SALIVA-BASED EARLY GASTRIC CANCER CLASSIFICATION USING DEEP SPARSE AUTOENCODERS

Abstract

Gastric cancer (GC) continues to represent a significant global health challenge, being one of the primary contributors to cancer-associated mortality.  Its poor prognosis is largely attributed to the absence of clear early-stage symptoms and the shortage of reliable biomarkers. Consequently, there is a critical demand for rapid, accurate, and non-invasive diagnostic methods that can support early detection, prognosis, and clinical decision-making. Saliva has recently emerged as a valuable biofluid due to its non-invasive collection and diverse molecular composition. In this work, we introduce a deep learning framework designed to classify individuals into three categories: patients with early gastric cancer (EGC), those with advanced gastric cancer (AGC), and healthy controls. Our dataset consists of 300 saliva samples, including 90 from EGC patients, 100 from AGC patients, and 110 from healthy volunteers. The findings demonstrate that the proposed approach effectively differentiates EGC and AGC cases from healthy individuals, highlighting its strong potential for clinical translation. The FC-DSAE + SMC neural network achieved outstanding performance, yielding an accuracy of 99%, selectivity of 98.9%, sensitivity of 100%, specificity of 99.57%, detection rate of 99.76%, and an F-measure of 99.5% for the EGC category.