Normalization of HE-stained histological images using cycle consistent generative adversarial networks

Abstract

Background: Histological images show strong variance (e.g. illumination, color, staining quality) due to differences in image acquisition, tissue processing, staining, etc. This can impede downstream image analysis such as staining intensity evaluation or classification. Methods to reduce these variances are called image normalization techniques.

Methods: In this paper, we investigate the potential of CycleGAN (cycle consistent Generative Adversarial Network) for color normalization in hematoxylin-eosin stained histological images using daily clinical data with consideration of the variability of internal staining protocol variations. The network consists of a generator network G_B that learns to map an image X from a source domain A to a target domain B, i.e. G_B:X_A→X_B. In addition, a discriminator network D_B is trained to distinguish whether an image from domain B is real or generated. The same process is applied to another generator-discriminator pair (G_A,D_A), for the inverse mapping G_A:X_B→X_A. Cycle consistency ensures that a generated image is close to its original when being mapped backwards (G_A(G_B(X_A))≈X_A and vice versa). We validate the CycleGAN approach on a breast cancer challenge and a follicular thyroid carcinoma data set for various stain variations. We evaluate the quality of the generated images compared to the original images using similarity measures. In addition, we apply stain normalization on pathological lymph node data from our institute and test the gain from normalization on a ResNet classifier pre-trained on the Camelyon16 data set.

Results: Qualitative results of the images generated by our network are compared to original color distributions. Our evaluation indicates that by mapping images to a target domain, the similarity training images from that domain improves up to 96%. We also achieve a high cycle consistency for the generator networks by obtaining similarity indices greater than 0.9. When applying the CycleGAN normalization to HE-stain images from our institute the kappa-value of the ResNet-model that is only trained on Camelyon16 data is increased more than 50%.

Conclusions: CycleGANs have proven to efficiently normalize HE-stained images. The approach compensates for deviations resulting from image acquisition (e.g. different scanning devices) as well as from tissue staining (e.g. different staining protocols), and thus overcomes the staining variations in images from various institutions.The code is publicly available at https://github.com/m4ln/stainTransfer_CycleGAN_pytorch . The data set supporting the solutions is available at https://doi.org/10.11588/data/8LKEZF .

Keywords: Deep learning; Digital pathology; Generative adversarial networks; HE-stain; Histology stain normalization; Style transfer; Unpaired image-to-image translation.