Identification of precancerous lesions by multispectral gastroendoscopy

Gastric cancer is one of the fifth most deadly cancers worldwide. Nowadays the diagnosis is performed through gastroendoscopy under white light and histological analysis. However, the precancerous lesions are multifocalized and present low differences with respect to healthy tissue. Several systems have been proposed based on light tissue interaction to improve the visualization of malignancies. However, these systems are limited to few wavelengths. In this paper, we propose a minimally invasive technique based on multispectral imaging and a methodology to identify malignancies in the stomach. We developed a multispectral gastroendoscopic system compatible with current gastroendoscopes, where only the illumination is changed. The spectra are extracted from the acquired multispectral images in order to compute statistical features that are used to classify the data in two classes: healthy and malignant. The features are ranked by pooled variance t test to train three classifiers. Neural networks using generalized relevance learning vector quantization, support vector machine (SVM) with a Gaussian kernel and k-nn are evaluated using leave one patient out cross-validation. Taking into consideration the data collected in this work, the quantitative results from the classification using SVM show high accuracy and sensitivity using a low number of features. These results show the ability to discriminate malignancies of the gastric tissue. Therefore, multispectral imaging could help in the identification of malignancies during gastroendoscopy.