Construction of mass spectra database and diagnosis algorithm for head and neck squamous cell carcinoma

Abstract

Objectives: Intraoperative identification of tumor margins is essential to achieving complete tumor resection. However, the process of intraoperative pathological diagnosis involves cumbersome procedures, such as preparation of cryosections and microscopic examination, thus requiring more than 30 min. Moreover, intraoperative diagnoses made by examining cryosections are occasionally inconsistent with postoperative diagnoses made by examining paraffin-embedded sections because the former are of poorer quality. We sought to establish a more rapid accurate method of intraoperative assessment.

Materials and methods: A diagnostic algorithm of head and neck squamous cell carcinoma (HNSCC) using machine learning was constructed by mass spectra obtained from 15 non-cancerous and 19 HNSCC specimens by probe electrospray ionization mass spectrometry (PESI-MS). The clinical validity of this system was evaluated using intraoperative specimens of HNSCC and normal mucosa.

Results: A total of 114 and 141 mass spectra were acquired from non-cancerous and cancerous specimens, respectively, using both positive- and negative-ion modes of PESI-MS. These data were fed into partial least squares-logistic regression (PLS-LR) to discriminate tumor-specific spectral patterns. Leave-one-patient-out cross validation of this algorithm in positive- and negative-ion modes showed accuracies in HNSCC diagnosis of 90.48% and 95.35%, respectively. In intraoperative specimens of HNSCC, this algorithm precisely defined the borders of the cancerous regions; these corresponded with those determined by examining histologic sections. The procedure took approximately 5 min.

Conclusion: This diagnostic system, based on machine learning, enables accurate discrimination of cancerous regions and has the potential to provide rapid intraoperative assessment of HNSCC margins.

Keywords: Ambient mass spectrometry; Diagnosis; Head and neck squamous cell carcinoma; Machine learning; Partial least squares logistic regression; Probe electrospray ionization; Real-time analysis.