Deep learning has shown its effectiveness in histopathology image analysis, such as pathology detection and classification. However, stain colour variation in Hematoxylin and Eosin (H&E) stained histopathology images poses challenges in effectively training deep learning-based algorithms. To alleviate this problem, stain normalisation methods have been proposed, with most of the recent methods utilising generative adversarial networks (GAN). However, these methods are either trained fully with paired images from the target domain (supervised) or with unpaired images (unsupervised), suffering from either large discrepancy between domains or risks of undertrained/overfitted models when only the target domain images are used for training. In this paper, we introduce a colour adaptive generative network (CAGAN) for stain normalisation which combines both supervised learning from target domain and unsupervised learning from source domain. Specifically, we propose a dual-decoder generator and force consistency between their outputs thus introducing extra supervision which benefits from extra training with source domain images. Moreover, our model is immutable to stain colour variations due to the use of stain colour augmentation. We further implement histogram loss to ensure the processed images are coloured with the target domain colours regardless of their content differences. Extensive experiments on four public histopathology image datasets including TCGA-IDH, CAMELYON16, CAMELYON17 and BreakHis demonstrate that our proposed method produces high quality stain normalised images which improve the performance of benchmark algorithms by 5% to 10% compared to baselines not using normalisation.
Read full abstract