Accurate and timely diagnosis of gastroesophageal disease and polyps from endoscopic images remains a demanding but crucial function in clinical practice due to visual analogy between pathological and non-pathological tissues. In this research, comparative evaluation of three deep learning models, ResNet-50, InceptionV3, and Ensemble Transfer Learning (Ensemble TL) framework, for multi-class classification of GERD, GERD Normal, Polyp, and Polyp Normal using highresolution endoscopic images is presented. The models were evaluated based on performance metrics: accuracy, precision, recall, F1 measure, and classification accuracy. ResNet-50 showed strong baseline performance with an accuracy of 73% and balanced performance on Polyp detection (F1: 0.78). Its GERD recall (0.65) and GERD Normal precision (0.64) reflected challenges in separating closely similar visual features. InceptionV3 achieved improved performance, particularly in GERD (F1: 0.78) and Polyp (F1: 0.83) detection, with a overall accuracy of 77 %. It did degrade its performance in GERD Normal sensitivity, though. The Ensemble TL model outperformed all the individual models, with a highest accuracy of 87 % and achieving improved F1 scores for all disease classes: GERD (0.85), GERD Normal (0.82), Polyp (0.94), and Polyp Normal (0.87). These enhancements are credited to the ability of the model to integrate high-level features from ResNet and Inception architectures, enhancing global and local feature learning. The paper underlines ensemble learning's role in complex medical imaging applications, giving consistent and transferable performance across diverse forms of diseases.