Scholarlink

<p>Imaging mass cytometry (IMC) is a high-dimensional imaging technology that allows the capture and quantification of up to fifty biomarkers at sub-cellular resolution. Each IMC dimension (‘channel’) corresponds to antibodies coupled to metal-tagged antigens and captures the imaging representation of biomarkers for each cell. Conventional IMC analysis relies on manually segmenting the cells across all IMC channels, which is a time-consuming process and prone to human error. Recent advances in computerized IMC analysis, using deep learning techniques such as convolutional neural networks (CNNs), enable automated segmentation by quantifying cellular structures through a data-driven approach. However, existing CNN-based methods concatenate and integrate IMC image channels at an early stage in the learning process, which potentially limiting the model's ability to retain meaningful correlations among different channels for training. In addition, the cluttered nature and inhomogeneous textures of cellular structures may further complicate the training process, making it even more difficult for CNNs to accurately segment the cell boundaries. In this study, we propose a Stacked Channel Learning Network (SCLN), a two-phase CNN-based method for IMC cell segmentation. SCLN uses a channel embedding approach with two training phases: Phase I generates imaging masks for cellular structures using a pre-trained Res-U-Net model, while Phase II refines the segmentation by incorporating these masks as an additional channel alongside the existing biomarker channels. In addition, SCLN preserves the image representations derived independently from each channel, enabling more detailed channel-level feature analysis. The proposed method was evaluated on the widely used IMC breast cancer METABRIC dataset. The experimental results show that our SCLN achieved a Dice coefficient score of 91.45%, which outperformed the existing segmentation methods by ~ 10%, especially for the challenging studies e.g., cluttered cells and cells with inhomogeneous textures. Our codes can be found at <a href="https://github.com/kongnet-djd/SCLN">https://github.com/kongnet-djd/SCLN</a>.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SCLN: Stacked Channel Learning Network for Imaging Mass Cytometry (IMC) cell segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年5月2日

查看全文
  • Jiadi Dong,
  • Tian Xia,
  • Pradeeba Sridar,
  • Mark Read,
  • Thomas Ashhurst,
  • Lei Bi,
  • Jinman Kim

摘要

Imaging mass cytometry (IMC) is a high-dimensional imaging technology that allows the capture and quantification of up to fifty biomarkers at sub-cellular resolution. Each IMC dimension (‘channel’) corresponds to antibodies coupled to metal-tagged antigens and captures the imaging representation of biomarkers for each cell. Conventional IMC analysis relies on manually segmenting the cells across all IMC channels, which is a time-consuming process and prone to human error. Recent advances in computerized IMC analysis, using deep learning techniques such as convolutional neural networks (CNNs), enable automated segmentation by quantifying cellular structures through a data-driven approach. However, existing CNN-based methods concatenate and integrate IMC image channels at an early stage in the learning process, which potentially limiting the model's ability to retain meaningful correlations among different channels for training. In addition, the cluttered nature and inhomogeneous textures of cellular structures may further complicate the training process, making it even more difficult for CNNs to accurately segment the cell boundaries. In this study, we propose a Stacked Channel Learning Network (SCLN), a two-phase CNN-based method for IMC cell segmentation. SCLN uses a channel embedding approach with two training phases: Phase I generates imaging masks for cellular structures using a pre-trained Res-U-Net model, while Phase II refines the segmentation by incorporating these masks as an additional channel alongside the existing biomarker channels. In addition, SCLN preserves the image representations derived independently from each channel, enabling more detailed channel-level feature analysis. The proposed method was evaluated on the widely used IMC breast cancer METABRIC dataset. The experimental results show that our SCLN achieved a Dice coefficient score of 91.45%, which outperformed the existing segmentation methods by ~ 10%, especially for the challenging studies e.g., cluttered cells and cells with inhomogeneous textures. Our codes can be found at https://github.com/kongnet-djd/SCLN.

Graphical Abstract