Generalizing to Unseen Domains in Histopathology: A Multi-Task Learning Approach

Deep neural networks achieved remarkable performance in automated histopathology image classification. These models however are trained based on the i.i.d. assumption between training and test data, which leads to significant performance drops under domain shift. Existing methods for addressing this phenomenon fail in improving by significant margin the performance of the baseline model on domains unseen at training time. To address this limitation, we propose to use multi-task learning (MTL) with tasks complementary and semantically related to the main task, to aid the model to learn representations of the data that are more stable to variations across domains. We evaluated our approach on the task of atypical mitosis recognition on two public datasets, by exploiting as auxiliary tasks two pixel-level classification tasks, namely binary segmentation and multi-class segmentation of the mitosis to classify. The conducted experiments demonstrated that our approach allows to improve model generalization capability on data from domains unseen at training time in statistically significant manner, outperforming existing state-of-the-art (SOTA) methods for addressing domain shift.