TABLE 3.

Performance comparison of KneeXNet with state-of-the-art methods on the test set and an independent dataset.

Method	Test set			Independent dataset
	Abnormality	ACL tear	Meniscal tear	Abnormality	ACL tear	Meniscal tear
SVM	0.872 $\pm$ 0.013	0.841 $\pm$ 0.016	0.836 $\pm$ 0.015	0.865 $\pm$ 0.014	0.833 $\pm$ 0.017	0.828 $\pm$ 0.016
RF	0.885 $\pm$ 0.011	0.857 $\pm$ 0.014	0.849 $\pm$ 0.013	0.878 $\pm$ 0.012	0.849 $\pm$ 0.015	0.841 $\pm$ 0.014
GBM	0.901 $\pm$ 0.010	0.873 $\pm$ 0.012	0.865 $\pm$ 0.011	0.894 $\pm$ 0.011	0.865 $\pm$ 0.013	0.857 $\pm$ 0.012
2D CNN	0.923 $\pm$ 0.008	0.896 $\pm$ 0.010	0.889 $\pm$ 0.009	0.916 $\pm$ 0.009	0.888 $\pm$ 0.011	0.881 $\pm$ 0.010
3D CNN	0.937 $\pm$ 0.007	0.912 $\pm$ 0.009	0.905 $\pm$ 0.008	0.930 $\pm$ 0.008	0.904 $\pm$ 0.010	0.897 $\pm$ 0.009
Transformer	0.948 $\pm$ 0.006	0.925 $\pm$ 0.007	0.919 $\pm$ 0.007	0.941 $\pm$ 0.007	0.917 $\pm$ 0.008	0.911 $\pm$ 0.008
SENet	0.956 $\pm$ 0.005	0.934 $\pm$ 0.006	0.928 $\pm$ 0.006	0.949 $\pm$ 0.006	0.926 $\pm$ 0.007	0.920 $\pm$ 0.007
KneeXNet	0.985 $\pm$ 0.003 a,b	0.972 $\pm$ 0.004 a,b	0.968 $\pm$ 0.004 a,b	0.978 $\pm$ 0.004 a,b	0.964 $\pm$ 0.005 a,b	0.960 $\pm$ 0.005 a,b
Additional evaluation metrics for KneeXNet
Accuracy	0.968 $\pm$ 0.004	0.951 $\pm$ 0.005	0.946 $\pm$ 0.006	0.960 $\pm$ 0.005	0.942 $\pm$ 0.006	0.937 $\pm$ 0.007
Precision	0.972 $\pm$ 0.005	0.958 $\pm$ 0.006	0.953 $\pm$ 0.006	0.964 $\pm$ 0.006	0.949 $\pm$ 0.007	0.944 $\pm$ 0.007
Recall	0.979 $\pm$ 0.004	0.965 $\pm$ 0.005	0.961 $\pm$ 0.005	0.971 $\pm$ 0.005	0.956 $\pm$ 0.006	0.952 $\pm$ 0.006
F1 score	0.975 $\pm$ 0.004	0.961 $\pm$ 0.005	0.957 $\pm$ 0.005	0.967 $\pm$ 0.005	0.952 $\pm$ 0.006	0.948 $\pm$ 0.006
Specificity	0.933 $\pm$ 0.008	0.918 $\pm$ 0.009	0.914 $\pm$ 0.010	0.920 $\pm$ 0.009	0.905 $\pm$ 0.010	0.901 $\pm$ 0.011
P-values for KneeXNet vs. best competing method (SENet)
	$p=0.0003$	$p=0.0007$	$p=0.0008$	$p=0.0011$	$p=0.0018$	$p=0.0022$

Statistical significance test results:

^a Significantly better than all traditional ML methods (SVM, RF, GBM) with $p<0.001$

^b Significantly better than all deep learning methods (2D CNN, 3D CNN, transformer, SENet) with $p<0.01$

The bold values represent the best-performing results for each respective metric or category.