Automated cytological detection of Barrett's neoplasia with infrared spectroscopy.

Abstract

BACKGROUND: Development of a nonendoscopic test for Barrett's esophagus would revolutionize population screening and surveillance for patients with Barrett's esophagus. Swallowed cell collection devices have recently been developed to obtain cytology brushings from the esophagus: automated detection of neoplasia in such samples would enable large-scale screening and surveillance. METHODS: Fourier transform infrared (FTIR) spectroscopy was used to develop an automated tool for detection of Barrett's esophagus and Barrett's neoplasia in esophageal cell samples. Cytology brushings were collected at endoscopy, cytospun onto slides and FTIR images were measured. An automated cell recognition program was developed to identify individual cells on the slide. RESULTS: Cytology review and contemporaneous histology was used to inform a training dataset containing 141 cells from 17 patients. A classification model was constructed by principal component analysis fed linear discriminant analysis, then tested by leave-one-sample-out cross validation. With application of this training model to whole slide samples, a threshold voting system was used to classify samples according to their constituent cells. Across the entire dataset of 115 FTIR maps from 66 patients, whole samples were classified with sensitivity and specificity respectively as follows: normal squamous cells 79.0% and 81.1%, nondysplastic Barrett's esophagus cells 31.3% and 100%, and neoplastic Barrett's esophagus cells 83.3% and 62.7%. CONCLUSIONS: Analysis of esophageal cell samples can be performed with FTIR spectroscopy with reasonable sensitivity for Barrett's neoplasia, but with poor specificity with the current technique.