Portal de Periódicos da CAPES

Diagnostic Potential of Imaging Flow Cytometry

2018; Elsevier BV; Volume: 36; Issue: 7 Linguagem: Inglês

10.1016/j.tibtech.2017.12.008

0167-9430

Minh Doan, Ivan A. Vorobjev, Paul Rees, Andrew Filby, Olaf Wolkenhauer, Anne E. Goldfeld, Judy Lieberman, Natasha S. Barteneva, Anne E. Carpenter, Holger Hennig,

Advanced Fluorescence Microscopy Techniques

Imaging flow cytometry (IFC) captures multichannel images of hundreds of thousands of single cells within minutes. IFC is seeing a paradigm shift from low- to high-information-content analysis, driven partly by deep learning algorithms. We predict a wealth of applications with potential translation into clinical practice. Imaging flow cytometry (IFC) captures multichannel images of hundreds of thousands of single cells within minutes. IFC is seeing a paradigm shift from low- to high-information-content analysis, driven partly by deep learning algorithms. We predict a wealth of applications with potential translation into clinical practice. Imaging flow cytometry (IFC) combines the high-throughput, multiparameter capabilities of conventional flow cytometry with morphological and spatial information, all at single-cell resolution. Multichannel digital images of hundreds of thousands of individual cells can be captured within minutes (Figure 1), and include several fluorescence channels as well as bright field (transmitted light) and dark field (scattered light). The throughput of IFC means that it is especially well suited to the analysis of rare cell types such as circulating tumor cells (which are cancer cells that escaped from a primary tumor and circulate in the bloodstream) [1Ogle L.F. et al.Imagestream detection and characterisation of circulating tumour cells – a liquid biopsy for hepatocellular carcinoma?.J. Hepatol. 2016; 65: 305-313Abstract Full Text Full Text PDF PubMed Scopus (75) Google Scholar] and transition states, such as cell cycle phases (mitosis) [2Blasi T. et al.Label-free cell cycle analysis for high-throughput imaging flow cytometry.Nat. Commun. 2016; 710256Crossref PubMed Scopus (175) Google Scholar]. By extracting information from these digital images (Figure 1, middle panel), IFC can quantify multiple properties of constituents of interest (including proteins, nucleic acids, glycolipids) in multiple subcellular compartments (nucleus, mitochondria, etc.). The rich information makes IFC ideal for high-content analysis, as well as machine learning, raising the possibility to profile complex cell phenotypes, identify rare cells and transition states, and, importantly, discover useful targets for disease diagnosis, personalized medicine, and drug development. Here, we discuss significant recent developments in the IFC field and a perspective on where IFC could be adopted as a diagnostic tool in clinical practice. Conventional (zero-spatial-resolution) flow cytometry, that is, without imaging, saw its first commercial instruments in the 1970s. Thanks to continuous standardization and improvement, flow cytometry is now routinely used as a diagnostic instrument for health disorders, especially hematologic diseases [3Betters D.M. Use of flow cytometry in clinical practice.J. Adv. Pract. Oncol. 2015; 6: 435-440PubMed Google Scholar]. Although conventional flow cytometry is considered high throughput because it analyzes up to 100 000 cells per second, it is considered to be low in information content because typically only a single feature (integrated intensity) is measured per fluorescence marker. The current trend is to increase the number of parameters that can be simultaneously measured by developing instruments with more lasers and detectors in combination with new fluorochromes that can be used in concert with one another. Mass cytometry can measure in excess of 40 markers simultaneously using antibodies tagged with rare earth metals [4Giesen C. et al.Highly multiplexed imaging of tumor tissues with subcellular resolution by mass cytometry.Nat. Methods. 2014; 11: 417-422Crossref PubMed Scopus (986) Google Scholar]. This platform significantly increases the number of parameters measured beyond what is currently achievable with conventional flow cytometry and has driven the adoption of machine learning techniques when analyzing such multidimensional data, however, it is still limited to intensity-based features. Equipped with 20×, 40×, or 60× objectives and up to two charge-coupled device cameras, IFC allows thousands of morphological and spatial properties to be measured for each individual cell. These include bright field, dark field, and up to ten fluorescent channels (Figure 1) [5Han Y. et al.Review: imaging technologies for flow cytometry.Lab Chip. 2016; 16: 4639-4647Crossref PubMed Google Scholar]. Similar to its flow cytometry-based siblings, IFC is well-suited to image nonadherent or dissociated cells, key for many clinical applications such as analyses of bodily fluids like blood, whose structures can be distorted (smeared) by placement onto a slide. IFC suffers from a 'content gap': although the images are high in information content, containing rich morphological and spatial information (even in a single bright-field channel), data analyses often have low information content, that is, analyses are based on only very few selected features, which are often manually identified by applying binary gates on cell populations of interest. These approaches are highly subjective, require significant user interaction, and only utilize a few morphological features instead of the hundreds that are inherently present in the data, for example, only using cell size to denote cell growth when there are likely tens or hundreds of other features that could describe and resolve the biological differences in a more reliable and powerful way. Such analyses make use of far less information than is present in the images. Yet, even with these limitations, some IFC applications are already heading toward the clinic [6Barteneva N.S. Vorobjev I.A. Imaging Flow Cytometry: Methods and Protocols. Springer, 2016Crossref Google Scholar], such as the diagnostic assessment of acute leukemia [7Grimwade L.F. et al.Applications of imaging flow cytometry in the diagnostic assessment of acute leukaemia.Methods. 2017; 112: 39-45Crossref PubMed Scopus (38) Google Scholar]. IFC data are now beginning to benefit from significant advancements in high-content analysis from the world of high-throughput microscopy [8Zanella F. et al.High content screening: seeing is believing.Trends Biotechnol. 2010; 28: 237-245Abstract Full Text Full Text PDF PubMed Scopus (322) Google Scholar]. High-content analysis might unveil disease states hidden to the eye of even specialists, leading to entirely new diagnostic capabilities. In image-based profiling, once single cells are identified and segmented, thousands of quantitative metrics can be extracted, including shape, intensity, texture, and object relationships. The general strategy is to 'measure everything first, then ask questions later'. A morphological profile is created, which is like a fingerprint of each cell. Then the similarities (correlations) between profiles can be compared to define cell subpopulations or identify disease-specific phenotypes [9Caicedo J.C. et al.Data-analysis strategies for image-based cell profiling.Nat. Methods. 2017; 14: 849-863Crossref PubMed Scopus (301) Google Scholar]. This profiling strategy might be particularly useful when applying IFC for clinical phenotyping, where the cell changes present in a chronic disease often spread throughout a spectrum of phenotypes, from pluripotent precursors to nascent immature cells to terminally differentiated cells, each with distinct morphologies and associated functions. However, advanced IFC analysis pipelines often involve multistep workflows challenging to those who are not computational experts. Momentum to enable a broader group of biomedical researchers and clinicians to carry out complex analyses is growing; for example, a user-friendly open-source high-content IFC protocol based on machine learning is now available [2Blasi T. et al.Label-free cell cycle analysis for high-throughput imaging flow cytometry.Nat. Commun. 2016; 710256Crossref PubMed Scopus (175) Google Scholar, 10Hennig H. et al.An open-source solution for advanced imaging flow cytometry data analysis using machine learning.Methods. 2017; 112: 201-210Crossref PubMed Scopus (52) Google Scholar]. Deep learning is revolutionizing computer vision across many domains, including computational biology [11Angermueller C. et al.Deep learning for computational biology.Mol. Syst. Biol. 2016; 12: 878Crossref PubMed Scopus (794) Google Scholar]. In particular, deep convolutional neural networks have proven to be very powerful (Box 1). The nature of data obtained using IFC is perfectly suited for deep learning for several reasons. Deep learning requires a large number of examples to train the network and IFC can quickly produce millions of single-cell images. In addition, deep learning operates at the pixel level and does not depend on prerequisite preprocessing or object segmentation, which is often prone to errors. Deep learning has shown success in identifying colon cancer cells [5Han Y. et al.Review: imaging technologies for flow cytometry.Lab Chip. 2016; 16: 4639-4647Crossref PubMed Google Scholar] and reconstructing cell cycle and disease progression [12Friedrich M.J. Going with the flow: the promise and challenge of liquid biopsies.JAMA. 2017; 318: 1095-1097Crossref PubMed Scopus (17) Google Scholar]. However, major challenges for applying deep learning to IFC in a clinical setting are the development of user-friendly workflows, satisfying heavy computational requirements, evaluation of diagnostic accuracy, and approval for clinical use. The Food and Drug Administration (FDA) approved the first clinical deep learning application for health care in 2017 (www.forbes.com/sites/bernardmarr/2017/01/20/first-fda-approval-for-clinical-cloud-based-deep-learning-in-healthcare).Box 1Image Processing and Deep LearningBioimaging research has recently made great strides thanks to deep learning, a subclass of machine learning. A deep learning network consists of an input layer, a number of hidden layers, and an output layer. The larger the number of hidden layers, the deeper the network. Among several deep learning methods, convolutional neural networks have shown the most impressive results in object detection [11Angermueller C. et al.Deep learning for computational biology.Mol. Syst. Biol. 2016; 12: 878Crossref PubMed Scopus (794) Google Scholar].Convolutional neural networks operate on the raw pixels of each image where the network learns to enhance complex abstractions in images, such as eyes, faces, or in the case of IFC, a cellular phenotype or signature of disease, while suppressing irrelevant information in the images (see Figure I). By rastering a window over the input image, the feature maps are created (by a mathematical operation called 'convolution'). The feature maps represent how the input image looks when seen through the convolutional filter. Pooling refers to downsampling the feature maps to reduce the number of pixels while keeping the relevant information [11Angermueller C. et al.Deep learning for computational biology.Mol. Syst. Biol. 2016; 12: 878Crossref PubMed Scopus (794) Google Scholar]. The output layer is, the classification probability which assigns a class to each input image with a certain probability, for example, Class 1 and 2 could correspond to leukemic and normal cells, respectively.In contrast to deep learning, in conventional image analysis, the images are first preprocessed, then cellular objects are identified (segmented) by analysis software, followed by the extraction of hundreds of human-engineered features per channel and object (middle panel in Figure 1), including shape, intensity, and texture [8Zanella F. et al.High content screening: seeing is believing.Trends Biotechnol. 2010; 28: 237-245Abstract Full Text Full Text PDF PubMed Scopus (322) Google Scholar, 9Caicedo J.C. et al.Data-analysis strategies for image-based cell profiling.Nat. Methods. 2017; 14: 849-863Crossref PubMed Scopus (301) Google Scholar]. While classical machine learning techniques rely on human-engineered features, deep learning filters can have a much higher level of abstraction and complexity.Recent open-source frameworks, such as TensorFlow, Keras, PyTorch and CAFFE, provide convenient and effective means to adopt powerful deep learning architectures into bioimaging research. With a wide range of practical applications emerging in various industries, we expect that such frameworks will become increasingly stable and robust, and thereby suitable for clinical applications. Bioimaging research has recently made great strides thanks to deep learning, a subclass of machine learning. A deep learning network consists of an input layer, a number of hidden layers, and an output layer. The larger the number of hidden layers, the deeper the network. Among several deep learning methods, convolutional neural networks have shown the most impressive results in object detection [11Angermueller C. et al.Deep learning for computational biology.Mol. Syst. Biol. 2016; 12: 878Crossref PubMed Scopus (794) Google Scholar]. Convolutional neural networks operate on the raw pixels of each image where the network learns to enhance complex abstractions in images, such as eyes, faces, or in the case of IFC, a cellular phenotype or signature of disease, while suppressing irrelevant information in the images (see Figure I). By rastering a window over the input image, the feature maps are created (by a mathematical operation called 'convolution'). The feature maps represent how the input image looks when seen through the convolutional filter. Pooling refers to downsampling the feature maps to reduce the number of pixels while keeping the relevant information [11Angermueller C. et al.Deep learning for computational biology.Mol. Syst. Biol. 2016; 12: 878Crossref PubMed Scopus (794) Google Scholar]. The output layer is, the classification probability which assigns a class to each input image with a certain probability, for example, Class 1 and 2 could correspond to leukemic and normal cells, respectively. In contrast to deep learning, in conventional image analysis, the images are first preprocessed, then cellular objects are identified (segmented) by analysis software, followed by the extraction of hundreds of human-engineered features per channel and object (middle panel in Figure 1), including shape, intensity, and texture [8Zanella F. et al.High content screening: seeing is believing.Trends Biotechnol. 2010; 28: 237-245Abstract Full Text Full Text PDF PubMed Scopus (322) Google Scholar, 9Caicedo J.C. et al.Data-analysis strategies for image-based cell profiling.Nat. Methods. 2017; 14: 849-863Crossref PubMed Scopus (301) Google Scholar]. While classical machine learning techniques rely on human-engineered features, deep learning filters can have a much higher level of abstraction and complexity. Recent open-source frameworks, such as TensorFlow, Keras, PyTorch and CAFFE, provide convenient and effective means to adopt powerful deep learning architectures into bioimaging research. With a wide range of practical applications emerging in various industries, we expect that such frameworks will become increasingly stable and robust, and thereby suitable for clinical applications. Many potential clinical uses of IFC are conceivable, for instance a differential diagnosis of acute and chronic lymphocytic/myeloid leukemia, possibly using fewer biomarkers, or ideally unstained and unmanipulated blood samples. Reducing the number of required, descriptive biomarkers would greatly simplify the laboratory sample preparation and help preserve the intact 'nativeness' of samples, which are often fragile in hematological diseases. Recently, IFC was shown to deliver integrated leukemia diagnostics in one test [7Grimwade L.F. et al.Applications of imaging flow cytometry in the diagnostic assessment of acute leukaemia.Methods. 2017; 112: 39-45Crossref PubMed Scopus (38) Google Scholar]. Another potential clinical use of IFC is to analyze bodily fluids for rare cells, for example, the typically small number of leukemic cells that remain in the patient during treatment, known as minimal residual disease (Figure 1, right panel) [6Barteneva N.S. Vorobjev I.A. Imaging Flow Cytometry: Methods and Protocols. Springer, 2016Crossref Google Scholar]. There is substantial interest in liquid biopsy, the analysis of circulating tumor cells, which are extremely rare [12Friedrich M.J. Going with the flow: the promise and challenge of liquid biopsies.JAMA. 2017; 318: 1095-1097Crossref PubMed Scopus (17) Google Scholar]. Liquid biopsy might detect cancer at an early stage, circulating metastatic or drug-resistant neoplastic cells, clotting abnormalities of platelet microparticles, or fetal abnormalities. A major advantage of the liquid biopsy is that it can be carried out in a simple, noninvasive way. One reason it took the liquid biopsy so long to develop is that circulating tumor cells are found in low concentrations in the bloodstream [12Friedrich M.J. Going with the flow: the promise and challenge of liquid biopsies.JAMA. 2017; 318: 1095-1097Crossref PubMed Scopus (17) Google Scholar]. IFC in combination with machine learning has the potential to identify a single tumor cell out of millions of cells with unprecedented accuracy. Blood specimens can then be sampled in serial fashion (Figure 1, right panel), providing a more comprehensive monitoring of a patient's cancer than can be obtained through traditional methods, as was demonstrated for hepatocellular carcinoma [1Ogle L.F. et al.Imagestream detection and characterisation of circulating tumour cells – a liquid biopsy for hepatocellular carcinoma?.J. Hepatol. 2016; 65: 305-313Abstract Full Text Full Text PDF PubMed Scopus (75) Google Scholar]. Liquid biopsies could also be analyzed by imaging technologies other than IFC: Recently, multiplex protein detection on circulating tumor cells using imaging mass cytometry has been demonstrated [13Gerdtsson E. et al.Multiplex protein detection on circulating tumor cells from liquid biopsies using imaging mass cytometry.Converg. Sci. Phys. Oncol. 2018; 4: 015002Crossref Google Scholar]. Despite this promise, IFC is currently primarily used in research rather than clinical practice. We see the data analysis as the primary hurdle: it is often prone to variation, manual tuning, and interpretation. These issues might be overcome with machine learning approaches. As well, there is a need for standardization of IFC, which should include standard operating procedures and standardized quality control of hardware performance. Although a common practice for conventional flow cytometry, this has not yet been implemented as such in IFC. User-friendly, robust, and standardized workflows that can facilitate machine learning, especially deep learning, will accelerate the paradigm shift from low- to high-content analysis in IFC. Furthermore, cloud computing can overcome the computational infrastructure hurdles. These developments are key for practical IFC applications to reach the clinic, fueling the applicability of IFC as a diagnostic, prognostic, and therapeutic tool. This work was supported in part by the US National Science Foundation/UK Biotechnology and Biological Sciences Research Council under a joint grant NSF DBI 1458626 and BB/N005163 (A.E.C. and P.R.).