Table 2.

Algorithm descriptions and hyper parameters

Team	Architecture	Input size (slide layer level)	Optimization (learning rate)	Augmentation real-time	Pre-processing	Post-processing; inference for confidence
Fiffeb	Inception v3, RFC	256×256×3 (6) Patch	SGD (0.9)	Color augmentation, horizontal flip, random rotation	Otsu thresholding, tumor (> 90%) and non-tumor (0% and > 20%)	Generation of heat map with image level 7 and feeding morphological information into FRC; RFC output
DoAI	U-Net	512×512×3 (0) Patch	SGD (1e-1, decay 0.1 each 2 epochs)	Rotation, horizontal and vertical flip	None	De-noising for false-positive reduction; CNN output
GoldenPass	U-Net, Inception v3	256×256×3 (4) Patch	Adam (1e-3, 5e-4)	Rotation, horizontal and vertical flip, brightness (0.5-1)	Otsu thresholding, tumor (> 100%)	None; Max value for heat-map
SOG	Simple CNN	300×300×3 (4) Slide	Adadelta (1e-3)	None	None	None; CNN output

SGD, stochastic gradient descent; RFC, random forest classifier; CNN, convolutional neural network.