Poster Session

Ashutosh Raman, Tanner J. Zachem, Ravi Prakash, Guangshen Ma, Dr. Weston A. Ross, Dr. Patrick Codd 

Department of Biomedical Engineering, Duke University; Department of Mechanical Engineering and Materials Science, Duke University; Duke University Medical Center, Duke University

Intro: Effectiveness of precision tumor removal procedures is greatly diminished when all cancer cells are not detected, leading to more frequent recurrences of disease. It is especially important to provide this accurate intraoperative diagnosis in a minimally invasive manner, so as not to damage surrounding healthy tissue. Fluorescence spectroscopy offers particular advantages for this, due to its ability to quickly collect biological information in a noncontact manner. Moreover, endogenous fluorescence enables the use of spectroscopy without external dyes, thus decreasing toxicity and preparation time. In this study we use intrinsic fluorescence spectroscopy (IFS) on a murine model, to highlight metabolic differences in cancerous sarcoma and healthy tissue. Thereafter, we utilize post-processing and machine learning techniques to classify newly excised tissue based solely on endogenous fluorescence. Methods: A 405nm, 180mW excitation laser with .75mm spot size and a CCD-Spectrometer with an integration time of .5s were used to capture spectral emissions for this study. Soft-tissue sarcoma tissue from the left hind legs of 6 newly sacrificed LSL-KrasG12D/+; p53Flox/Flox (KP) mice was excised, verified as cancerous by a subject matter expert, and then placed on our portable spectrometer stage setup. Healthy tissue from the right hind leg was also imaged as a control sample. The laser spot was systematically positioned on discrete, non-overlapping areas of the tissue sample, and spectra were collected. 513 samples were collected, with 395 sarcoma spectra taken. Wavelength collection ranges spanned from 200-1000nm, however, ranges were trimmed to 460-1000nm to correct for excitation light reflection. Spectra with an average intensity below .05 were disregarded to maintain realistic Signal-to-Noise Ratio. Lastly, spectra were normalized and smoothed. Principal Component Analysis (PCA) was used for feature engineering, followed by classification with K-Nearest Neighbors, Logistic Regression, Support Vector Machines, and a simple Artificial Neural Network (ANN). The former three were subjected to nested Grid Search Cross Validation, to ensure optimal parameters and generalizability. Results and Discussion: Sarcoma and healthy tissue spectra exhibited expected intensity differences predicted by their physiology, namely in the locations of 470nm (free NADH), and 570nm (Free FAD). Use of ANNs is substantiated, with accuracies around 94%, though their computational intensivity and training time remain prohibitive. For the three other ML pipelines, sensitivity, specificity, F1 score, accuracy, and AUC are compared, and PCA-SVM is shown to be the best performer, with testing accuracy of 88%, recall of 90%, and specificity of 83%. Moreover, PCA-SVM showed negligible difference from other classifiers in its training time, warranting its use in intraoperative settings, and its preference over ANNs. Conclusion/Future: Our study highlights the ability of a portable non-invasive fluorescence spectroscopy device to accurately classify sarcoma tissue from healthy tissue, and also shows the relative ease of classification with machine learning methods. Future work should focus on integration into medical robotics and further optimization of machine learning models. With the demonstration of this point-of-care method, we intend to explore the diagnostic capabilities of IFS in conjunction with other modes of sensing, to create a multimodal framework for intraoperative tissue evaluation.

ashutosh.raman@duke.edu