Label‐Free Leukemia Monitoring by Computer Vision

Minh Doan, Marian Case, Dino Mašić, Holger Hennig, Claire McQuin, Juan Carlos Caicedo, Shantanu Singh, Allen Goodman, Olaf Wolkenhauer, Huw D. Summers, David Jamieson, Frederik W. van Delft, Andrew Filby, Anne E. Carpenter, Paul Rees, Julie IrvingCytometry 2020 · DOI

TL;DR

This paper introduces a novel "label-free" method for monitoring leukemia in children using deep learning and imaging flow cytometry. By analyzing only bright-field and dark-field images of cells, the system can identify leukemia cells with high accuracy, potentially simplifying and democratizing crucial diagnostic tests.

What Problem Does This Paper Address?

Accurately assessing the presence of residual leukemia cells after treatment is critical for guiding therapy and predicting relapse in children with Acute Lymphoblastic Leukemia (ALL). Current methods, like morphology-based assessment and flow cytometry using multiple antibody markers, are effective but can be labor-intensive, costly, require specialized training and equipment, and sometimes struggle with subtle distinctions or rare cells. This can be particularly challenging in resource-limited settings. The paper aims to develop a faster, cheaper, and more accessible method for this vital monitoring.

Methods

The core of this research lies in applying deep learning, specifically a ResNet50 architecture, to analyze images from imaging flow cytometry (IFC). The key innovation is the label-free approach: instead of relying on fluorescent markers (antibodies) to identify leukemia cells, the model is trained to recognize them using only the bright-field (transmitted light) and dark-field (side scatter equivalent) images of individual cells.

Here's a breakdown:

  • Imaging Flow Cytometry (IFC): This technique captures images of individual cells as they flow through a laser beam, combining the high throughput of flow cytometry with the spatial information of microscopy.
  • Deep Learning (ResNet50): A convolutional neural network architecture known for its effectiveness in image recognition tasks. The model was trained to classify cells into three categories: ALL blasts, normal B lymphocytes, and "other" nucleated cells.
  • Feature Extraction: The model learns to extract relevant morphological features directly from the bright-field and dark-field images, bypassing the need for fluorescent labels.
  • Data Augmentation and Training: Techniques like random undersampling, horizontal/vertical flips, shifts, and rotations were used to create a robust training dataset and prevent overfitting.
  • Comparison: The label-free method's performance was compared against traditional antibody-based flow cytometry and classical machine learning approaches (linear SVM) using morphological features.

Datasets

The study utilized bone marrow samples from children diagnosed with B lineage ALL. These samples were collected at the time of diagnosis ("presentation") and during follow-up treatment phases.

  • Training/Validation Set: Pooled data from 19 samples across 15 patients.
  • Test Sets:
    • Test 1: Manually gated ground-truth populations for leukemic blasts, normal lymphocytes, and other cell types.
  • Test 2 (MRD): Unannotated cells to validate the model's ability to derive a residual disease readout (percentage of leukemic cells).
  • Test 3: Stained/unstained samples with laser-on/laser-off conditions to assess the impact of fluorescence and lasers.
  • Test 4 (Held-out): A completely separate set of patient data, held out until the final verification stage.

The paper does not explicitly mention introducing new benchmark datasets but rather uses clinical samples for training and evaluation.

Key Insights and Ideas

  • Morphology is Rich: The study powerfully demonstrates that even without fluorescent labels, the morphological features captured in bright-field and dark-field images contain sufficient information for accurate leukemia cell identification. This challenges the long-held reliance on specific biomarkers.
  • Deep Learning's Advantage: Deep learning models, particularly ResNet50, are adept at learning complex patterns from raw image data, simplifying the workflow by potentially automating feature extraction and selection, which are often manual and time-consuming in classical methods.
  • Potential for Accessibility: The label-free nature of the method opens doors for simpler, potentially laser-free cytometers, making advanced monitoring more accessible in resource-limited settings.
  • Robustness to Variation: The model showed robustness to technical and patient heterogeneity, suggesting its potential for real-world clinical deployment.

Key Results

  • High Accuracy: The deep learning model achieved an accuracy of >88% in identifying ALL cells using only label-free images on patient samples excluded from training.
  • Label-Free Performance: Even when using only bright-field and dark-field images, the method achieved 90.3% accuracy in leave-one-out testing and 88.6% on held-out patient samples.
  • Comparable to Biomarkers: The label-free method's performance was comparable to using a full panel of four specific antigenic biomarkers and human-curated ALL counts for residual disease assessment.
  • Excellent Sensitivity/Specificity at 25% Cutoff: At the clinically meaningful cutoff of 25% MRD, the label-free method demonstrated 100% sensitivity and specificity.
  • Efficiency: While training took weeks on a GPU, the inference speed for cell images was estimated to be around 375,000 cells/min, suggesting rapid analysis once trained.

Limitations and Future Work

  • Accuracy Gap: While promising, the authors acknowledge that the label-free method, at this stage, does not completely replace morphology or residual disease tests in terms of absolute accuracy. However, it is acceptable for clinically meaningful cutoffs.
  • Clinical Deployment Challenges: Successful clinical deployment would require careful data collection across multiple sites and operators to train robust models for different instruments and protocols.
  • Future Directions:
    • Adaptation to Other Leukemias: The authors suggest the method is feasible for adaptation to other types of leukemia, which would revolutionize residual disease monitoring.
  • Simplified Cytometers: The potential to adapt this to simple, laser-free, hand-held imaging cytometers for automated, point-of-care testing is a significant future prospect.
  • Expanding Training Data: Improving accuracy and robustness by expanding the training dataset is a clear path forward.
  • Diagnosing ALL and Lymphoma: The approach could potentially be used to diagnose ALL and differentiate it from lymphoblastic lymphoma.

Summary updated

Abstract

Acute lymphoblastic leukemia (ALL) is the most common childhood cancer. While there are a number of well-recognized prognostic biomarkers at diagnosis, the most powerful independent prognostic factor is the response of the leukemia to induction chemotherapy (Campana and Pui: Blood 129 (2017) 1913-1918). Given the potential for machine learning to improve precision medicine, we tested its capacity to monitor disease in children undergoing ALL treatment. Diagnostic and on-treatment bone marrow samples were labeled with an ALL-discriminating antibody combination and analyzed by imaging flow cytometry. Ignoring the fluorescent markers and using only features extracted from bright-field and dark-field cell images, a deep learning model was able to identify ALL cells at an accuracy of >88%. This antibody-free, single cell method is cheap, quick, and could be adapted to a simple, laser-free cytometer to allow automated, point-of-care testing to detect slow early responders. Adaptation to other types of leukemia is feasible, which would revolutionize residual disease monitoring. © 2020 The Authors. Cytometry Part A published by Wiley Periodicals, Inc. on behalf of International Society for Advancement of Cytometry.
BibTeX
@inproceedings{doan2020labelfree,
  author = {Minh Doan and Marian Case and Dino Mašić and Holger Hennig and Claire McQuin and Juan Carlos Caicedo and Shantanu Singh and Allen Goodman and Olaf Wolkenhauer and Huw D. Summers and David Jamieson and Frederik W. van Delft and Andrew Filby and Anne E. Carpenter and Paul Rees and Julie Irving},
  title = {{Label‐Free Leukemia Monitoring by Computer Vision}},
  booktitle = {Cytometry},
  year = {2020},
  doi = {10.1002/cyto.a.23987},
}
Figure