Table 2

Performance of the deep learning models

A. Performance of the DL model for MH aetiology classification	AUC	ACC	SPE	SEN
Training	1.000	0.998	1.000	1.000
Validation	0.997	0.986	0.994	0.964
Testing	0.965	0.950	0.870	0.938

B. Performance of the models for postoperative MH status prediction	AUC	ACC	SPE	SEN
MDFN
Training	0.928	0.855	0.897	0.808
Validation	0.881	0.826	0.746	0.912
Testing	0.904	0.825	0.977	0.766
VGG
Training	0.953	0.901	0.855	0.922
Validation	0.805	0.778	0.887	0.581
Testing	0.804	0.758	0.872	0.656
FCN
Training	0.807	0.789	0.759	0.723
Validation	0.776	0.791	0.550	0.936
Testing	0.797	0.813	0.652	0.829

C. Performance of the models for postoperative IMH status prediction	AUC	ACC	SPE	SEN
MDFN
Training	0.999	0.988	0.989	0.987
Validation	0.974	0.901	1.000	0.865
Testing	0.947	0.875	0.815	0.979
VGG
Training	0.969	0.901	0.955	0.880
Validation	0.891	0.840	0.782	0.873
Testing	0.836	0.755	0.800	0.762
FCN
Training	0.873	0.782	0.733	0.876
Validation	0.926	0.828	0.800	0.954
Testing	0.768	0.717	0.625	0.892

ACC, accuracy; AUC, the area under the receiver operating characteristic curve; FCN, fully connected network; IMH, idiopathic macular hole; MDFN, multimodal deep fusion network; MH, macular hole; SEN, sensitivity; SPE, specificity; VGG, Visual Geometry Group.