Table 1.

Comparison on the PGT dataset (average ACC % and PR % for each disease. IF, FF, and DF represent input-level fusion, feature-level fusion and decision-level fusion, respectively). All the results are presented as mean ± standard deviation. The throughput is calculated as the time consumption (second) of predicting 100 images, which is measured using 2080Ti GPU.

Method	Dim	Fusion	Params	Throughput	ACC	PA	WT	MT	BCA	OBL
P3D	3D	IF	67 M	15.8	76.1 ± 5.5	59.9 ± 23.1	84.3 ± 5.3	69.7 ± 19.0	71.4 ± 7.3	78.0 ± 14.0
C3D	3D	IF	28 M	32.3	71.0 ± 4.1	68.3 ± 38.9	81.3 ± 15.4	67.8 ± 10.0	71.4 ± 7.5	84.5 ± 12.4
Resnet34	2D	IF	22 M	4.0	69.9 ± 4.0	81.0 ± 11.0	77.6 ± 7.4	61.4 ± 10.8	53.8 ± 15.5	68.1 ± 7.9
Resnet152	2D	IF	58 M	6.3	69.0 ± 3.5	50.5 ± 18.0	74.1 ± 10.1	64.3 ± 20.0	62.9 ± 8.9	75.2 ± 11.4
3D Resnet34	3D	IF	64 M	23.8	73.3 ± 5.1	69.2 ± 16.2	86.8 ± 5.2	75.3 ± 18.1	68.7 ± 13.8	65.7 ± 6.4
BoTNet50	3D	IF	21 M	4.2	77.4 ± 2.0	82.6 ± 14.4	84.0 ± 4.1	76.1 ± 6.9	76.0 ± 12.4	70.8 ± 8.4
[15]	3D	FF	45 M	27.9	74.2 ± 2.9	76.0 ± 24.8	86.2 ± 10.0	72.9 ± 16.3	75.2 ± 24.6	80.0 ± 15.2
[16]	3D	FF	130 M	30.8	73.3 ± 2.4	46.2 ± 13.2	78.4 ± 7.9	70.2 ± 15.5	69.8 ± 15.8	79.0 ± 10.7
P3D	3D	DF	136 M	22.1	74.8 ± 4.6	50.5 ± 20.0	85.1 ± 4.4	70.5 ± 20.2	69.5 ± 8.7	73.4 ± 14.3
C3D	3D	DF	57 M	41.3	71.0 ± 3.3	58.3 ± 33.3	70.7 ± 9.3	74.0 ± 8.5	78.9 ± 20.0	73.2 ± 6.7
Resnet34	2D	DF	45 M	5.6	71.3 ± 4.5	72.7 ± 21.7	75.3 ± 8.1	72.5 ± 10.3	60.9 ± 16.8	70.3 ± 9.7
Resnet152	2D	DF	116 M	9.5	72.2 ± 5.5	63.5 ± 18.3	75.6 ± 10.4	73.4 ± 18.7	83.2 ± 16.3	69.6 ± 11.5
3D Resnet34	3D	DF	128 M	34.6	72.1 ± 3.5	64.7 ± 14.4	81.5 ± 9.3	66.8 ± 8.4	69.6 ± 8.9	72.1 ± 14.9
BoTNet50	3D	DF	44 M	6.6	78.8 ± 3.4	72.6 ± 6.2	82.9 ± 4.3	73.2 ± 8.2	76.9 ± 16.2	87.9 ± 8.2
TransMed-T	2D	——	17 M	4.2	87.0 ± 2.6	80.1 ± 13.8	87.3 ± 3.0	90.7 ± 5.1	82.5 ± 15.3	93.6 ± 3.3
TransMed-S	2D	——	43 M	4.5	88.9 ± 3.0	90.1 ± 12.2	89.2 ± 6.8	92.0 ± 4.4	82.9 ± 9.3	88.3 ± 6.1
TransMed-B	2D	——	110 M	6.3	87.4 ± 2.1	86.2 ± 15.2	88.4 ± 3.8	88.2 ± 7.0	84.8 ± 13.8	92.2 ± 8.0