Machine Learning Approaches in Non-Contact Autofluorescence Spectrum Classification
Ashutosh P. Raman, Tanner J. Zachem, Sarah Plumlee, Christine Park, William Eward, Patrick J. Codd, Weston Ross
Abstract
Manual surgical resection of soft tissue sarcoma tissue can involve many challenges, including the critical need for precise determination of tumor boundary with normal tissue and limitations of current surgical instrumentation, in addition to standard risks of infection or tissue healing difficulty. Substantial research has been conducted in the biomedical sensing landscape for development of non-human contact sensing devices.
One such point-of-care platform, previously devised by our group, utilizes autofluorescence-based spectroscopic signatures to highlight important physiological differences in tumorous and healthy tissue. The following study builds on this work, implementing classification algorithms, including Artificial Neural Network, Support Vector Machine, Logistic Regression, and K-Nearest Neighbors, to diagnose freshly resected murine tissue as sarcoma or healthy.
Introduction
In precision oncologic surgeries, it is important to be able to continuously identify tumor margins and classify tissue during the course of a surgical resection. This includes the ability in real-time to make this determination even as the anatomic landscape changes as tissue is removed or additional anatomic regions are exposed. Sarcoma, a connective tissue cancer, is one of a variety of malignant tumors that require precise resection to maximize survival and reduce disease progression while salvaging healthy surrounding tissue to minimize damage to important neurovascular and biomechanically important tissues.
Methods
To produce the samples for evaluating the spectral signature classification hypothesis of the spectrometer, ex vivo surgically explanted murine sarcoma samples were compared with otherwise healthy murine tissue in an equivalent anatomic location.
To determine an approximate number of mice necessary to attain a comparable classification accuracy to that of a surgeon using visual cues, an inverse power law regression was fit to validation accuracy results at various sample sizes for an artificial neural network, and then extrapolated to 90% classification accuracy [40]. This resulted in a prediction of approximately 500 total samples of healthy and sarcoma data- with reasonable interclass balance- needed to achieve comparable accuracy to a surgeon using visual assessment.
Results
Spectral signatures acquired from each class of tissue with the point-of-care fluorescence device were plotted according to the experimental set-up. This yielded the following set of healthy and sarcoma tumor tissue spectra, outlined in Fig 4A and 4B with appropriate cutoffs, smoothing, and normalization at the biologically relevant 500 nm wavelength region.
As is seen in Fig 4A and 4B, it is difficult to visually ascertain major differences, aside from some noticeable unique peaks in the sarcoma spectra. Thus, all spectra were afterwards averaged and presented in a single plot with standard errors (Fig 5). To generate data for Fig 5A, no normalization was done to individual collected spectra; rather, all raw spectra in a given class were averaged, and the resulting two averaged spectral class curves were divided by the global maximum average intensity value, which was the maximum value of the averaged healthy spectra curve in this study.
Citation: Raman AP, Zachem TJ, Plumlee S, Park C, Eward W, Codd PJ, et al. (2024) Machine learning approaches in non-contact autofluorescence spectrum classification. PLOS Digit Health 3(10): e0000602. https://doi.org/10.1371/journal.pdig.0000602
Editor: Frank Rudzicz, Dalhousie University Faculty of Computer Science, CANADA
Received: May 2, 2023; Accepted: June 10, 2024; Published: October 9, 2024
Copyright: © 2024 Raman et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The data (i.e. spectroscopic signatures and labels) supporting the findings in this paper are available from the Duke Research Data Repository (https://doi.org/10.7924/r4vt1vh11).
Funding: A.R. received funding from the National Science Foundation Graduate Research Fellowship (https://www.nsfgrfp.org/). T.Z. received funding from the Goldwater Scholarship (https://goldwaterscholarship.gov/). W.R., P.C, and W.E. received funding from National Institutes of Health R01 Grant: EB030982 (https://www.nih.gov/). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing Interests: The authors have declared that no competing interests exist.
Source: https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000602#sec015
