Correlative multimodal chemical imaging via machine learning
Inventors
OVCHINNIKOVA, Olga S. • Ievlev, Anton V. • Lorenz, Matthias • Borodinov, Nikolay • King, Steven T.
Assignees
UT Battelle LLC • University of Tennessee Research Foundation
Publication Number
US-12057304-B2
Publication Date
2024-08-06
Expiration Date
2041-06-03
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
Machine learning approach can combine mass spectral imaging (MSI) techniques, one with low spatial resolution but intact molecular spectra and the other with nanometer spatial resolution but fragmented molecular signatures, to predict molecular MSI spectra with submicron spatial resolution. The machine learning approach can perform transformations on the spectral image data of the two MSI techniques to reduce dimensionality, and using a correlation technique, find relationships between the transformed spectral image data. The determined relationships can be used to generate MSI spectra of desired resolution.
Core Innovation
The invention provides a machine learning approach for combining spectral imaging data from two mass spectrometry imaging (MSI) techniques: one technique delivering low spatial resolution but intact molecular spectra and another providing higher (nanometer) spatial resolution with fragmented molecular signatures. By performing transformations to reduce the dimensionality of both spectral imaging datasets and then applying a correlation technique, the invention determines relationships between the transformed data. The learned relationships are used to generate molecular MSI spectra with submicron spatial resolution that would otherwise be practically infeasible with direct measurement.
The main problem addressed is the inability of existing single imaging approaches to directly achieve both high spatial resolution and intact component spectra for chemical imaging. While high spatial resolution MSI techniques exist, their combination of spatial and spectral limitations makes direct high-resolution molecular imaging impractical. By leveraging correlative multimodal imaging and machine learning, the invention enables the reconstruction of high-spatial resolution molecular images that combine the advantages of both modalities.
The system uses a hardware instrument to acquire two kinds of spectral image data cubes representing different spatial and spectral characteristics of a sample. A processor co-registers and transforms these data cubes into component abundance maps using techniques like non-negative matrix factorization (NMF), then applies spatial correlation analysis (such as canonical correlation analysis) to link the abundance maps from each modality. Once trained, the model can generate high-resolution molecular spectral images from new high-resolution fragmented spectral data, effectively predicting intact molecular spectra at spatial resolutions previously possible only in fragmented data.
Claims Coverage
There are two independent claims covering the inventive features: a system and a computer-implemented method for generating high-resolution spectral images by fusing multimodal MSI data via machine learning.
System for multimodal chemical imaging via machine learning
The system includes: - An instrument for acquiring two types of spectral image data cubes, with different spatial resolutions and spectral properties (first-type with lower spatial resolution/intact molecular spectra, second-type with higher spatial resolution/fragmented signatures). - A processor configured to train a machine learning model by: 1. Co-registering and matching the spatial resolutions of the two types of data cubes. 2. Transforming each data cube into abundance maps of respective components (intact molecules or fragments) using dimensionality reduction methods. 3. Spatially correlating these abundance maps and storing the learned relationships as model parameters. - The processor also: - Receives new high-resolution fragmented data. - Generates a spectral image, represented as a first-type spectral data cube at the high (second-type) spatial resolution, by applying the trained machine learning model.
Computer-implemented method for generating high-resolution spectral images
The method consists of: - Receiving a first-type spectral-data cube (lower spatial resolution, intact molecular composition) and a second-type spectral-data cube (higher spatial resolution, molecular fragments) from a sample. - Training a machine learning model that: 1. Co-registers and spatially matches the two data cubes. 2. Transforms each into abundance maps of spatially coexisting components. 3. Spatially correlates the abundance maps and saves these relationships as learned parameters. - Receiving a new data cube of second-type (high resolution, fragmented spectra). - Generating a spectral image of first-type (intact molecules) at second-type (higher) spatial resolution by applying the trained model.
The independent claims provide comprehensive coverage for a machine learning-based system and method that integrates and correlates multimodal MSI data, using dimensionality reduction and abundance map correlation, to achieve high spatial resolution chemical imaging with reconstructed molecular spectra.
Stated Advantages
Enables automated prediction of molecular mass spectra at sub-micrometer spatial resolution, combining advantages of different MSI modalities.
Overcomes practical infeasibility of direct high spatial resolution measurements while preserving intact molecular spectra.
Allows reconstruction of spectral data with improved spatial resolution based on co-registered multimodal imaging.
Can reveal fine details of molecular distributions within complex systems.
Improves signal-to-noise ratio and preserves spatial details in generated data through dimensionality reduction techniques.
Documented Applications
Subcellular imaging of biological systems to reveal fine molecular distributions.
Chemical visualization in biological tissues, functional nanomaterials, and polymer blends.
Incorporation of additional imaging modalities, such as Raman and Fourier-transform infrared spectroscopy (FTIR) imaging, following similar mixing rules.
Interested in licensing this patent?