Dice-XMBD: Deep learning-based cell segmentation for imaging mass cytometry

Xu Xiao, Ying Qiao, Yudi Jiao, Na Fu, Wenxian Yang, Liansheng Wang, Rongshan Yu, Jiahuai HanbioRxiv 2021 · DOI

TL;DR

This paper introduces Dice-XMBD, a deep learning model for segmenting individual cells in Imaging Mass Cytometry (IMC) data. It addresses the challenge of accurate cell segmentation in high-dimensional IMC images, which is crucial for biological analysis, by leveraging a U-Net architecture and a novel approach to combine marker information.

What Problem Does This Paper Address?

Imaging Mass Cytometry (IMC) is a powerful technique that allows researchers to simultaneously measure up to 40 different protein markers within a single tissue sample at subcellular resolution. This capability is invaluable for understanding complex biological systems, such as tumor microenvironments. However, a major hurdle in utilizing IMC data is the accurate and efficient segmentation of individual cells. Existing methods often struggle with the high dimensionality and complexity of IMC images, leading to inaccurate cell boundaries or requiring extensive manual annotation, which is both time-consuming and prone to user variability. This bottleneck hinders the widespread adoption and application of IMC in biological research.

Methods

The core of Dice-XMBD is a U-Net architecture, a popular deep learning model for image segmentation. The novelty lies in how it processes the multi-channel IMC data. Instead of treating each protein marker channel independently, Dice-XMBD first merges them into two simplified channels: a "nuclear channel" (containing proteins localized to the nucleus) and a "cell channel" (containing proteins from the cytoplasm and cell membrane). This approach makes the model more robust to variations in specific marker expression across different studies.

To overcome the challenge of obtaining large amounts of manually annotated data, the researchers employed a knowledge distillation technique. They used Ilastik, an interactive image analysis tool, with manual annotations from experts to generate "teacher" probability maps. These maps then served as training labels for the Dice-XMBD model. Finally, the output of the U-Net (pixel probability maps) is fed into CellProfiler, a widely used image analysis software, to generate the final cell segmentation masks using a propagation method.

Datasets

The study utilized four IMC image datasets for training, validation, and testing:

  • BRCA1 and BRCA2: Datasets from breast cancer patients, with 37 and 35 markers respectively.
  • T1D1 and T1D2: Datasets from Type I Diabetes patients, with 35 and 33 markers respectively.

The BRCA1 dataset was used for training and internal validation, while all four datasets were used to evaluate the model's generalization capabilities across different studies and labs.

Key Insights and Ideas

  • Channel Consolidation is Key: The strategy of merging multiple protein channels into just two (nuclear and cell) significantly simplifies the input for the deep learning model, making it more adaptable to diverse IMC datasets without requiring retraining for each new study.
  • Knowledge Distillation Bridges the Gap: Using Ilastik with expert annotations to generate training labels effectively transfers human expertise to the deep learning model, reducing the need for massive, fully annotated datasets.
  • U-Net's Power for High-Dimensional Data: The U-Net architecture proves highly effective in segmenting cells even when dealing with complex, multi-channel IMC data, demonstrating its versatility beyond standard image analysis.
  • Open-Source for Reproducibility: Making the code and datasets publicly available is crucial for advancing the field and ensuring that other researchers can build upon and validate these findings.

Key Results

Dice-XMBD demonstrated superior performance compared to existing methods, including CellProfiler's default algorithms (distance, watershed, propagation) and the Mesmer deep learning model.

  • Pixel-level Accuracy: Achieved a high F1 score of 0.92 and Jaccard index of 0.85 on the BRCA1 dataset.
  • Cell-level Accuracy: Showed excellent performance with a Pearson correlation of 0.998 for cell counts and a mean IOU of 0.815 for matched cell pairs.
  • Generalization: The model performed well across different datasets from various labs, indicating its robustness and applicability to diverse IMC studies without retraining.
  • Downstream Analysis: Protein profiling correlations between Dice-XMBD predictions and ground truth were highest among tested methods, suggesting minimal impact on subsequent biological analyses.

Limitations and Future Work

  • Manual Annotation Still Needed (for Teacher Model): While Dice-XMBD significantly reduces the annotation burden, the initial generation of the "teacher" model still requires some manual annotation using tools like Ilastik.
  • Ignoring Ambiguous Locations: The paper mentions that "ambiguous locations are ignored by Dice-XMBD for segmentation as they contribute little to the segmentation results." This might lead to missing some subtle cellular structures or boundaries in complex regions.
  • Future Work: The authors suggest that the open-source nature of Dice-XMBD will facilitate further development and application in various biomedical studies. Future work could explore optimizing the channel consolidation strategy or further reducing the reliance on initial manual annotations.

Summary updated

Abstract

Abstract Highly multiplexed imaging technology is a powerful tool to facilitate understanding cells composition and interaction in tumor microenvironment at subcellular resolution, which is crucial for both basic research and clinical applications. Imaging mass cytometry (IMC), a multiplex imaging method recently introduced, can measure up to 40 markers simultaneously in one tissue section by using a high-resolution laser with a mass cytometer. However, due to its high resolution and large number of channels, how to process and interpret the image data from IMC remains a key challenge for its further applications. Accurate and reliable single cell segmentation is the first and a critical step to process IMC image data. Unfortunately, existing segmentation pipelines either produce inaccurate cell segmentation results, or require manual annotation which is very time-consuming. Here, we developed Dice-XMBD, a Deep learnIng-based Cell sEgmentation algorithm for tissue multiplexed imaging data. In comparison with other state-of-the-art cell segmentation methods currently used in IMC, Dice-XMBD generates more accurate single cell masks efficiently on IMC images produced with different nuclear, membrane and cytoplasm markers. All codes and datasets are available at https://github.com/xmuyulab/Dice-XMBD .
BibTeX
@misc{xiao2021dice,
  author = {Xu Xiao and Ying Qiao and Yudi Jiao and Na Fu and Wenxian Yang and Liansheng Wang and Rongshan Yu and Jiahuai Han},
  title = {{Dice-XMBD: Deep learning-based cell segmentation for imaging mass cytometry}},
  year = {2021},
  doi = {10.1101/2021.06.05.447183},
  howpublished = {bioRxiv},
}
Figure