Centroid-Aware Gaussian Prompt Learning with Erosion-Guided Accumulator for Robust Semantic Cell Segmentation

Nur Suriza Syazwany, Su Jung Kim, Ju-Hyeon Nam, Sang-Chul Lee

Keywords: cell segmentation, prompt

TL;DR: A novel centroid-guided Gaussian map-based accumulator prompting approach for nuclei segmentation.

Abstract: Accurately segmenting cell images remains challenging due to variations in cell size, shape, and overlapping structures. Existing approaches often struggle with densely packed and overlapping cell regions, leading to inconsistent performance. While recent methods, such as the Segment Anything Model (SAM), have shown promise, they rely heavily on manual prompting, which can be time-consuming and inconsistent for densely packed nuclei datasets. To address these limitations, we propose a novel centroid-guided Gaussian map-based accumulator prompting approach for robust nuclei segmentation. Our method constructs Gaussian maps by accumulating centroids across multiple erosion iterations, capturing the frequency and spatial distribution of nuclei centroids. These maps serve as informative priors to guide the segmentation model, enhancing its ability to localize cell structures while maintaining adaptability to varying cell sizes. By integrating these Gaussian-based prompts into a transformer-based segmentation model, our approach enables refined predictions with improved spatial awareness. We validate our method on two challenging, densely packed datasets, DSB18 and ConSeP, demonstrating robust and superior segmentation performance over state-of-the-art methods.

Submission Number: 18