Raman spectroscopy and artificial intelligence to predict the Bayesian probability of breast cancer
Abstract This study addresses the core issue facing a surgical team during breast cancer surgery: quantitative prediction of tumor likelihood including estimates of prediction error. We have previously reported that a molecular probe, Laser Raman spectroscopy (LRS), can distinguish healthy and tumor...
Guardado en:
| Autores principales: | , , , , , , , , , , |
|---|---|
| Formato: | article |
| Lenguaje: | EN |
| Publicado: |
Nature Portfolio
2021
|
| Materias: | |
| Acceso en línea: | https://doaj.org/article/74b8e0603ab6427490c20c75e6ad68da |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Sumario: | Abstract This study addresses the core issue facing a surgical team during breast cancer surgery: quantitative prediction of tumor likelihood including estimates of prediction error. We have previously reported that a molecular probe, Laser Raman spectroscopy (LRS), can distinguish healthy and tumor tissue. We now report that combining LRS with two machine learning algorithms, unsupervised k-means and stochastic nonlinear neural networks (NN), provides rapid, quantitative, probabilistic tumor assessment with real-time error analysis. NNs were first trained on Raman spectra using human expert histopathology diagnostics as gold standard (74 spectra, 5 patients). K-means predictions using spectral data when compared to histopathology produced clustering models with 93.2–94.6% accuracy, 89.8–91.8% sensitivity, and 100% specificity. NNs trained on k-means predictions generated probabilities of correctness for the autonomous classification. Finally, the autonomous system characterized an extended dataset (203 spectra, 8 patients). Our results show that an increase in DNA|RNA signal intensity in the fingerprint region (600–1800 cm−1) and global loss of high wavenumber signal (2800–3200 cm−1) are particularly sensitive LRS warning signs of tumor. The stochastic nature of NNs made it possible to rapidly generate multiple models of target tissue classification and calculate the inherent error in the probabilistic estimates for each target. |
|---|