Label-free detection of rare circulating tumor cells by image analysis and machine learning

Shen Wang, Yuyuan Zhou, Xiaochen Qin, Suresh Nair, Xiaolei Huang, Yaling LiuScientific 2020 · DOI

TL;DR

This paper introduces a novel, label-free method for detecting rare circulating tumor cells (CTCs) in patient blood using bright-field microscopy images and deep learning. By leveraging a Convolutional Neural Network (CNN), the approach significantly simplifies CTC identification, moving away from complex and potentially cell-damaging fluorescent labeling techniques. This has the potential to make CTC analysis more accessible, faster, and less invasive for cancer diagnosis and monitoring.

What Problem Does This Paper Address?

Detecting and counting circulating tumor cells (CTCs) in patient blood is crucial for early cancer diagnosis, monitoring disease recurrence, and predicting patient survival. However, CTCs are extremely rare, existing at concentrations of only a few cells per 10 billion blood cells. This scarcity, coupled with the heterogeneity of CTCs, makes their identification a significant challenge. Current methods often rely on fluorescent labeling and immunocytochemistry, which are labor-intensive, require specialized equipment and expertise, and can negatively impact cell viability, hindering further analysis like culturing or drug screening. The paper aims to overcome these limitations by developing a simpler, faster, and label-free method for CTC detection.

Methods

The core of this work is a deep learning approach using a Convolutional Neural Network (CNN) for image classification. The process begins with collecting bright-field microscopy images of patient blood samples.

  1. Image Pre-processing: Raw images undergo several steps to isolate individual cells. This includes applying Otsu's thresholding for edge detection, flood-fill operations, and morphological opening to segment cells and remove artifacts. A watershed transformation is used for final segmentation. Crucially, the researchers developed a software toolbox for this process. After segmentation, individual cell images are cropped.
  2. Data Augmentation: To address the challenge of limited data, especially for rare CTCs, the researchers employed data augmentation techniques. This involved applying traditional affine and elastic transformations (rotation, shear, reflection) to create synthetic variations of existing images, thereby increasing the diversity and size of the training dataset.
  3. CNN Architecture: A pre-trained ResNet-50 architecture is utilized for feature extraction. This model is fine-tuned with a fully connected layer, a dropout layer, and a softmax activation function for binary classification (CTC vs. White Blood Cell - WBC). The use of pre-trained weights from ImageNet helps reduce the need for extensive training data.
  4. Training and Evaluation: The model is trained using a five-fold cross-validation strategy to ensure robust performance evaluation. This involves splitting the data into training, validation, and testing sets, with augmentation applied to the training and validation data.

Datasets

The study utilized two main types of datasets:

  • Cultured Cell Lines: A dataset comprising 1745 single-cell images (436 cultured cells and 1309 WBCs) was used for initial training and testing.
  • Patient Blood Samples: A dataset of 95 single-cell images from metastatic renal cell carcinoma (mRCC) patients was used for training and testing. This dataset was augmented to create a larger training set of 1000 CTCs and 1000 WBCs for each training experiment.

The paper does not explicitly mention introducing a new benchmark dataset but rather uses existing patient samples and cultured cell lines for training and evaluation.

Key Insights and Ideas

  • Label-Free Detection is Feasible: The most significant insight is that accurate CTC detection can be achieved using only standard bright-field microscopy images, eliminating the need for fluorescent labels. This drastically simplifies the experimental setup and reduces costs.
  • Data Augmentation for Rare Events: The successful application of data augmentation techniques is key to overcoming the challenge of limited sample sizes for rare CTCs, demonstrating its effectiveness in improving model robustness.
  • Pre-trained CNNs for Efficiency: Leveraging pre-trained CNN architectures like ResNet-50 significantly reduces the training time and data requirements, making the approach more practical.
  • t-SNE for Visualization: The use of t-SNE to visualize the high-dimensional data in a 2D space clearly shows the separability of CTCs and WBCs, providing confidence in the model's ability to learn discriminative features.

Key Results

  • High Accuracy: The model achieved high accuracy in classifying cells:
  • Cultured Cells: 97.5% accuracy.
  • Patient Blood Samples: 88.4% overall accuracy with a five-fold cross-validation.
  • Robust Performance: The F-score for patient blood samples was 0.89, indicating a good balance between precision and recall.
  • AUC: The averaged Area Under the Curve (AUC) for the ROC curve was 0.923, signifying excellent discriminative ability.
  • Simplified Workflow: The method requires minimal data pre-processing and an easy experimental setup, making it faster and simpler than traditional fluorescent methods.

Limitations and Future Work

  • Data Size Limitation: The authors acknowledge that the limited size of the original patient blood sample dataset is a challenge. While data augmentation helps, they suggest that further improvements can be made with more data.
  • Accuracy Gap: The accuracy on patient blood samples (88.4%) is lower than on cultured cell lines (97.5%). This is attributed to the inherent differences and potential noise in real patient samples compared to controlled cell cultures.
  • Misclassification Causes: Misclassifications are hypothesized to be due to noise, manual labeling errors, or inherent overlap in cell characteristics.
  • Future Directions:
    • Improved Model: Further improvement of the machine learning model with more data.
  • Integration with Sorting: Combining this label-free detection with sorting techniques (e.g., acoustic sorting) for simultaneous isolation and characterization, potentially improving cell viability.
  • Clinical Application: Integrating CTC count information with molecular characterization for personalized therapy selection, especially for challenging cases like RCC where traditional methods are less effective.

Summary updated

Abstract

Detection and characterization of rare circulating tumor cells (CTCs) in patients' blood is important for the diagnosis and monitoring of cancer. The traditional way of counting CTCs via fluorescent images requires a series of tedious experimental procedures and often impacts the viability of cells. Here we present a method for label-free detection of CTCs from patient blood samples, by taking advantage of data analysis of bright field microscopy images. The approach uses the convolutional neural network, a powerful image classification and machine learning algorithm to perform label-free classification of cells detected in microscopic images of patient blood samples containing white blood cells and CTCs. It requires minimal data pre-processing and has an easy experimental setup. Through our experiments, we show that our method can achieve high accuracy on the identification of rare CTCs without the need for advanced devices or expert users, thus providing a faster and simpler way for counting and identifying CTCs. With more data becoming available in the future, the machine learning model can be further improved and can serve as an accurate and easy-to-use tool for CTC analysis.
BibTeX
@inproceedings{wang2020label,
  author = {Shen Wang and Yuyuan Zhou and Xiaochen Qin and Suresh Nair and Xiaolei Huang and Yaling Liu},
  title = {{Label-free detection of rare circulating tumor cells by image analysis and machine learning}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-69056-1},
}
Figure