Color deconvolution of histological images

Abstract

Cancer is one of the main health problems worldwide. When dealing with this disease, the earlier the diagnosis is produced, the better the prognostic of the patient can be. For that reason, in the last years, digital pathology has been emerging as main diagnosis technique and, in consequence, the number of artificial intelligence research studies about it has also increased exponentially. The Institut Català de la Salut (ICS) has created the project DigiPatICS, in collaboration with the Universitat Politècnica de Catalunya (UPC), with the aim of developing computer vision models to process histopathological images and helping to speed up their analysis by the pathologist to produce faster but precise diagnosis. Nevertheless, the variations between the images from the same type but from different hospitals make the development of a global model for all the hospitals an impossible task. In this project, we develop a model that corrects one of this variations: the color variations of the HER2 stain between the hospitals of Vall d Hebron and Bellvitge. For that reason, we trained different models of domain adaptation between the images of these hospitals in order to be able to apply the same nuclei segmentation and classification model to both domains. The results show that we have been able to obtain a good classification using a Vall d Hebron style trained model over the transformed Bellvitge images, but we think that these results can be improved.