SpectrAi: How deep learning enables more rapid Raman microscopy
SpectrAi is a Python/Matlab open-source package that enables faster Raman imaging. The principle of improving acquisition speed up to 40-140x with SpectrAI is based on the published work https://www.impopen.com/jsi-abstract/I11_a7. In SpectrAI we speed up Raman spectroscopy by giving the microscope access to prior chemical information about a specific sample through deep learning. This allow us to reduce integration time significantly, specifically:
To practically implement SpectrAI for an application the following workflow is required (for instance, to realize rapid imaging of tissue sections):
5. Train the Raman data set using SpectrAI for spectral and spatial reconstruction. Assess when to stop training using the validation data set and benchmark the performance using the (independent) test set.
6. The developed SpectrAI model can now be used for inference to realize rapid imaging by using low integration time and lower sampling resolution.
Notes: Samples must be representative and span the experimental variability for the application. Once a model has been developed it can only be assumed it will work for a specific application (e.g., for a specific tissue type). However, transfer learning enables direct and rapid extension to other applications with minimal effort. For this reason SpectrAI can be used for deployment of higher throughput applications (e.g., pathology, tissue engineering monitoring, cell sorting or rapid in vivo diagnostics etc.) while the application for explorative discovery driven Raman spectroscopy remains limited.
A comprehensive introduction to SpectrAI and DeepeR can be found in the following publications:
https://www.impopen.com/jsi-abstract/I11_a7
pubs.acs.org/doi/10.1021/acs.analchem.1c02178.
Citation
If you find this resource helpful in your work, please cite the following work:
Conor C. Horgan and Mads S. Bergholt, "spectrai: a deep learning framework for spectral data.", arXiv preprint arXiv:2108.07595 (2021).
- Training of the Raman microscope to predict high signal to noise (SNR) Raman spectra from low SNR (low integration time) Raman spectra.
- Training of the Raman microscope for computational super resolution imaging allows it to predict neighboring pixel values thereby reducing spatial resolution and imaging time.
- Combining the above two principles synergistically improves Raman imaging reconstruction significantly.
To practically implement SpectrAI for an application the following workflow is required (for instance, to realize rapid imaging of tissue sections):
- Keep measuring parameters constant (microscope objective, laser wavelength, tissue section thickness, laser power etc).
- Ensure calibration remains constant (regularly monitor the wavelength calibration and system response).
- Acquire a large training data/validation/test data set (e.g., 80 Raman images) for a specific application with following characteristics:
- 20 Raman images (e.g., 200x200 pixels) with low SNR (e.g., 0.1 sec). The same 20 Raman images should be measured with high SNR (e.g., 2 sec)
- 20 Raman images with low resolution (e.g., 200x200) and the same 20 Raman images with higher resolution (e.g., 400x400)
5. Train the Raman data set using SpectrAI for spectral and spatial reconstruction. Assess when to stop training using the validation data set and benchmark the performance using the (independent) test set.
6. The developed SpectrAI model can now be used for inference to realize rapid imaging by using low integration time and lower sampling resolution.
Notes: Samples must be representative and span the experimental variability for the application. Once a model has been developed it can only be assumed it will work for a specific application (e.g., for a specific tissue type). However, transfer learning enables direct and rapid extension to other applications with minimal effort. For this reason SpectrAI can be used for deployment of higher throughput applications (e.g., pathology, tissue engineering monitoring, cell sorting or rapid in vivo diagnostics etc.) while the application for explorative discovery driven Raman spectroscopy remains limited.
A comprehensive introduction to SpectrAI and DeepeR can be found in the following publications:
https://www.impopen.com/jsi-abstract/I11_a7
pubs.acs.org/doi/10.1021/acs.analchem.1c02178.
Citation
If you find this resource helpful in your work, please cite the following work:
Conor C. Horgan and Mads S. Bergholt, "spectrai: a deep learning framework for spectral data.", arXiv preprint arXiv:2108.07595 (2021).