Gastrointestinal (GI) diseases remain a significant global health challenge, particularly in low-income settings where diagnostic resources are often scarce. Endoscopic examination is essential for detecting and monitoring these diseases, yet the manual analysis of the resulting images is time-consuming, prone to observer variability, and demanding of clinical expertise. Recent advances in computer-aided diagnosis (CAD) using deep convolutional neural networks (CNNs) have shown promise in automating endoscopic image classification, but limited annotated data and the subtlety of GI findings continue to pose challenges. To address these constraints, this study proposes a two-level stacking ensemble framework that combines three pre-trained CNN architectures—ResNet50, DenseNet201, and MobileNetV3Large—with four classical machine-learning meta-classifiers (Logistic Regression, Random Forest, Support Vector Machine, and K-Nearest Neighbors). The KvasirV2 dataset, comprising 8,000 GI endoscopic images across eight classes, was used to train and evaluate the models. Results indicate that the stacking ensemble achieved a top accuracy of 94.33%, surpassing individual CNN baselines by 1–2%. Notably, this multi-level ensemble approach demonstrated improved diagnostic consistency for challenging classes like early-stage esophagitis and normal Z-line, suggesting that synergizing diverse CNN feature extractors can mitigate the limitations of single-network methods. These findings underscore the potential of ensemble-based transfer learning to enhance clinical decision support, reduce observer variability, and facilitate earlier, more accurate detection of GI diseases.