Table 2.

Diagnostic performance of the transformer-based model on a pixel level after 250 training epochs and additional finetuning.

Diagnostic category	Total pixel number (N × 106)	F1	IoU	Average precision	Accuracy
Diagnostic performance after 250 training epochs (baseline training)
Caries
Non-cavitation	6.096	0.595	0.423	0.683	0.983
Grayish translucency	0.260	0.347	0.210	0.420	0.999
Enamel breakdown	0.139	0.161	0.088	0.468	0.999
Dentin cavity	2.713	0.763	0.617	0.751	0.995
Fully destructed tooth	0.922	0.630	0.460	0.542	0.997
Molar–incisor hypomineralization (chalky teeth)
Demarcated opacity	3.715	0.586	0.423	0.657	0.990
Enamel disintegration	0.688	0.604	0.433	0.674	0.998
Atypical restoration	1.552	0.669	0.503	0.704	0.996
None	246.057	0.984	0.969	0.980	0.970
Total	262.142	0.962	0.937	0.961	0.964
Diagnostic performance after 250 training epochs + finetuning
Caries
Non-cavitation	6.386	0.773	0.630	0.813	0.990
Grayish translucency	0.292	0.746	0.595	0.743	0.999
Enamel breakdown	0.136	0.521	0.352	0.588	0.999
Dentin cavity	2.471	0.818	0.692	0.830	0.997
Fully destructed tooth	1.674	0.881	0.787	0.882	0.999
Molar–incisor hypomineralization (chalky teeth)
Demarcated opacity	4.758	0.804	0.672	0.827	0.993
Enamel disintegration	0.322	0.673	0.507	0.669	0.999
Atypical restoration	1.566	0.906	0.829	0.902	0.999
None	244.539	0.990	0.979	0.988	0.981
Total	262.144	0.977	0.959	0.977	0.978