US-12625057-B2 - Multi-spectral digital inline holography for biological particle classification

Abstract

A system and method for characterizing biological particles. A multi-spectral digital inline holographic includes a computing system, a camera and a light source having a. coherent multi-spectral beam of light. Tire light source illuminates a sample having one or more biological particles and the camera captures holograms produced by interference of (i) light from the coherent multi-spectral beam of light that was scattered by the sample with (ii) light from the coherent multi-spectral beam of light that was not scattered, by the sample, the captured holograms including holograms from two or more spectral bands. The computing system applies a machine learning model to the captured holograms to extract features of the biological particles m the sample from the captured holograms.

Inventors

• Jiarong HONG
• Ruichen He

Assignees

• REGENTS OF THE UNIVERSITY OF MINNESOTA

Dates

Publication Date

20260512

Application Date

20220429

Claims (7)

1. 1 . A method for characterizing biological particles, the method comprising: illuminating a sample with a coherent multi-spectral beam of light, the sample including one or more biological particles; capturing holograms defining a spectral response of the sample, the holograms produced by interference of (i) light from the coherent multi-spectral beam of light that was scattered by the sample with (ii) light from the coherent multi-spectral beam of light that was not scattered by the sample, wherein the captured holograms including holograms from two or more spectral bands; and applying a machine learning model to the spectral response defined by the captured holograms to extract features of the biological particles in the sample from the captured holograms.
2. 2 . The method of claim 1 , wherein the features include one or more of biological particle localization, morphology characterization including size and shape, classification of different biological particle types, or results of biochemical analysis of the biological particles.
3. 3 . The method of claim 2 , wherein classifications of different biological particle types including classifications of different species and different strains of the same species.
4. 4 . The method of claim 2 , wherein results include viability and vitality of the biological particles.
5. 5 . The method of claim 1 , wherein illuminating the sample with the coherent multi-spectral beam of light comprises feeding the multi-spectral beam of light into beam combining optics.
6. 6 . The method of claim 1 , wherein applying a machine learning model to the spectral response of the holograms includes applying a trained convolutional neural network (CNN) to the holograms.
7. 7 . The method of claim 1 , wherein applying a machine learning model to the spectral response of the holograms includes applying a trained you-only-look-once (YOLO) model to the holograms.

Description

This application is a national stage entry of International Patent Application No. PCT/US2022/072030, filed Apr. 29, 2022, which claims to and the benefit of U.S. Provisional Patent Application No. 63/201,477, filed Apr. 30, 2021, the entire contents of both applications are incorporated herein by reference. TECHNICAL FIELD This disclosure generally relates to imaging and classification of biological particles. BACKGROUND Characterization of the biochemical properties of biological particles is essential in research and industrial applications. Representative applications include differentiating cancer cells from normal cells in situ, characterizing the lipid content in biofuel production algae, examining the viability and concentration of yeast in the fermentation industry, and analyzing the viability of bacteria suspended in water and air. State-of-art and commercially available techniques for characterizing biochemical properties of biological particles include cytometer, lab culturing and manual counting, mass spectrometry, and quantitative polymerase chain reaction (qPCR). Cytometry uses light scattering from a single cell to determine cell properties including viability and genome size. Cytometry, however, suffers from low throughput and requires complex sample preparation, including fluorescent labeling. Fluorescent labeling may impact cell activities, limiting the ability of cytometry to precisely determine viability of some cells. The lab culturing and manual counting technique usually takes from two days to seven days to culture the cell colonies, which means this technique is not suitable for applications that require fast analysis, such as detecting intrusion of viable biological particles in surgical sites. Similarly, mass spectrometry ionizes the biomolecules (e.g., DNA, RNA, and protein) for identification of bacteria or fungi, while qPCR uses the polymerase chain reaction to amplify target DNA for quantification of gene expression to determine strains and types of biological particles. Both mass spectrometry and qPCR can only be applied to fragmented cells, however, and cannot, therefore, be used to characterize cell viability. The abovementioned techniques usually require expensive setups (cytometers ˜$100,000. qPCR & mass spectrometers>$60,000) and supplies for testing (˜˜100 per PCR test sample). Such limitations hinder the wider application of these techniques in industry settings. SUMMARY In general, the present disclosure describes techniques for high-throughput classification of different types of cells using machine learning and multi-spectral digital inline holographic (DIH) imaging. Cells to be classified are imaged using a digital inline holography technique operating at multiple wavelengths. The holograms acquired at the different wavelengths are then mixed and used to train a neural network to perform cell type classification. In one example approach, the described method has demonstrated a high level of accuracy of cell classification (>90%) for yeast cells and has improved >10% in comparison to methods that use a machine learning network trained by holograms generated via a single wavelength. Previous approaches to biological particle classification suffer from low throughput and low accuracy of classification, often requiring complex optical setups and long sample preparation. Current techniques rely on sophisticated and expensive optical setups (such as quantitative phase imaging and force spectrum microscopy) to capture information from biological particles for further analysis. But those methods require fine-tuning of the setups and are typically more expensive than the techniques described herein. In addition, those solutions capture only a very limited number of biological particles at the same time, which hinders their throughput. In addition, single-wavelength DIH captures only very limited features of biological particles and is not, therefore, accurate. As described herein, the present disclosure describes techniques for using multi-spectral DIH to generate images of biological particles for the classification of different types of biological particles. The present disclosure further describes techniques for mixing holograms generated from different wavelengths of illumination for using the mixed holograms to train a convolutional neural network. Such an approach improves cell classification accuracy in comparison to using the single-wavelength holograms. The proposed techniques may have high throughput with improved accuracy and potentially less sample preparation time compared to incumbent methods (e.g., quantitative phase imaging, force spectrum microscopy, and single wavelength DIU). These techniques classify cell types with high accuracy and high throughput in comparison to the methods described above. They also provide a viable solution for applications requiring fast and accurate detection of different types of particles (e.g., fast cancer detection, surgical