Motivation

Blood delivers a variety of insights into health conditions of an organism, especially in the field of in-vitro diagnostics. Therefore, hematological analysis such as complete blood count (CBC) represent a high share of laboratory tests in the health care sector. Despite the availability of largely automated analyzers, the gold standard for the routine diagnosis of hematological disorders is the tedious Giemsa stained blood smear [1] which suffers from inter-observer variations. Further drawbacks of current approaches are a high effort for sample preparation, material expenses, a long processing time or missing clinical relevance [2]. To overcome the mentioned issues and specially to reduce time-consuming manual intervention, the CellFACE project was created. Flow cytometry in combination with digital holographic microscopy (DHM) [3] enables researchers to record 3D images of each single cell, while sustaining a high throughput stream of blood cells. The overall goal is to establish this label- and reagent-free approach as a platform technology for automated blood cell diagnosis [4]. The best methodologies to analyze the recorded phase images are subject to research. Therefore, this work will investigate classical and modern techniques in the field of computer vision and machine learning on their suitability for interpreting the images and the contained information.

State of the Art

Today many diseases like Leukemia or other myeloproliferative neoplasms, which directly affect the composition of the blood or the morphology of blood cells, are analyzed with the mentioned blood smear technique. Here, the blood sample is spread to a thin film and then stained chemically [5]. A microscopist then hast to classify and count the cells manually to get a diagnosis, which can be time consuming and is prone the human errors [1]. Modern hematology analyzers have partly automatized this process. Such devices can be used to perform a five-part differential of white blood cells which may indicate an unhealthy distribution of leukocyte sub-types. Further analysis, e.g. on a morphological level, is not possible with these devices due to the inflexibility of this method [1]. An even higher amount of detail is obtained by fluorescence-activated cell sorting (FACS). Here, cells can be marked with fluorescent antibodies and therefore clearly identified and counted. Though this method only works for the major cell types but not for degenerated cell populations and is additionally very cost intensive [2].

Statistical significance and a high certainty play major roles in a clinical surrounding. By using a microfluidic channel in combination with a phase microscope, it is possible to capture images of a vast number of cells in only a few minutes which contain a high degree of detail. Based on in-line differential holography, it is even possible to perceive information of the internal structures e.g. the nucleus of the cell. The Center for Translational Cancer Research (TranslaTUM) hosts a specifically manufactured DHM setup which can take hundred phase images per second and provides a high-quality resolution over the whole image plane. A detailed description of the microscope, the alignment process and the sample handling are given in [6]. In Figure 2 you can see a reconstructed phase image from DHM which contains the three-dimensional relief of each blood cell. Since this approach is label-free, there are no markers attached to cells which would allow an easy differentiation of the distinct cell types. So, the classification and characterization of each cell only from the captured topology becomes a challenging task [7]. To get an visual impression Figure 3 contains three representative phase images of different leukocytes. Several approaches were proposed, which directly work on the morphological cell features to enable label-free differentiation of white blood cells. For three-part differentials of leukocytes Li et al. [8] used Linear Discriminant Analysis (LDA) to select features according to their importance and subsequently trained a Support Vector Machine (SVM) for classification. Vercruysse et al. [9] obtained similar results, but still could not reach the performance of modern haematology analyzers. Amir et al. [10] used t-Distributed Stochastic Nearest Neighbor Embedding in their tool viSNE, to reduce the high dimensional parameter space to two dimensions while preserving similar clusters. Roitshtain [11] and Ugele et al. [6] showed the advantages of Principal Component Analysis (PCA) for dimensionality reduction of the feature space. These analysis tools indicate a broad clinical applicability to overcome the limitations of today’s standard methods for the routine diagnosis of hematological disorders and make the proposed goal seem achievable [6, 12].


Problem Statement 

Blood is a complex substance where many diseases leave their traces in diverse ways. Hence, a modern hematology analyzer should be capable to detect a broad range of abnormalities and changes in the morphology and the composition of blood. To achieve a statistically significant result, even if the observable variations are infrequent, a high cell throughput must be sustained. The system required to work stand alone especially in remote areas, since many neglected diseases are present in
regions where no broadband internet uplink is available. The execution time should not take longer then a few minutes in order to obtain a quick result in urgent cases and to simplify the process for the doctor and the patient. Even if a system can provide the described services, it must fulfill three other major criteria to achieve clinical relevance. First, the approach has to satisfy certain quality standards be allowed in a diagnosis process. It must also be explainable. Blackbox solutions are widely unaccepted in the treatment of humans or animals. Also, the usability of the developed algorithms and systems plays an important role, as this is essential for a program to be used in a productive way. All in all, an End-to-End approach would be favorable which is able to assist physicians in diagnostics and treatment based on a single blood sample of a patient.

Scientific and Technical Goals

About 480,000 people per year in Germany suffer from cancer, which makes this disease the second highest cause of death. Therefore, the improvement of diagnostic and therapeutic options is an important task of cancer research. At TranslaTUM, scientists from the fields of medicine, engineering and natural sciences are working together to improve the chances of cancer patients being cured. Hence, the overarching idea of this project is to quickly transfer new findings from research into clinical practice, a process known as translation.

Various diseases leave traces on cells, in particular blood cells, in different ways. Modern haematology analysers should therefore be able to detect a large number of abnormalities and changes in the morphological structure of the cells and their composition of the cells or blood. Since the observed irregularities can sometimes be very rare, the cells have to be examined in high throughput in order to obtain a statistical result. Such a system is also expected to be autonomous, especially for use in remote areas where the so-called neglected or poverty-related diseases still occur. The duration of the analysis should be short not only in urgent cases, so that the process can be kept as simple as possible for the physician and the patient.

Even if a system is able to provide mentioned services, it must still fulfill at least three additional criteria in order to achieve clinical relevance. First and foremost, it must meet high safety and quality standards. In addition, the used approaches must be explainable, since black box solutions are largely unacceptable in the treatment of humans and animals. Last but not least, the user-friendliness of the proposed programs and devices plays a decisive role in ensuring that they can be used productively. Ultimately, technological progress is aimed at an end-to-end approach that allows a single patient sample to support physicians in the diagnosis and treatment of a variety of diseases.

Requirements

Since science and technology are still far away from such a comprehensive and flexible haematology analyzer, this project is a step in the right direction. As described, DHM in combination with a microfluidic channel enables a high throughput of cells and at the same time a relatively detailed view of the composition of the cell and their interior.

Diseases such as leukemia can not only destroy the natural balance of white blood cells, but also cause the release of unfinished precursor cells. Thus, not only the normal subgroups must be distinguished and counted but also the special cases, which cannot be carried out by conventional devices. 

This problem also exists in the differentiation and purification of cell cultures, which is hardly possible for the human eye without sample preparation.

In general, image-based detection, classification and analysis of cellular objects offers great potential in the fight against intra- and extracellular parasites. In the case of drug development for schistosomiasis, the viability of worm-like parasites should be monitored. The same applies to the detection and treatment of malaria, in which plasmodia infect red blood cells.

All the above mentioned use cases place the following demands on a predictive artificial intelligence solution:

  • Robust and efficient information retrieval about the digitized cells
  • Differentiation and classification of different cell types in suspension on the basis of holographic data
  • Analysis of different cell types for different clinical questions: cell cultures, blood cells, intra- and extracellular parasites, ...

The prototypical implementation of the CellPhaser software in this project will take this into account and will be evaluated according to its performance in these areas.

Process steps

The data basis for blood cell analysis is formed by 3D holograms and the phase images reconstructed from them (cf. Figure 2), which allow insights into the nature of the cell and its interior. The interpretation of these data requires the following steps:

Segmentation

Firstly, it must be determined which image areas contain relevant objects or structures. Then the relevant areas can be prepared for processing by sophisticated machine learning algorithms.

Labeling:

It is generally considered that modern learning algorithms require a large amount of training data. The task of building a suitable ground truth must be supported by the machine.

Feature extraction:

The main challenges for feature extraction are the different object groups (leukocytes, erythrocytes, tissue cells and parasites), the strongly varying object size and the problem type itself (e.g. malaria or leukemia detection). It is important to have features that are universally and robustly applicable, but do not lose their interpretability.

Classification:

Depending on the question, e.g. the assignment to a blood cell subgroup, the detection of malaria pathogens or the determination of the viability of parasites, different statements have to be made. The decisions made in this step must be precise and comprehensible in order to use the software effectively.

Representation:

The acceptance of a machine judgement is based on the transparency of the generating methodology and the linkability with existing forms of presentation.

Methodology

The implementation of the project is characterized by two aspects. These are the user-centered design to achieve clinical relevance and acceptance as well as the finding of suitable algorithms for robust information retrieval from cell data.


References

  1. J. J. Barcia, “The giemsa stain: Its history and applications,” International Journal of Surgical Pathology, vol. 15, no. 3, pp. 292–296, 2007.
  2. A. Filby, “Sample preparation for flow cytometry benefits from some lateral thinking,” Cytometry Part A: the journal of the International Society for Analytical Cytology, vol. 89, pp. 1054–1056, 2016.
  3. Y. Jo, H. Cho, S. Yun Lee, G. Choi, G. Kim, H.-s. Min, and Y. Park, “Quantitative phase imaging and artificial intelligence: A review,” IEEE Journal of Selected Topics in Quantum Electronics, vol. 25, pp. 1–14, 2018.
  4. Z. El-Schich, S. Kamlund, B. Janicke, K. Alm, and A. Wingren, “Holography: The usefulness of digital holographic microscopy for clinical diagnostics,” in Holographic Materials and Optical Systems, pp. 319– 333, Intech, 2017.
  5. S. Prieto, A. Powless, J. W. Boice, S. Sharma, and T. Muldoon, “Proflavine hemisulfate as a fluorescent contrast agent for point-of-care cytology,” PLoS ONE, vol. 10, p. e0125598, 2015.
  6. M. Ugele, M. Weniger, M. Stanzel, M. Bassler, S. W. Krause, O. Friedrich, O. Hayden, and L. Richter, “Labelfree high-throughput leukemia detection by holographic microscopy,” Advanced Science, vol. 1800761, 2018.
  7. M. Rinehart, H. Sang Park, and A. Wax, “Influence of defocus on quantitative analysis of microscopic objects and individual cells with digital holography,” Biomedical Optics Express, vol. 6, no. 6, pp. 2067–2075, 2015.
  8. Y. Li, B. Cornelis, A. Dusa, G. Vanmeerbeeck, D. Vercruysse, E. Sohn, K. Blaszkiewicz, D. Prodanov, P. Schelkens, and L. Lagae, “Accurate label-free 3-part leukocyte recognition with single cell lens-free imaging flow cytometry,” Computers in biology and medicine, vol. 96, pp. 147–156, 2018.
  9. D. Vercruysse, A. Dusa, R. Stahl, G. Vanmeerbeeck, K. de Wijs, C. Liu, D. Prodanov, P. Peumans, and L. Lagae, “Three-part differential of unlabeled leukocytes with a compact lens-free imaging flow cytometer,” Nature biotechnology, vol. 15, pp. 1123–1132, 2013.
  10. E.-A. D. Amir, K. L. Davis, M. Tadmor, E. Simonds, J. Levine, S. Bendall, D. Shenfeld, S. Krishnaswamy, G. Nolan, and D. Pe’er, “visne enables visualization of high dimensional single-cell data and reveals phenotypic heterogeneity of leukemia,” Lab on a Chip, vol. 31, no. 6, pp. 545–52, 2013.
  11. D. Roitshtain, L. Wolbromsky, E. Bal, H. Greenspan, L. L. Satterwhite, and N. T. Shaked, “Quantitative phase microscopy spatial signatures of cancer cells,” Cytometry Part A: the journal of the International Society for Analytical Cytology, vol. 91, pp. 482–493, 2017.
  12. J. Yoon, Y. Jo, M.-h. Kim, K. Kim, S. Lee, S.-J. Kang, and Y. Park, “Identification of non-activated lymphocytes using three-dimensional refractive index tomography and machine learning,” Scientific reports, vol. 7, no. 1, p. 6654, 2017.
  • Keine Stichwörter