Table 3.

Performance of few-shot learning models and radiologists in Cohort #1, round #3.

Round #3	Confusion Matrix					Metrics
		VGG16+XGBoost
		Normal	OP	OS	Total	Accuracy	0.660
Actual	Normal	39	10	30	79	Sensitivity (OP)	0.659
	OP	4	60	27	91	Sensitivity (OS)	0.777
	OS	8	17	87	112	Sensitivity (mean)	0.718
	Total	51	87	144	282	Specificity	0.494
		ResNet50+XGBoost
		Normal	OP	OS	Total	Accuracy	0.589
Actual	Normal	33	21	25	79	Sensitivity (OP)	0.560
	OP	10	51	30	91	Sensitivity (OS)	0.732
	OS	14	16	82	112	Sensitivity (mean)	0.646
	Total	57	88	137	282	Specificity	0.418
		Xception+XGBoost
		Normal	OP	OS	Total	Accuracy	0.730
Actual	Normal	46	11	22	79	Sensitivity (OP)	0.648
	OP	7	59	25	91	Sensitivity (OS)	0.902
	OS	7	4	101	112	Sensitivity (mean)	0.775
	Total	60	74	148	282	Specificity	0.582
		Junior radiologist
		Normal	OP	OS	Total	Accuracy	0.496
Actual	Normal	51	14	14	79	Sensitivity (OP)	0.418
	OP	23	38	30	91	Sensitivity (OS)	0.455
	OS	28	33	51	112	Sensitivity (mean)	0.436
	Total	102	85	95	282	Specificity	0.646
		Senior radiologist
		Normal	OP	OS	Total	Accuracy	0.596
Actual	Normal	56	10	13	79	Sensitivity (OP)	0.451
	OP	19	41	31	91	Sensitivity (OS)	0.634
	OS	9	32	71	112	Sensitivity (mean)	0.542
	Total	84	83	115	282	Specificity	0.709

Bold figures indicate the highest numeric values.

OP, osteopenia; OS, osteoporosis; XGBoost, eXtreme gradient boosting.