Table 1.

Summary of neural networks used for pixel-wise semantic segmentation of retina layers in OCT images.

Network	Task	Dataset	Loss Function	Data Augmentation
ReLayNet [35]	7 layers and fluid	Duke SD-OCT public DME dataset [37] 11 B-scans each (512 × 740 px), 110 images in total	Weighted Dice and Cross Entropy Loss	horizontal flip, spatial translation, cropping
3D ReLayNet [38]	7 layers	13 volumes (13 normal subjects), 10 B-scans each, 130 images in total	Cross Entropy Loss	none
FCN8 [39]	4 layers	10 volumes (5 patients with CSC, 5 normal eyes), 128 B-scans each (512 × 1024 px), 1280 images in total	Weighted Cross Entropy Loss	none
LFUNet [40]	5 layers and fluid	58 volumes (25 diabetic patients, 33 healthy subjects), 245 B-scans each (245 × 245 px), 14210 images in total	Weighted Dice and Cross Entropy Loss	horizontal flip, rotation, scaling
DRUNet [41]	6 regions	100 scans (40 healthy, 41 POAG, 19 PACG), single B-scan through ONH each (468 px width), 100 images in total	Jaccard Loss	horizontal flip, rotation, intensity shifts, white noise, speckle noise, elastic deformation, occluding patches
Uncertainty UNet (U2-Net) [42]	photoreceptor layer	50 volumes (50 patients: 16 DME, 24 RVO, 10 AMD+CNV), 49 B-scans each (512 × 496 px), 2450 images in total	Cross Entropy Loss	none
UNet with pretrained ResNet weights [43]	4 layers	23 volumes (23 AMD patients), 128 B-scans each (1024 × 512 px), 1270 images in total	Weighted Log Loss	horizontal flip, rotation
2 cascaded UNets with residual blocks [44]	8 layers and pseudocysts	35 volumes (35 patients: 21 with macula sclerosis, 14 healthy), 49 B-scans each (496 × 1024 px), 1715 images in total	1st: Dice Loss, 2nd: MSE Loss	horizontal flip, vertical scaling