Table 2.

3D DenseNet architectures for EGFR mutation classification

Layer	Tensor size	Building blocks
Input	48 × 48 × 48 × 1
Convolution	48 × 48 × 48 × 32	3 × 3 × 3 conv
Pooling	24 × 24 × 24 × 32	2 × 2 × 2 average pool
Dense Block (1)	24 × 24 × 24 × 80	$\left[\begin{array}{c}\text{bn}-\text{leaky}\text{relu}-1\times 1\times 1\text{conv}\\ \text{bn}-\text{leaky}\text{relu}-3\times 3\times 3\text{conv}\end{array}\right]\times 3$
Compression and Pooling (1)	12 × 12 × 12 × 40	$\left[\begin{array}{c}\text{bn}-\text{leaky}\text{relu}-1\times 1\times 1\text{conv}\\ 2\times 2\times 2\text{average}\text{pool}\end{array}\right]$
Dense Block (2)	12 × 12 × 12 × 136	$\left[\begin{array}{c}\text{bn}-\text{leaky}\text{relu}-1\times 1\times 1\text{conv}\\ \text{bn}-\text{leaky}\text{relu}-3\times 3\times 3\text{conv}\end{array}\right]\times 6$
Compression and Pooling (2)	6 × 6 × 6 × 68	$\left[\begin{array}{c}\text{bn}-\text{leaky}\text{relu}-1\times 1\times 1\text{conv}\\ 2\times 2\times 2\text{average}\text{pool}\end{array}\right]$
Dense Block (3)	6 × 6 × 6 × 132	$\left[\begin{array}{c}\text{bn}-\text{leaky}\text{relu}-1\times 1\times 1\text{conv}\\ \text{bn}-\text{leaky}\text{relu}-3\times 3\times 3\text{conv}\end{array}\right]\times 4$
Compression and Pooling (3)	3 × 3 × 3 × 66	$\left[\begin{array}{c}\text{bn}-\text{leaky}\text{relu}-1\times 1\times 1\text{conv}\\ 2\times 2\times 2\text{average}\text{pool}\end{array}\right]$
Dense Block (4)	3 × 3 × 3 × 114	$\left[\begin{array}{c}\text{bn}-\text{leaky}\text{relu}-1\times 1\times 1\text{conv}\\ \text{bn}-\text{leaky}\text{relu}-3\times 3\times 3\text{conv}\end{array}\right]\times 3$
Global Pooling (DLR)	114	3 × 3 × 3 average pool
Output	1	sigmoid