In the context of esophageal cancer diagnosis through gastrointestinal endoscopy, this study aims to examine the issues that are related with variability and misinterpretation of lesions observed under white-light endoscopy. To overcome these obstacles, the study introduces a convolutional neural network (CNN) that combines bilinear pooling and attention mechanisms. The CNN, built on the ResNet50 architecture, incorporates a novel global channel attention module and utilizes bilinear pooling to enhance feature representation by fusing multiple feature layers. The results, based on a dataset of 2101 cases from various clinical centers and white-light endoscopy images, demonstrate the effectiveness of this approach. It achieves high accuracy in classifying esophageal lesions, with classification rates of 94.2% at the image level and 96.9% at the patient level. In terms of esophageal cancer detection, it shows a sensitivity and specificity of 95.4% and 98.8% at the image level and 98.7% and 95.9% at the patient level. Notably, this approach outperforms recent models and methods in comparative experiments. These results show that the suggested network greatly raises the diagnostic accuracy for esophageal cancer under white-light endoscopy while maintaining its robustness.