A radiographic, deep transfer learning framework, adapted to estimate lung opacities from chest x-rays

Table 8 Performance analysis of best model

Population	N	Avg MA- MAE	Lt MA-MAE	Rt MA- MAE	Avg F1	Lt F1	Rt F1	Avg Pr	Lt Pr	Rt Pr	Avg Rec	Lt Rec	Rt Rec
All Patients	2839		0.4769	0.3855		62.11	65.21		62.24	66.22		63.08	65.41
RACE
Asian	184	0.3971	0.4536	0.3406	64.25	62.72	65.78	65.56	62.57	68.54	65.76	64.67	66.84
Black	453	0.4626	0.5408	0.3844	64.34	60.98	67.69	65.54	62.75	68.33	65.45	62.69	68.21
White	1346	0.438	0.4736	0.4025	64.08	63.34	64.82	64.36	63.26	65.45	64.37	63.96	64.78
Other	739	0.4196	0.4579	0.3813	62.00	60.37	63.64	63.39	60.83	65.95	62..65	61.43	63.87
Unknown	117	0.4039	0.4510	0.3569	64.45	60.73	68.16	65.59	61.45	69.73	65.81	62.39	69.23
SEX
Male	1625	0.4295	0.4937	0.3652	63.16	60.07	66.25	64.03	60.63	67.43	63.75	61.16	66.33
Female	1214	0.4129	0.4482	0.3775	66.94	65.68	68.20	67.56	65.64	69.48	67.58	66.63	68.53
COVID
COVID +ve	1513	0.4216	0.4783	0.3649	61.58	55.40	67.76	62.44	69.64	66.32	62.12	55.71	68.53
COVID -ve	1326	0.4584	0.5262	0.3907	67.44	69.46	65.43	67.98	56.29	68.59	67.87	70.58	65.16

The final model was comprised of a fine-tuned ResNet-50 architecture with no ROI segmentation and an undersampling scheme. The performance analysis is performed across race, sex and COVID-19 status. Abbreviations used in the table include Pr = Precision, Rec = Recall, F1 = F1 score, and MA-MAE = Macro-averaged Mean Absolute Error

ISSN: 2332-8886