Stain transfer using Generative Adversarial Networks and disentangled features

With the digitization of histopathology, machine learning algorithms have been developed to help pathologists. Color variation in histopathology images degrades the performance of these algorithms. Many models have been proposed to resolve the impact of color variation and transfer histopathology images to a single stain style. Major shortcomings include manual feature extraction, bias on a reference image, being limited to one style to one style transfer, dependence on style labels for source and target domains, and information loss. We propose two models, considering these shortcomings. Our main novelty is using Generative Adversarial Networks (GANs) along with feature disentanglement. The models extract color-related and structural features with neural networks; thus, features are not hand-crafted. Extracting features helps our models do many-to-one stain transformations and require only target-style labels. Our models also do not require a reference image by exploiting GAN. Our first model has one network per stain style transformation, while the second model uses only one network for manyto-many stain style transformations. We compare our models with six state-of-the-art models on the Mitosis-Atypia Dataset. Both proposed models achieved good results, but our second model outperforms other models based on the Histogram Intersection Score (HIS). Our proposed models were applied to three datasets to test their performance. The efficacy of our models was also evaluated on a classification task. Our second model obtained the best results in all the experiments with HIS of 0.88, 0.85, 0.75 for L-channel, a-channel, and b-channel, using the Mitosis-Atypia Dataset and accuracy of 90.3% for classification.

Automated summary

With the use of Generative Adversarial Networks (GANs) and feature disentanglement, we propose two models to address color variation in histopathology images. Our models extract features automatically using neural networks and can perform many-to-one or many-to-many stain style transformations. In experiments, our second model achieved good results and demonstrated higher accuracy in the classification task.