Generating BlobCell Label from Weak Annotations for Precise Cell Segmentation

Cell segmentation in histopathological image analysis is critical for identifying cancer cells and predicting disease severity. However, manual cell labeling is time-consuming. Many experiments speed up the generation of cell data by annotating central cell points and classes, generating cell segmentation labels with a fixed radius. However, the accuracy of this method depends on the specified given radius, which is problematic due to the variety of cell sizes and shapes, including elongated ovals and linear shapes. The use of a fixed radius is considered in-accurate and unreasonable. To address this, we propose BlobCell labeling, which uses blob extraction for accurate cell labeling based on central coordinates, resulting in a +9.02% improvement in the dice score. Furthermore, to improve cell detection from cell segmentation results such as the proposed challenge baseline [1], we designed a new network architecture that utilizes BlobCell information within the Injection model structure, we achieved a significant performance improvement of +12.11% in mF1 score on the test set.