Show abstract - PDF - Reviews

Performance of state-of-the-art deep learning methods is often impacted when evaluated on data coming from unseen acquisition settings, hindering their approval by the regulatory agencies and incorporation to the clinic. In recent years, several techniques have been proposed for improving the generalizability of models by using the target data and their corresponding ground truths. Some of those approaches have been adopted in histopathology, however they either focus on pixel-level predictions or simple tile level classification tasks with or without target labels. In this work, we investigate adversarial strategies in weakly supervised learning frameworks in digital pathology domain without access to the target labels, thereby strengthening the generalizability to unlabeled target domains. We evaluate several strategies on Camelyon dataset for metastatic tumor detection tasks and show that some methods can improve the average F1-score over 10% for the target domain.
Hide abstract