Dynamic Hard Task-Guided Meta-Transfer Learning for Pathological Cell Segmentation

Cell segmentation is a key task in digital pathology, providing important information for cancer prognosis and precision medicine. However, it is challenging to construct accurate and efficient segmentation models due to the scarcity of labeled data, staining differences, multi-organ variants, and cell overlap. In this paper, we propose an MTL-Segmentation model that combines Meta-Transfer Learning (MTL) and Dynamic Hard Task batch strategy (DHT), aiming to improve the feature extraction ability and generalization performance of the model. Experimental results show that the model performs well with limited data and significantly outperforms existing methods.