mSphere of Influence: the Rise of Artificial Intelligence in Infection Biology

Artur YakimovichmSphere 2019 · DOI

TL;DR

This commentary highlights how artificial intelligence (AI), specifically deep learning (DL) and convolutional neural networks (CNNs), is revolutionizing infection biology by enabling rapid, accurate, and label-free analysis of microscopic pathogens. Two key papers demonstrate AI's power in distinguishing bacterial spores and counting colonies, showcasing its potential to automate complex tasks and accelerate discovery in the field.

What Problem Does This Paper Address?

Modern microscopy has made it easier than ever to visualize pathogens and host-pathogen interactions. However, the sheer volume of data generated by these advanced techniques necessitates automated and unbiased quantification and analysis. Traditional methods are often time-consuming and require specialized reagents. This commentary addresses the challenge of efficiently extracting meaningful information from complex biological images, particularly in the context of infection biology, where accurate identification and enumeration of pathogens are crucial.

Methods

The commentary focuses on two papers that leverage AI for image analysis in infection biology.

  • Holographic Deep Learning for Rapid Optical Screening of Anthrax Spores (Jo et al.): This work employs a DL algorithm, specifically a Convolutional Neural Network (CNN) named HoloConvNet, inspired by the mammalian visual cortex. CNNs learn hierarchical features from images through multiple layers of operations. The novelty here lies in applying this to label-free holographic microscopy data. Holographic microscopy captures the refractive index distribution of spores, providing a unique signature without the need for staining or labeling.
  • Bacterial Colony Counting with Convolutional Neural Networks in Digital Microbiology Imaging (Ferrari et al.): This paper utilizes CNNs for automated bacterial colony counting from conventional photographic images. The approach involves training a CNN on a large dataset of segmented colony images.

Datasets

  • Jo et al.: The study used label-free holographic microscopy images of Bacillus anthracis spores and four related species (Bacillus thuringiensis, Bacillus cereus, Bacillus atrophaeus, and Bacillus subtilis). The exact size of this dataset is not specified, but the focus was on distinguishing individual spores.
  • Ferrari et al.: This work utilized a large dataset of conventional photographic images of bacterial colonies from 28,500 blood agar plates. This represents a significant benchmark for colony counting.

Key Insights and Ideas

  • AI's Versatility Across Scales: AI algorithms, particularly CNNs, can effectively tackle image analysis problems across vastly different scales, from individual bacterial spores to entire colonies.
  • Label-Free Imaging Synergy: Combining label-free imaging techniques (like holographic microscopy) with AI offers a powerful approach. It reduces the need for costly and time-consuming labeling reagents while still achieving high accuracy.
  • Transforming Data Analysis: AI has the potential to fundamentally change how data is analyzed in infection biology, moving from manual, tedious tasks to automated, high-throughput discovery.
  • "Data-Hungriness" vs. Accuracy: While CNNs often require large datasets (the "data-hungriness" problem), the accuracy achieved, even with relatively "shallow" networks, is unprecedented and suggests significant future potential.
  • Accelerating Discovery: By automating routine tasks like quantification and classification, AI can free up researchers' time to focus on higher-level scientific questions and accelerate the pace of discovery.

Key Results

  • Jo et al.: The HoloConvNet model achieved greater than 96% accuracy in distinguishing B. anthracis spores from other related species, leveraging the refractive index distribution captured by holographic microscopy.
  • Ferrari et al.: The CNN-assisted bacterial colony counting achieved 92% accuracy on a large dataset of images from 28,500 blood agar plates, demonstrating improved detection and separation of colonies.

Limitations and Future Work

  • Data Requirements: The "data-hungriness" of CNNs remains a challenge, requiring substantial annotated datasets for training.
  • Algorithm Complexity: While the Ferrari et al. paper used a "relatively shallow" CNN, the commentary implies that deeper, more complex networks could further improve accuracy with the same training data.
  • Future Directions: The commentary strongly suggests that AI is poised to usher in a "new era for infection biology, the era of machines." Future work will likely involve applying these AI techniques to a wider range of pathogens and host-pathogen interactions, further refining algorithms, and integrating AI into routine laboratory workflows for accelerated discovery. The author also mentions their own work employing a comparable pathogen-centric view for Toxoplasma gondii host-pathogen interactions.

Summary updated

Abstract

Artur Yakimovich works in the field of computational virology and applies machine learning algorithms to study host-pathogen interactions. In this mSphere of Influence article, he reflects on two papers "Holographic Deep Learning for Rapid Optical Screening of Anthrax Spores" by Jo et al. (Y. Jo, S. Park, J. Jung, J. Yoon, et al., Sci Adv 3:e1700606, 2017, https://doi.org/10.1126/sciadv.1700606) and "Bacterial Colony Counting with Convolutional Neural Networks in Digital Microbiology Imaging" by Ferrari and colleagues (A. Ferrari, S. Lombardi, and A. Signoroni, Pattern Recognition 61:629-640, 2017, https://doi.org/10.1016/j.patcog.2016.07.016). Here he discusses how these papers made an impact on him by showcasing that artificial intelligence algorithms can be equally applicable to both classical infection biology techniques and cutting-edge label-free imaging of pathogens.
BibTeX
@inproceedings{yakimovich2019msphere,
  author = {Artur Yakimovich},
  title = {{mSphere of Influence: the Rise of Artificial Intelligence in Infection Biology}},
  booktitle = {mSphere},
  year = {2019},
  doi = {10.1128/msphere.00315-19},
}
Figure