Gastric cancer is the fourth most lethal malignancy worldwide. Esophagogastroduodenoscopy is the first choice procedure for diagnosis of upper gastrointestinal lesions, especially early gastric cancer. The success of this procedure depends on endoscopist's skill and the rigorous exploration of the zones with high probability of being affected. It has been documented most gastric neoplasias are lesions already existent at the examination time and unobserved when early detection is possible. For a second reader, automatic strategies must first recognize gastric anatomic regions. The aim of this paper is to assess the performance of convolutional neural networks at classifying anatomical regions. 2.054 raw upper gastrointestinal endoscopic images from 96 patients were collected and labeled as six representative sub-anatomical stomach regions. The networks were trained with transfer learning, data augmentation, and two efficient learning methods: warm-up and fine-tuning. The top-10 macro F1-score rates of the testing dataset were 84% to 87%. These preliminary tests suggest the trained networks showed good performance in recognizing sub-anatomical stomach regions of esophagogastroduodenoscopy images.