0.792	0.809

CNN: Convolutional neural network.

Table 5 Performance of the ternary classifiers

Dataset	Plain scan	Arterial phase	Venous phase	χ2 value	P value
Accuracy	0.820568	0.790633	0.788076	1.074	0.585
Specificity	0.985721	0.984770	0.990305	0.577	0.749
Sensitivity (cancer at the tail/body of pancreas)	0.520122	0.411098	0.360272	1.841	0.398
Sensitivity (cancer at the head/neck of pancreas)	0.462148	0.852390	0.728743	16.651	< 0.001