TopoUT: Enhancing Cell Segmentation Through Efficient Topological Regularization
Cell segmentation plays a crucial role in biological image analysis, serving as a fundamental step with downstream applications in pathology, clinical diagnosis, and drug discovery. Recent advancements in deep learning, especially the Unet architecture, have shown promising results in medical image segmentation. However, accurately segmenting cell boundaries, especially the nuclei, remains challenging due to diverse cell morphology. This paper presents a novel approach that incorporates topological data analysis methods into deep neural networks for cell segmentation. Our proposed method introduces a pair of topological regularization terms, namely loop penalty and cohesion penalty, which efficiently estimate zero and one-dimensional topological features, respectively. To evaluate the efficacy of our model, we conduct experiments on microscopy images featuring various types of human tissue cells. By capturing rich geometric patterns, our approach outperforms baseline models, offering enhanced accuracy in segmentation.
