Detection of Gastric Cancer in Endoscopic Images Using Deep Convolutional Neural Networks

In finding features of gastric cancer in endoscopic images, difficulties arise from early manifestations and variance in the appearance of lesions. The objective of this study is to design an automobile deep convolutional neural network (CNN) that would help enhance the rate of accurate diagnosis of gastric cancer. In the current work, using the dataset of endoscopic images reviewed by senior gastroenterologists, transfer learning is applied on the base of the ResNet-50 based deep CNN the result of data augmentation is extended. By evaluating the model on measures of training, validation, and testing sets, high performance has been realized as depicted by accuracy of 87. 4%, precision of 87. 5%, recall of 87%, and f1-score of 87%. Research on the comparison of the proposed method with the conventional CAD systems and machine learning models confirmed the efficiency of the deep learning method. The utilization of Grad-CAM visualizations brought interpretability in improving match between regions of focus and clinically relevant areas therefore improving the trust in the model among the clinicians. To foster efficiency, the implementation incorporated Python and deep learning tools, including TensorFlow, the model of which was trained on GPUs. Thus, this paper demonstrates the feasibility of the studied CNNs in promoting the early diagnosis of gastric cancer, suggesting the adaptation of AI tools as complementary tools for clinicians in everyday practice. Future studies should follow up on how false negative results may occur; additionally, future studies should devise better approaches on how to increase the generalizability of these models across different patients.