Nuclei Segmentation in Histopathological Images with Enhanced U-Net3+

Abstract

In the rapidly evolving field of nuclei segmentation, there is an increasing trend towards developing a universal segmentation model capable of delivering top-tier results across diverse datasets. While achieving this is the ultimate goal, we argue that such a model should also outperform dataset-specific specialized models. To this end, we propose a task-specific feature sensitive U-Net model, that sets a baseline standard in segmentation of nuclei in histopathological images. We meticulously select and optimize the underlying U-Net3+ model, using adaptive feature selection to capture both short- and long-range dependencies. Max blur pooling is included to achieve scale and position invariance, while DropBlock is utilized to mitigate overfitting by selectively obscuring feature map regions. Additionally, a Guided Filter Block is employed to delineate fine-grained details in nuclei structures. Furthermore, we apply various data augmentation techniques, along with stain normalization, to reduce inconsistencies and thus resulting in significantly outperforming the state-of-the-art performance and paving the way for precise nuclear segmentation essential for cancer diagnosis and possible treatment strategies.