MoE-NuSeg: Enhancing nuclei segmentation in histology images with a two-stage Mixture of Experts network

Abstract

Accurate nuclei segmentation is essential for extracting quantitative information from histology images to support disease diagnosis and treatment decisions. However, precise segmentation is challenging due to the presence of clustered nuclei, varied morphologies, and the need to capture global spatial correlations. While state-of-the-art Transformer-based models employ tri-decoder architectures to decouple the segmentation task into nuclei, edges, and cluster edges segmentation, their complexity and long inference times hinder clinical integration. To address this, we introduce MoE-NuSeg, a novel Mixture of Experts (MoE) network that consolidates the tri-decoder into a single decoder. MoE-NuSeg employs three specialized experts for nuclei segmentation, edge delineation, and cluster edge detection, thereby mirroring the functionality of tri-decoders while surpassing their performance and reducing parameters by sharing attention heads. We propose a two-stage training strategy: the first stage independently trains the three experts, and the second stage fine-tunes their interactions to dynamically allocate the contributions of each expert using a learnable attention-based gating network. Evaluations across three datasets demonstrate that MoE-NuSeg outperforms the state-of-the-art methods, achieving an average increase of 0.99% in Dice coefficient, 1.14% in IoU and 0.92% in F1 Score, while reducing parameters by 30.1% and FLOPs by 40.2%. The code is available at https://github.com/deep-geo/MoE-NuSeg.

Keywords

• Nuclei segmentation
• Histology image
• Swin transformer
• Mixture of Experts (moEs)