Show abstract - Show schedule - Proceedings - PDF - Reviews

To develop deep learning-based models for automatic analysis of histopathology whole slide images (WSIs), the atomic entities to be directly processed are often the smaller patches cropped from WSIs as it is not always possible to feed a whole WSI to a model given its enormous size. However, a trained model tends to relate the slide-specific characteristics to diagnosis results because a large number of patches cropped from the same WSI will share common slide features and thus have strong correlations between them, resulting in deteriorated generalization capability of the trained model. Current approaches to alleviate this issue include data pre-processing (stain normalization or color augmentation) and adversarial learning, both of which introduce extra complications in computations. Alternatively, we propose to reduce the impact of this issue by introducing a new regularization term to the standard loss function to reduce the correlation of the patches from the same WSI. It is intuitive and easy-to-implement and introduces comparably smaller computation overhead compared to existing approaches. Experimental results prove that the proposed regularization term is able to enhance the generalization capability of learning models and consequently to achieve better performance. The code is available in: \url{https://github.com/hrzhang1123/SlideCorrelationReduction}.
Hide abstract

Thursday 8th July
G1-9 (short): Interpretability and Explainable AI - 16:45 - 17:30 (UTC+2)
Hide schedule