Table 2. Algorithm testing on 1431 images with a patient split used for training.

		Deep learning framework
		Colon_aphtae	_debris	_fissure	_non-ulc. Inflamm.	_normal	_ulcer	SB_aphtae	_fissure	_non-ulc. Inflamm.	_normal	_debris	_lymph. hyp.	_ulcer	Total
Gold standard	Colon_aphtae	34	2	0	3	0	3	3	0	0	0	0	0	0	45
	_debris	0	223	0	0	15	0	0	0	0	0	5	0	0	243
	_fissure	6	1	0	6	0	20	0	9	0	0	0	0	0	42
	_non-ulc. Inflamm.	21	2	1	39	1	2	1	0	4	0	1	0	0	72
	_normal	0	6	0	0	34	0	0	0	0	0	0	0	0	40
	_ulcer	9	1	0	8	0	27	0	5	0	0	0	0	1	51
	SB_aphtae	0	0	0	0	0	0	109	0	4	0	0	0	11	124
	_fissure	0	0	1	0	0	5	3	50	1	0	3	0	14	77
	_non-ulc. Inflamm.	0	0	0	1	0	0	2	2	0	0	0	0	0	5
	_normal	0	0	0	0	0	0	0	0	0	186	16	0	0	202
	_debris	0	0	0	0	0	0	0	0	0	17	374	0	1	392
	_lymph. hyp.	0	1	0	0	0	0	0	0	0	0	1	8	0	10
	_ulcer	1	0	0	1	0	8	46	39	0	0	4	0	29	128
	Total	71	236	2	58	50	65	164	105	9	203	404	8	56	1431

The matrix displays the number of images according to their classification with the gold standard and deep learning framework depending on the location. Lesions were assigned to one of 13 predefined categories. The inter-modality agreement was substantial (κ = 0.74).