Table 3.

Detection Errors of AI According to the Quantity of Learning Data and the Number of Detecting Targets Implemented During the Training Process^a

		Number of Learning Data (N)
Number of Landmarks (M)		n = 50	n = 100	n = 200	n = 400	n = 800	n = 1200	n = 1600	n = 2000
m = 19	First trial	3.74	2.54	2.80	1.89	2.19	1.84	1.66	1.63
	Second trial	2.78	2.71	2.39	2.07	1.73	1.96	1.55	1.64
	Third trial	2.82	2.88	2.68	2.56	2.17	2.09	1.76	1.59
	Fourth trial	2.98	2.59	2.49	2.45	1.84	1.72	1.50	1.58
	Average	3.08	2.68	2.59	2.24	1.98	1.90	1.62	1.61
m = 40	First trial	3.31	2.9	2.67	1.98	2.18	1.75	1.74	1.52
	Second trial	3.79	2.63	2.99	1.95	1.87	1.89	1.96	1.89
	Third trial	4.48	3.55	2.89	2.43	1.82	1.99	1.80	1.79
	Fourth trial	2.93	2.98	2.63	2.50	2.27	1.79	2.05	1.85
	Average	3.63	3.02	2.80	2.22	2.04	1.86	1.89	1.76
m = 80	First trial	3.51	3.38	3.08	2.52	2.14	1.98	2.27	1.76
	Second trial	3.21	3.27	2.43	2.55	2.26	2.13	1.98	1.94
	Third trial	3.75	2.78	3.15	2.44	2.32	2.15	2.09	2.15
	Fourth trial	3.58	3.12	3.19	2.99	2.41	2.36	1.91	1.76
	Average	3.51	3.14	2.96	2.63	2.28	2.16	2.06	1.90

^aThe values are mean radial error (MRE) by each trained AI in millimeter units for 200 test data sets. For each N, M combination, four random samples of size N were drawn from 2200 images.