Table 7.

The statistical analysis of KL-grade-based KOA classification and MOAKS score prediction using different aggregation methods for patch-level trabecular parameters.

Methods	Metrics	MLP(CT)	CNN (CT)	Transformers (CT)	Transformers (MR)
KL-grade-based KOA classification	Precision	0.843 ± 0.058	0.846 ± 0.051	0.863 ± 0.053	0.887 ± 0.061
	Recall	0.836 ± 0.074	0.899 ± 0.065	0.903 ± 0.068	0.935 ± 0.037
	F1 score	0.844 ± 0.067	0.868 ± 0.050	0.879 ± 0.056	0.910 ± 0.044
	AUC	0.742 ± 0.032	0.872 ± 0.033	0.883 ± 0.029	0.914 ± 0.048
MOAKS score prediction	Mean Average Error	1.076 ± 0.146	0.945 ± 0.052	0.893 ± 0.041	0.876 ± 0.038
	R-Square	0.837 ± 0.062	0.866 ± 0.044	0.881 ± 0.037	0.903 ± 0.030

The classification is implemented using regressed trabecular parameters on image patches as features. Here we compare different aggregation methods to employ regressed trabecular parameters. Our CT-SMA employs the proposed Transformers layer for aggregation. The first four rows are results of KL-grade-based KOA classification.

Bold values indicate the best performance among the CT-based methods.