Table 1.

Architecture of the mitosis detection model.

Feature extractor and mitosis classifier					Domain classifier
	Layer	Size	Filter	Rec. F.	Layer	Output	Filter	Rec. F.
$\mathcal{F}$	Input	64 × 64 × 3		1 × 1
	Conv	60 × 60 × 16	5 × 5	5 × 5
	Max pool	30 × 30 × 16	2 × 2	6 × 6
	Conv	28 × 28 × 16	3 × 3	10 × 10
	Max pool	14 × 14 × 16	2 × 2	12 × 12	Bifurcation	14 × 14 × 16		12 × 12
$\mathcal{C}$	Conv	12 × 12 × 16	3 × 3	20 × 20	Conv	12 × 12 × 32	3 × 3	20 × 20	$\mathcal{D}$
	Max pool	6 × 6 × 16	2 × 2	24 × 24	Conv	10 × 10 × 64	3 × 3	24 × 24
	Conv	4 × 4 × 16	3 × 3	40 × 40	Softmax	10 × 10 × 8	1 × 1	24 × 24
	Max pool	2 × 2 × 16	2 × 2	48 × 48
	Conv	1 × 1 × 64	2 × 2	64 × 64
	Sigmoid	1 × 1 × 1	1 × 1	64 × 64

The feature extractor $\mathcal{F}$ and mitosis classifier $\mathcal{C}$ form a 10-layer CNN with a single class-probability output. The domain classifier $\mathcal{D}$ is a 3-layer network bifurcated at the second max-pooling layer of $\mathcal{F}$ and outputs a 8-domain probability vector.