Look in different views: Multi-scheme regression guided cell instance segmentation

Cell instance segmentation constitutes a formidable challenge that involves simultaneous detection and segmentation for individual cell within an image. A multitude of instance segmentation techniques have been applied to address this challenge. Despite notable strides, existing approaches grapple with certain limitations, particularly the accurate localization of cell center points. In scenarios involving densely packed cells, the potential for conflating multiple cells exists, while the identification of elongated cells runs the risk of erroneously recognizing them as multiple entities. To address these issues, we introduce a novel cell instance segmentation network founded on the principles of multi-scheme regression guidance. Our proposed network leverages the versatility of multi-scheme regression guidance, enabling a comprehensive analysis of each cell from multiple perspectives. Specifically, a Gaussian guidance attention mechanism is meticulously crafted to harness Gaussian labels, directing the network's attention effectively. Subsequently, a point-regression module is integrated to facilitate the precise regression of cell centers. Finally, the outputs from the aforementioned modules were utilized to further bootstrap the instance segmentation process, thereby enhancing its performance. The multi-scheme regression guidance strategy empowers us to harness the distinguishing attributes of different cell regions, notably the central cell region. Thorough experimentation across five benchmark datasets underscores the potency of our approach. Our model outperforms the previous state-of-the-art (SOTA) method by margins of 1.2% and 2.3% (AP50) on DSB2018 and CA2.5, respectively. Similarly, our method surpasses the prior SOTA by 1.5% (AJI) on MoNuSeg and 0.3% lower on CPM17. Notably, our method exhibits a substantial performance leap, boasting a 3.0% higher AP50 on SCIS. These encouraging results collectively demonstrate the effectiveness of our approach. Furthermore, our approach's interpretability is substantiated through visualization and comprehensive analysis. The source code is available at https://github.com/cv516Buaa/MSRNet.