The role of unpaired image-to-image translation for stain color normalization in colorectal cancer histology classification

Background: Histological assessment of colorectal cancer (CRC) tissue is a crucial and demanding task for pathologists. Unfortunately, manual annotation by trained specialists is a burdensome operation, which suffers from problems like intra-and inter-pathologist variability. Computational models are revolution-izing the Digital Pathology field, offering reliable and fast approaches for challenges like tissue segmen-tation and classification. With this respect, an important obstacle to overcome consists in stain color variations among different laboratories, which can decrease the performance of classifiers. In this work, we investigated the role of Unpaired Image-to-Image Translation (UI2IT) models for stain color normal-ization in CRC histology and compared to classical normalization techniques for Hematoxylin-Eosin (H&E) images. Methods: Five Deep Learning normalization models based on Generative Adversarial Networks (GANs) belonging to the UI2IT paradigm have been thoroughly compared to realize a robust stain color normal-ization pipeline. To avoid the need for training a style transfer GAN between each pair of data domains, in this paper we introduce the concept of training by exploiting a meta-domain, which contains data coming from a wide variety of laboratories. The proposed framework enables a huge saving in terms of training time, by allowing to train a single image normalization model for a target laboratory. To prove the applicability of the proposed workflow in the clinical practice, we conceived a novel perceptive qual-ity measure, which we defined as Pathologist Perceptive Quality (PPQ). The second stage involved the classification of tissue types in CRC histology, where deep features extracted from Convolutional Neural Networks have been exploited to realize a Computer-Aided Diagnosis system based on a Support Vector Machine (SVM). To prove the reliability of the system on new data, an external validation set composed of N = 15,857 tiles has been collected at IRCCS Istituto Tumori "Giovanni Paolo II". Results: The exploitation of a meta-domain consented to train normalization models that allowed achiev-ing better classification results than normalization models explicitly trained on the source domain. PPQ metric has been found correlated to quality of distributions (Frechet Inception Distance - FID) and to similarity of the transformed image to the original one (Learned Perceptual Image Patch Similarity - LPIPS), thus showing that GAN quality measures introduced in natural image processing tasks can be linked to pathologist evaluation of H&E images. Furthermore, FID has been found correlated to accuracies of the downstream classifiers. The SVM trained on DenseNet201 features allowed to obtain the highest classification results in all configurations. The normalization method based on the fast variant of CUT (Contrastive Unpaired Translation), FastCUT, trained with the meta-domain paradigm, allowed to achieve the best classification result for the downstream task and, correspondingly, showed the highest FID on the classification dataset.