Detection of Early Gastric Cancer and Lesion Segmentation Based on Deep Learning

Background: Early detection and precise identification of gastric cancer tumors significantly enhance patient outcomes. Conventional methods often rely on manual interpretation of endoscopic images, which can be time-consuming and subjective. Recent advancements in deep learning offer promising alternatives for automating and improving these diagnostic processes. Objective: The primary objective of this study is to explore the effectiveness of a deep learning model in detecting early gastric cancer and segmenting lesions from endoscopic images. Methods: This retrospective study was conducted at Khalifa Gul Nawaz Teaching Hospital /BMC Bannu, from December 2021 to November 2022. Medical records and endoscopic images from 180 patients with suspected or confirmed gastric lesions were analyzed. Images were diversified in terms of lesion types and characteristics. Preprocessing steps including standardization and enhancement techniques were applied to improve image quality for analysis. Results: The deep learning model achieved an accuracy of 90% in identifying gastric cancer lesions, with a sensitivity of 85% and specificity of 92%. The area under the curve (AUC-ROC) was calculated to be 0.61, indicating a good discriminative performance of the model. Conclusion: The deep learning model demonstrated significant potential for enhancing the detection and segmentation of early gastric cancer from endoscopic images, providing a valuable tool for gastroenterologists in the early diagnosis and treatment planning.