ACANet: Adaptive Contour Aware Nucleus Segmentation Network

Abstract

Accurate segmentation and analysis of nucleus are crucial for cancer diagnosis and prognosis. For nuclear images with blurred boundaries and low color contrast, existing methods often rely on additional contour-prediction branches and final feature fusion modules. However, these approaches encounter challenges such as the loss of contour information during the downsampling process and the introduction of redundant information and noisy artifacts during feature fusion. To overcome these limitations, we propose an Adaptive Contour Aware Nucleus Segmentation Network (ACANet), which integrates Sobel convolution and Fourier transform to enhance the contours of the nuclei. Additionally, a novel multi-scale adaptive feature learning module is employed to extract multi-scale discriminative features of nuclei. Finally, a loss function guided by Grad-CAM is proposed to achieve better segmentation performance and enhance model interpretability. Experimental results demonstrate that the proposed architecture achieves superior accuracy in nucleus segmentation tasks.