Skin Lesion Segmentation Based on Edge Attention Vnet with Balanced Focal Tversky Loss

Table 1

ET-Vnet 2D models’ details. Inputs (batch size, height, width, depth, Channels) : (4,512,512,1,1).

Layer	Layers’ details	Output size	Activate/Norm/Weight initializer
Layer 1	Conv 3 × 3	(4,512,512,1,32)	GN, ReLU, Xavier
	ResNet block
Down 1	Filter 3 × 3	(4,256,256,1,64)	GN, ReLU, Xavier
	Filter 3 × 3
Layer 2	Filter 3 × 3		GN, ReLU, Xavier
	ResNet block
Down 2	Filter 3 × 3	(4,128,128,1,128)	GN, ReLU, Xavier
	Filter 3 × 3
Layer 3	Filter 3 × 3		GN, ReLU, Xavier
	Filter 3 × 3
	ResNet block
Down 3	Filter 3 × 3	(4,64,64,1,256)	GN, ReLU, Xavier
	Filter 3 × 3
Layer 4	Filter 3 × 3		GN, ReLU, Xavier
	Filter 3 × 3
	ResNet block
Down 4	Filter 3 × 3	(4,32,32,1,512)	GN, ReLU, Xavier
	Filter 3 × 3
Layer 5	Filter 3 × 3		GN, ReLU, Xavier
	Filter 3 × 3
	ResNet block
Upsample 1	Filter 3 × 3	(4,64,64,1,256)	GN, ReLU, Xavier
	Filter 3 × 3
Layer 6	Filter 3 × 3		GN, ReLU, Xavier
	Filter 3 × 3
	ResNet block
Unsample 2	Filter 3 × 3	(4,128,128,1,128)	GN, ReLU, Xavier
	Filter 3 × 3
Layer 7	Filter 3 × 3		GN, ReLU, Xavier
	Filter 3 × 3
	ResNet block
Upsample 3	Filter 3 × 3	(4,256,256,1,64)	GN, ReLU, Xavier
	Filter 3 × 3
Layer 8	Filter 3 × 3		GN, ReLU, Xavier
	Filter 3 × 3
	ResNet block
Upsample 4	Filter 3 × 3	(4,512,512,1,32)	GN, ReLU, Xavier
Layer 9	Filter 3 × 3
	Filter 3 × 3		GN, ReLU, Xavier
	ResNet block
Layer 10	ET-Net2D	(4,512,512,1,1) + ET-Net2D	GN, ReLU, Xavier
3D single output	Filter 1 × 1