Domain Adaptation Strategies Based on Color Normalisation for Histological Image Classification

Digital pathology has been recently powered by digital scanners at high magnification. Several open histological image datasets are available nowadays for developing machine learning models. However, these models, often trained on a supervised setting, are prone to significant drops in performance when evaluated on images that display distinct visual characteristics from those observed on the training/evaluation sets. This limits the deployment of machine learning models in clinical settings with histological images manually acquired at low magnification under a camera-mounted microscope. In this study, we simulated the acquisition of whole slide images (WSI) at low magnification by navigating the slides as done with a camera-mounted microscope, adding a degree of blur and using a stitching method to recover a full slide image at low magnification. We trained five EfficientNet models using varying training budgets of the PANDA challenge dataset (20%, 25%, 35%, 50%, and 100%), consisting of 10,616 prostate cancer images acquired by digital scanners. Then the performance of these models was evaluated using WSI acquired at low magnification via the mentioned simulation process. Our models achieved Quadratic Weighted Kappa (QWK) metrics of 0.65, 0.76, 0.84, 0.85, and 0.93, respectively. Afterward we used Reinhard's and Vahadane's color normalisation methods to better adapt the images used to train the models before applying the machine learning classifiers. This process improved the QWK metrics to 0.88, 0.88, 0.88, 0.93 and 0.96 respectively, thus suggesting that color normalisation is a useful strategy for the domain adaptation of manually acquired low magnification images.