Table 2 Comparison of the performance of the proposed FKD-CSS model and different ophthalmologists on six external validation datasets

Ophthalmologist	Northern		Southern		Central		Northwestern		Southwestern		Northeastern		All
Ophthalmologist	Pearson	AUC	Pearson	AUC	Pearson	AUC	Pearson	AUC	Pearson	AUC	Pearson	AUC	Pearson	AUC
Junior A	0.771	0.622	0.809	0.676	0.821	0.624	0.840	0.678	0.783	0.568	0.795	0.563	0.800	0.655
Junior B	0.752	0.578	0.801	0.714	0.733	0.682	0.820	0.644	0.661	0.650	0.784	0.511	0.729	0.608
Mid-level A	0.830	0.705	0.801	0.712	0.858	0.731	0.854	0.749	0.891	0.756	0.842	0.675	0.855	0.729
Mid-level B	0.835	0.725	0.824	0.680	0.876	0.746	0.849	0.723	0.900	0.756	0.797	0.671	0.864	0.733
Senior A	0.884	0.740	0.834	0.714	0.914	0.772	0.883	0.749	0.915	0.806	0.881	0.740	0.898	0.762
Senior B	0.890	0.730	0.840	0.691	0.867	0.749	0.885	0.753	0.859	0.742	0.802	0.620	0.865	0.734
FKD-CSS	0.871	0.845	0.874	0.846	0.899	0.883	0.855	0.851	0.887	0.862	0.844	0.804	0.882	0.860

The best results are highlighted in bold a and the second-best results are underlined. All p-value of Pearson test are statistically significant (p-value < 0.001).
FKD fine-grained knowledge distillation, CSS corneal staining score, AUC area under the curve.

Quick links

Search