A Lightweight-Based Architecture for Nuclei Segmentation in Histopathology Images

Abstract

The segmentation of histopathology images is crucial in diagnosing and assessing various cell types, including cancer cells, normal cells, and bacteria. However, analyzing these images manually is time-consuming and can lead to subjective results. To address this, automated segmentation methods based on deep learning and AI techniques have been developed. Supervised learning methods with encoder-decoder architectures, such as U-Net, have shown promising results in medical image segmentation. However, these models often struggle with the imbalance of information in heterogeneous medical images. To overcome this, new models like Unet++ and ResUnet were introduced, combining skip connections and dense connections to balance semantic information. Transformer architectures, such as ViT, have emerged as a new paradigm in computer vision and have shown excellent results when combined with Unet in the medical field. However, these models are computationally expensive and not feasible for mobile devices. To address this, LiReNet, a lightweight U-Net++-based architecture, was proposed to reduce storage and training costs while maintaining good learning and prediction capabilities. Our model LiReNet achieves superior results compared to other state of the arts, with the Dice scores up to 92.3% on the Data Science Bowl 2018 datset, and 88.3% on the glaS dataset. Code is available at https://github.com/PVTHust/LiReNet.