Table 3.

Classification results for PIRC, QRDR and PIMEC with varying input image sizes on the primary validation set.

Grading system	Input size	Macro-AUC	Accuracy	Quadratic-Weighted Kappa
PIRC	256	0.901	0.751	0.772
PIRC	299	0.919	0.785	0.834
PIRC	512	0.951	0.838	0.894
PIRC	1024	0.961	0.870	0.915
PIRC	2095*	0.962	0.869	0.910
PIRC	6 × 512a	0.958	0.944	0.904
QRDR	256	0.977	0.912	0.901
QRDR	299	0.981	0.922	0.914
QRDR	512	0.989	0.937	0.930
QRDR	1024	0.991	0.938	0.932
QRDR	2095*	0.991	0.925	0.914
QRDR	6 × 512a	0.991	0.962	0.938
PIMEC	256	0.959	0.928	0.813
PIMEC	299	0.970	0.923	0.803
PIMEC	512	0.979	0.935	0.832
PIMEC	1024	0.978	0.937	0.846
PIMEC	2095*	0.981	0.934	0.856
PIMEC	6 × 512a	0.983	0.973	0.871

Macro-AUC refers to area under macro average of ROC for each class one-vs-all manner.

*Trained with model using instance normalization layers and an optimizer with accumulation of 15 mini-batches.

^aEnsemble of six classifiers trained on same data with same input size.