Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

High School Student with a Traditionally Underrepresented Background

HS Poster

Poster 15 - Analysis of Collagen Fiber Imaging Through Coherent Vibrational Microscopy and Machine Learning

Saturday, October 26, 2024
10:00 AM - 12:00 PM EST
Location: Exhibit Hall E, F & G

Introduction: The protein collagen is primarily found in connective tissues providing structural support in the human body. In oncological pathology, collagen fibers exhibit distinct aberrations, including increased density, altered orientation, and changes in the collagen structure. Yet, current molecular-level insights into collagen fibers are lacking. In this study, we employed the noninvasive, label-free sum-frequency generation (SFG) microspectroscopy to explore the structure of collagen fibrillar and identify tissue disease-related molecular signatures. We developed a deep learning model for facilitating rapid detection of cancerous cells, which has the ability to distinguish between cancerous and healthy cells with high accuracy, offering a novel approach for early cancer diagnostics.

Materials and

Methods: Sum-frequency generation (SFG) is a spectroscopic technique involving overlapping infrared (IR) and visible laser beams. The IR frequency matches the frequency of the vibrational mode of the C-H and N-H stretch modes (2900 cm-1 and 3300 cm-1). The sum frequency of those two lights is generated from the surface, providing surface specificity.
Singular value decomposition (SVD) is a matrix factorization technique used to isolate variance in data. We use SVD to analyze the SFG spectral image to obtain the represented spectra and the spatial distribution of their amplitudes.
Deep learning is a machine learning method that utilizes neural networks to make predictions about unseen data. We employ a semi-supervised training methodology, using data from SFG images of the normal and cancerous mouse and human cells.

Results, Conclusions, and Discussions: To investigate the molecular packing and interactions, we record hyperspectral SFG images of the collagen samples in control and tumor samples. The transmission bright field (BF) image and the corresponding SFG image of the human tumor tissue are shown in Figs. 1(a) and 1(b), respectively. The SFG image indicates that collagen’s molecular structure and orientation are inhomogeneously distributed. Using SVD analysis, we found that the first five components are quite different between the control and tumor samples, suggesting that their microscopic environments are distinct.
To further analyze the SFG image, a logistic regression deep learning model was implemented to classify pixels from the spectral images as tumors and non-tumors. A weakly supervised learning approach was used to label pixels from the tumor image. The metric for inexact supervision was determined by the distribution of the ratios of C-H and N-H peaks in both control and tumor images. The mean and median ratios were 1.7649 and 1.5091, respectively, for the control, and 1.1892 and 1.0104, respectively, for the tumor. The total data amounted to 15,250 data points, and 40% of it was set aside to test the model. The deep learning algorithm’s average test accuracy after 100 epochs was 89.47%. The training and validation loss curves shown in Fig. 1(c) are nearly equal to each other, suggesting that the model is reasonably well-tuned. However, the loss values stabilize at around 0.4, which is not particularly high, but indicates room for improvement. It is likely that more data is needed to further verify the accuracy of this model.
The collagen fiber in tumors and control samples demonstrated distinct SFG signals. Using multivariate analysis and machine learning algorithms, we can accurately identify the tumor regions. This may pave the way for early cancer diagnosis.

Acknowledgements (Optional): This work was supported by a generous gift from Beckman Laser Institute Inc. (L. S.), National Institutes of Health & National Institute of General Medical Sciences (W. X. and C. C. Y.). The high school students (*) participated in the UCSD IEM OPALS internship. Special thanks to Dr. Shu Chien from UCSD Bioengineering and Dr. Lizhu Chen from CorDx Inc. for their donations

References (Optional):