Detection and Characterization of Gastric Cancer Using Cascade Deep Learning Model in Endoscopic Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Endoscopic examination for gastric cancer requires significant expertise.
• Previous image segmentation methods faced challenges with false positives and high computational demands.
• Diagnostic support tools are crucial for improving endoscopic accuracy.

Purpose of the Study:

• To develop a cascaded deep learning model for classifying endoscopic images and identifying gastric cancer invasion.
• To overcome limitations of previous methods, including false positives and computational costs.
• To enhance the accuracy and efficiency of gastric cancer diagnosis via endoscopy.

Main Methods:

• A convolutional neural network was used to classify endoscopic images into normal, early gastric cancer, or advanced gastric cancer categories.
• Two U-Net models performed segmentation to identify the extent of cancer invasion in classified positive images.
• The model was evaluated on a dataset of 1208 normal, 533 early gastric cancer, and 637 advanced gastric cancer images.

Main Results:

• Image classification achieved 97.0% sensitivity and 99.4% specificity for gastric cancer detection.
• Case-based evaluation demonstrated 100% sensitivity and specificity.
• The model successfully identified the extent of cancer invasion with acceptable accuracy.

Conclusions:

• The proposed cascaded deep learning model effectively classifies endoscopic images and identifies gastric cancer invasion.
• This AI-driven approach offers improved diagnostic accuracy and efficiency compared to previous methods.
• The method shows potential as a valuable tool for supporting gastric cancer diagnosis in clinical settings.