Table 7 Constitution of the training sets on the basis of Table 4, which aims to cover the variation in the magnification levels for model development.

From: Critical evaluation of artificial intelligence as a digital twin of pathologists for prostate cancer pathology

Finding/models

Considered datasets with magnification levels as training set

Augmentation features

(Augmentation probability: 50%)

GP3

 

Random brightness and contrast

Random image compression rates (variable image resolution)

Random flip (horizontal and vertical)

Random rotation (between -90 and 90)

Random hue saturation value

Random Gaussian noise

Random clip limits in the Contrast limited adaptive histogram equalization54

Model 1

5×, 10×, 20×

Model 2

10×, 20×

GP4

 

Model 1

10×, 20×

Model 2

Model 3

Model 4

Model 5

Model 6

GP5

 

Model 1

10×

Model 2

  1. Various approaches for patch augmentation were applied to increase the variation in patch appearance to increase the likelihood of obtaining more generalizable models.
Back to article page