Table 2.

Summary of Unet based DL models along with the lightweight model, ‘LWBNA_Unet’ developed in the present study.

Types of deep learning	Total	File size	Extra contour		Dice coefficient (D)
Model	Parameters	(MB)	Total contour	Total images	Predicted	SD	Largest contour	SD
Unet	28,340,931	332	51	41	0.9581	0.0186	0.9586	0.0185
Unet_AB	29,830,275	350	24	18	0.9584	0.0186	0.9586	0.0185
Unet_AB_Upsampling	24,466,627	287	23	21	0.9589	0.0197	0.9594	0.0194
Unet_AB_Upsampling_Add	12,259,523	144	23	16	0.9557	0.0315	0.9561	0.0312
Unet_AB_128_Upsampling_Add	2,673,795	32	45	35	0.9576	0.0232	0.958	0.023
Unet_AB_64_Upsampling_Add	673,347	8.3	204	46	0.9555	0.0371	0.9571	0.0355
LWBNA_Unet (developed in current study)	2,958,819	35	29	27	0.9578	0.0195	0.9584	0.019

Performance of the trained DL models was tested on a testing dataset of 145 OCTA images. Dice similarity coefficient (D) was calculated between the manually segmented image, and the image predicted by the DL model. The largest contour D is calculated by selecting the biggest segmented area in the predicted image (when multiple FAZ region exist). SD is the standard deviation. All the models were trained for fixed number of epochs (500).

AB Attention block, Unet and Unet_AB are standard DL models without and with attention block. The numbers 128 or 64 mean the model has fixed channels/filters (128 or 64) in each 2D convolution layer. Unet or Unet_AB uses transpose convolution and concatenate instead of upsampling and add layers.