[Paper Review] Neural Stain-Style Transfer Learning using GAN for Histopathological Images

Hyungjoo Cho, Sungbin Lim+2 more|arXiv (Cornell University)|Oct 23, 2017

AI in cancer detection13 references77 citations

TL;DR

The paper proposes a stain-style transfer (SST) network based on conditional GANs with a feature-preserving loss to transfer stain styles across institutes while preserving tumor classifier performance, evaluated on CAMELYON16 data.

ABSTRACT

Performance of data-driven network for tumor classification varies with stain-style of histopathological images. This article proposes the stain-style transfer (SST) model based on conditional generative adversarial networks (GANs) which is to learn not only the certain color distribution but also the corresponding histopathological pattern. Our model considers feature-preserving loss in addition to well-known GAN loss. Consequently our model does not only transfers initial stain-styles to the desired one but also prevent the degradation of tumor classifier on transferred images. The model is examined using the CAMELYON16 dataset.

Motivation & Objective

• Motivate the problem that classifier performance varies with stain styles across institutes.
• Propose SST as a GAN-based stain-style transfer to unify stain styles while preserving classifier features.
• Integrate reconstruction, conditional GAN, and feature-preserving losses to train the style generator.
• Demonstrate that SST preserves classifier performance and improves tumor classification accuracy on cross-institute data.

Proposed method

• Define stain-style as a per-institute distribution over color images.
• Use a gray normalization G and a stain-style generator zeta to form a stain-transfer tau = zeta o G.
• Train tau with three losses: reconstruction loss to keep intra-style consistency, conditional GAN loss to learn color mapping, and feature-preserving loss to maintain classifier features.
• Employ a fully convolutional FusionNet-based generator and a VGG-like discriminator in a GAN framework.
• Leverage a ResNet-34 classifier as both discriminator signal and feature extractor for the FP loss.

Experimental results

Research questions

• RQ1 Can a GAN-based stain-style transfer learn to map between institute stain styles without label or reference implementations from the target institute?
• RQ2 Does incorporating a feature-preserving loss prevent degradation of tumor classifier performance after stain-style transfer?
• RQ3 How does SST compare to traditional stain normalization methods in terms of tumor classification metrics on cross-institute data?

Key findings

• SST achieves highest tumor classification AUC (0.9185) on transferred images compared to other methods and original images.
• SST maintains classifier performance with AUC 0.9760 on target-like data, while original images drop to 0.8900.
• Compared with WSICS, HS, Macenko, and Reinhard, SST yields substantially better Precision, Recall, and Specificity on the cross-institute test set.
• SST outperforms visual stain normalization approaches while preserving histopathological patterns relevant to classification.