Authors:
Gagandeep B. Daroach
1
;
Josiah A. Yoder
1
;
Kenneth A. Iczkowski
2
and
Peter S. LaViolette
2
Affiliations:
1
Electrical Engineering and Computer Science, Milwaukee School of Engineering, Milwaukee, WI 53202, U.S.A.
;
2
Radiology and Biomedical Engineering, Medical College of Wisconsin, Wauwatosa, WI 53226, U.S.A.
Keyword(s):
Medical Imaging, Histology, Prostate, Prostate Cancer, Deep Learning, Machine Learning, Latent Space, GAN, StyleGAN, StyleGAN2, Generative Adversarial Networks, Gleason.
Abstract:
For use of deep learning algorithms in clinical practice, detailed justification for diagnosis is necessary. Convolutional Neural Networks (CNNs) have been demonstrated to classify prostatic histology using the same diagnostic signals as pathologists. Using the StyleGAN series of networks, we demonstrate that recent advances in high-resolution image synthesis with Generative Adversarial Networks (GANs) can be applied to prostatic histology. The trained network can produce novel histology samples indistinguishable from real histology at 1024x1024 resolution and can learn disentangled representations of histologic semantics that separates at a variety of scales. Through blending of the latent representations, users have the ability to control the projection of histologic semantics onto a reconstructed image. When applied to the medical domain without modification, StyleGAN2 is able to achieve a Fréchet Inception Distance (FID) of 3.69 and perceptual path length (PPL) of 33.25.