Background: Histopathology is generally considered a ‘Gold Standard’ in diagnosis of cancers. But it is invasive and time consuming. So cytology-based non-invasive diagnosis is recommended. However, conventional cytology suffers from low sensitivity and specificity. Therefore Raman spectroscopy has been considered as an alternative for reliable cytodiagnosis as it is a non-invasive, label-free diagnostic technique with high specificity. Recently, artificial intelligence (AI) and machine learning has transformed many fields but is still considered a Blackbox when applied to vibrational spectroscopy as it suffers from incomplete understanding of molecular signatures for automated yet reliable decision making. In this research, we extract Raman signatures responsible for classification by explainable AI models.<br/><br/>Methods: Raman spectra were obtained from 30 human mammary epithelial cells (HMEpC) and breast cancer cells (MCF-7) cells with 632.8 nm excitation. Both conventional univariate and multivariate statistical methods including multivariate curve resolution (MCR-ALS) were utilized to either classify or model pure component Raman spectra. Neural network models were constructed followed by application of various feature extraction methods.<br/><br/>Results and Conclusions: Average Raman spectra of HMEpC and MCF-7 cells is shown in Fig.1A. There is a marked difference in univariate ratios of DNA/Lipids (Fig.1B) and Protein/Lipids (Fig.1C). However, this procedure is ambiguous as multiple components may contribute to each band. Conventional multivariate methods discriminate breast cancer cells efficiently but lack detailed molecular information. MCR-ALS analysis throws light on molecular changes in cancer cells but requires human inputs. Therefore we developed explainable AI models with superior performance and revealed important Raman signatures for reliable classification. Our results demonstrate that RM + AI, as an adjunct to morphology, has the potential to become an excellent cytodiagnostic tool that can both accurately and objectively discriminate cancer from normal cells.

Explainable AI assisted identification of molecular signatures in Raman spectroscopy for reliable and automated classification of breast cancer cells

Where thousands of chemistry professionals meet to share ideas and advance scientific and technical knowledge. Spring 2023 abstracts, posters, and presentations will launch at the end of March.

ACS Spring 2023

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.