Gastrointestinal (GI) endoscopy is the preferred technique for detecting lesions and important landmarks in the GI tract, but manually examining these images is laborious and subjective. To further propel the development of computeraided diagnosis, we conduct an extensive evaluation of deep learning models on the Kvasir-v3 dataset, which contains 7,976 high-quality annotated endoscopic images from eight nearly balanced classes. In our experiments, we compare several deep learning architectures: conventional CNNs (VGG16, VGG19, InceptionV3, Xception, DenseNet169), a standalone Vision Transformer (ViT) and hybrid models using feature-level fusion of ViT + InceptionV3, ViT + DenseNet169. Of these models, DenseNet169 achieves the best performance, yielding a macroaverage F1-score and accuracy of 0.89, surpassing the others in the experiments. In contrast, the standalone Vision Transformer achieves lower performance, at an accuracy of 0.76 to 0.80, owing to the relatively small dataset. The best performance is achieved by the hybrid model ViT + DenseNet169, yielding macro-averaged precision, recall, F 1 -score, and accuracy of 0.94, an improvement of 5 to 6% relative to DenseNet169, and an improvement of about 22% relative to ViT. To further ensure the trustworthiness of the proposed model, we also apply LIME and saliency maps for visual explanation, showing that the model is focusing on the right features, such as the texture of the mucosal layer, pathological regions, and procedural landmarks, making DenseNet169 a trustworthy solution for multi-class GI disease and landmark classification in endoscopic procedures.