AI-ENHANCED SINGLE-CELL DATA ANALYTICS FOR LABEL-FREE CYTOMETRY

Description:
The synergistic convergence of microfluidics and AI is expected to have a transformative role in single-cell analysis. Combining microfluidics with AI promises to enable highly automated, high-throughput, and multiparametric studies, boosting advancements in diagnostics and personalized medicine.

Within the realm of microfluidic single-cell approaches, label-free image flow cytometry and impedance flow cytometry are presently undergoing a revolution in how their data are processed by leveraging innovative AI-based solutions. In this workshop, after a comprehensive overview of this lively research field, exemplary state-of-the-art approaches will be discussed, as well as open issues and perspectives. Some specific challenges will be considered in detail by means of examples developed in widely used programming environments (Phyton, Matlab). Topics include e.g.: handling of so-called "unknown unknowns" in the automatic classification of cells, robustness with respect to different chips/laboratories, and related concepts of data normalization and data augmentation.

The workshop will include three lectures.

Lecture 1 (Prof. Yaxiaer Yalikun)
Due to their capability to provide rich information and enable high-throughput analysis, AI-empowered imaging and impedance flow cytometry are increasingly becoming potential replacements for standard flow cytometry as label-free techniques.

This introductory lecture will offer a comprehensive overview of strategies for implementing AI approaches in the imaging and impedance flow cytometry. We will first begin by outlining the commonly employed setups for acquiring data (i.e., images or signals) from large quantities of biological cells. Subsequently, we will delve into the processes necessary for extracting features from the acquired image or signal data. Specifically, we will systematically summarize how these features can be utilized for cell phenotyping through the application of machine learning algorithms.

Additionally, we will discuss existing challenges and provide insights into the future perspectives of intelligent imaging and impedance flow cytometry. Moreover, we will introduce and discuss some of the state-of-the-art research advances in this domain, illustrated with examples.

Lecture 2 (Dr. David Dannhauser)
The second lecture will focus on the issue of "unknown unknowns", with specific reference to unknown cell class prediction from single-cell images. In other words, we don't know what we don't know, which often results in classification models that are confident about their choice of classifying a never-before-seen cell class, but they actually perform wrong. To overcome this issue, a neural network needs to define first an in-distribution of known cell classes to later distinguish out-of-distribution unknown cell classes. In fact, a good classification model should not only produce accurate predictions of known data, but also detect unknown cell classes and reject or classify them in a new class. Accordingly, we will first define step-by-step a neural network to characterize different known cell classes with high prediction accuracy (in-distribution). Next, we will implement a new network structure based on auxiliary open-set risk, which enable us to detect unknown cell classes (out-of-distribution) from known ones. We will look together on the prediction performance of the network and which model parameters influence most the prediction outcome of unknown versus known cell classes using Python language.

Lecture 3 (Prof. Federica Caselli)
The third lecture will focus on AI-approaches for microfluidic impedance cytometry. Specifically, recurrent and convolutional neural networks will be designed in Matlab to predict single-cell properties from raw impedance data streams (which are time-domain current signals, at possibly multiple frequencies). The challenges associated to the definition of target values for supervised training and opportunities offered by synthetic data streams will be considered. Moreover, examples of supervised and unsupervised machine learning approaches for cell population analysis based on impedance features (e.g., amplitude, phase, opacity) will be discussed. Some AI-workflows tailored to specific applications (i.e., fast dielectric spectroscopy, coincidence resolution, and multimodal analysis) will also be presented. Finally, analogies with other domains characterized by time-domain signals (e.g., signals from nanopores) will be highlighted.

Participants Will Need the Following:
For Lecture 2, attendees willing to participate actively should bring a laptop with Python.