Poster R8 - Molecule Detection with Stimulated Raman Scattering and SVM spectral classification

Introduction: Stimulated Raman scattering (SRS) microscopy has been leveraged in biomedical imaging to identify the concentration of certain molecular bonds within a sample, with hyperspectral imaging capabilities allowing for a pixel's Raman spectra to be captured across space. This Raman spectra contains peaks that correspond to specific rovibrational modes of molecular bonds and can be helpful in determining the bonds present within that pixel. However, there still remains a reliable method for detecting the molecular composition within a certain pixel due to the prevalence of multiple bond peaks within the Raman spectrum. Support vector machines (SVMs) are machine learning tools that employs unsupervised learning to separate data with multidimensional hyperplanes for both linear and nonlinear classification. By representing the hyperspectral image as a series of 1-dimensional vectors, we can train a support vector machine on the spontaneous Raman spectrum of pure biomolecule samples to classify their presence in a 2-dimensional space with SRS. Using the SVM hyperplane to define a metric to be used as a measurement for probability we are able to present a probability distribution for 26 different biomolecules in different tissues from C57BL/6 mice, specifically focusing on the mouse brain and liver.

Materials and

Methods: The tissues utilized in this experiment are from C57BL/6 mice pups, 4 different animals. Several organs were dissected and extracted, which were then embedded in 4% agar and sliced into 120 um sections via a CompressTome (Precisionary Instruments). Specifically, the mouse pup brain and liver were sliced and mounted onto a glass slide with 10um of PBS and imaged utilizing our SRS system. The SRS system utilizes an upright laser scanning microscope with synchronized tunable pump beam (720-990 nm, 5-6 ps pulse width, and 80 MHz repetition rate) and Stokes beam (1032 nm,6 ps pulse width, and 80 MHz repetition rate) from a picoEmerald system (Applied Physics & Electronics). The beams are collected by a high NA condenser and high O.D. shortpass filter (Thorlabs) to isolate the pump beam, which could terminated, filtered, and demodulated by a lock-in amplifier (HF2LI, Zurich instruments). Each hyperspectral image was acquired with a pixel dwell time of 100us and a time constant of 60us. The images were 512x512x65 images collected from 2700-3100 cm-1.
The SVM algorithm was written on Jupyter Notebook based mainly on the Scikit-Learn module utilizing a one-to-one classifier to train 26 different biomolecule spontaneous Raman spectra collected by a confocal Raman microscope (XploRA PLUS, Horiba). The pixel spectra from the HSI images were preprocessed to remove background signal and noise. The SVM algorithm was then used to classify the processed spectra into the 26 different labels/biomolecules.

Results, Conclusions, and Discussions: WIth the results from the SVM classification algorithm, we were able to determine the heterogeneity in expression of these 26 different molecules for the different mice organs. Specifically, we see a closer expression of cholesterol in the mouse hippocampus and omega-3 lipid subtypes in the mouse lateral lobes. This is expressed as number of pixels predicted by the trained SVM model to correspond to one of the different molecules. We then utilize mean-squared-error (MSE) as a metric to determine the accuracy between predicted spectra and the reference spontaneous Raman spectrum. Using MSE, we see that the error for the different classes in brain is 0.055 on average and 0.069 for liver, with some biomolecules reaching MSE less than 0.01. Using the SVM algorithm, we were able to classify pixel spectra from hyperspectral imaging with SRS to spontaneous Raman spectra of pure samples from several common biomolecules. Due to the nature of the SVM boundary decision/hyperplane, a distance from hyperplane would be indicative of how likely it is that the spectra would be classified into a specific label. This metric can be used to recreate a 2D image of the sample as a probability distribution of the likelihood to be classified as a certain biomolecule. These two metrics allow for us to view the prediction distribution of certain molecules in a 2D image of cells or tissue and a metric for accuracy on this prediction.

Acknowledgements (Optional):