Table 3.

Details of the unsupervised lesion detector at each ConvNet’s layer (unit: pixel).

Layer Type			Kernel Attribute	Num of Filters
Image Input Layer
Lesion Detector	Conv1	Convolutional Layer	3 × 3, stride = 1, padding = same	64
		ReLU Layer
		Convolutional Layer	3 × 3, stride = 1, padding = same	64
		ReLU Layer
		Max Pooling		2 × 2
	Conv2	Convolutional Layer	3 × 3, stride = 1, padding = same	128
		ReLU Layer
		Convolutional Layer	3 × 3, stride = 1, padding = same	128
		ReLU Layer
		Max Pooling		2 × 2
	Conv3	Convolutional Layer	3 × 3, stride = 1, padding = same	256
		ReLU Layer
Lesion Detector	Conv3	Convolutional Layer	3 × 3, stride = 1, padding = same	256
		ReLU Layer
		Convolutional Layer	3 × 3, stride = 1, padding = same	256
		ReLU Layer
		Max Pooling		2 × 2
	Conv4	Convolutional Layer	3 × 3, stride = 1, padding = same	512
		ReLU Layer
		Convolutional Layer	3 × 3, stride = 1, padding = same	512
		ReLU Layer
		Convolutional Layer	3 × 3, stride = 1, padding = same	512
		ReLU Layer
	Conv5	Convolutional Layer	3 × 3, stride = 1, padding = same	512
		ReLU Layer
		Convolutional Layer	3 ×3, stride = 1, padding = same	512
		ReLU Layer
		Convolutional Layer	3 × 3, stride = 1, padding = same	512
		ReLU Layer
	Conv6_1 - Conv6_5	Convolutional Layer	1 × 1, stride = 0, no padding	7 × 7 × 6
		PS ROI Pooling		7 × 7
		Fully Connected Layer