Table 8.

Dice score and Mean Absolute Error results of 4 class pixel segmentation with various kernel sizes for UNet models after removing anomalous data samples and expanding data set with 4 augmentation techniques.

Kernel [px]		Dice Score				MAE [px]			TII [%]
Size	Area	Vitreous	Preretinal Space	Retina	Space under Retina	PCV	ILM	RPE
3 × 3	9	0.9968	0.9591	0.9944	0.9984	1.33	0.50	0.57	15.1
5 × 5	25	0.9980	0.9615	0.9945	0.9988	1.08	0.47	0.54	4.6
7 × 7	49	0.9981	0.9635	0.9948	0.9988	0.92	0.45	0.52	4.6
5 × 3	15	0.9978	0.9615	0.9944	0.9987	1.17	0.48	0.58	5.7
7 × 3	21	0.9981	0.9628	0.9946	0.9988	0.95	0.47	0.54	2.8
9 × 3	27	0.9981	0.9626	0.9947	0.9988	0.93	0.47	0.52	2.4
3 × 5	15	0.9977	0.9620	0.9945	0.9987	1.22	0.47	0.55	9.7
3 × 7	21	0.9980	0.9631	0.9946	0.9987	0.98	0.46	0.55	6.8
3 × 9	27	0.9981	0.9640	0.9946	0.9987	0.90	0.46	0.54	5.1