Cascade and Fusion of Multitask Convolutional Neural Networks for Detection of Thyroid Nodules in Contrast-Enhanced CT

Table 1

Architecture of DenseU-Net, where Dense-1_/32 represents a Dense block structure, and its growth rate is 32.

Encoder	Output size	Skip connection	Decoder	Output size
Input	256^∧2 × 1		Conv6	256^∧2 × 1
Dense-1/32	256^∧2 × 64	Res1	Conv5	256^∧2 × 2
Pooling	128^∧2 × 64		Up4	256^∧2 × 64
Dense-2/32	128^∧2 × 128	Res2	Conv4	128^∧2 × 64
Pooling	64^∧2 × 128		Up3	128^∧2 × 128
Dense-2/64	64^∧2 × 256	Res3	Conv3	64^∧2 × 128
Pooling	32^∧2 × 256		Up2	64^∧2 × 256
Dense-4/64	32^∧2 × 512	Res4	Conv2	32^∧2 × 256
Pooling	16^∧2 × 512		Up1	32^∧2 × 512
Dense-4/128	16^∧2 × 1024		Conv1	16^∧2 × 512

Pooling indicates that the maximum pooling layer has a pooling window of 2^∧2 and a step size of 2. Conv represents the convolutional layer. Up indicates the upsampling layer. Res stands for Residual block.