Table 2.

CNN performance for detection and differentiation of multiple small bowel lesions with distinct haemorrhagic potential

	Sensitivity	Specificity	PPV	NPV
Overall, mean %±SD	87.8±8.1	99.4±3.7	86.8%±9.0%	99.4%±3.9%
P0L vs all, %	88.3	99.1	91.4	98.7
P0X vs all, %	85.3	99.7	91.7	99.5
P1E vs all, %	72.7	99.2	78.5	98.9
P1U vs all, %	81.0	99.4	85.2	99.2
P2U vs all, %	94.1	99.7	94.5	99.7
P1RS vs all, %	79.3	99.2	78.3	99.2
P2V vs all, %	90.7	99.4	80.4	99.7
P1P vs all, %	93.4	98.8	72.7	99.8
P2P vs all, %	94.3	99.9	95.7	99.9
B vs all, %	98.7	99.8	99.5	99.5
P0 vs all, %	87.9	99.6	97.2	98.2
P1 vs all, %	86.9	97.1	84.1	97.7
P2 vs all, %	93.8	99.1	91.0	99.4
P0 vs P1, %	95.2	98.0	97.8	95.7
P0 vs P2, %	99.6	99.2	99.4	99.5
P1 vs P2, %	97.3	96.5	97.6	96.0

The prefixes P0, P1, and P2 refer to the haemorrhagic potential according to Saurin’s classification.

CNN, convolutional neural network; NPV, negative predictive value; P1E, mucosal erosions; P0L, lymphangiectasia; P1P, P1 protruding lesions; P2P, P2 protruding lesions; PPV, positive predictive value; P1RS, P1 red spots; P1U, P1 ulcers; P2U, P2 ulcers; P2V, P2 vascular lesions; P0X, xanthomas.