A Hybrid Deep Learning Framework for Gastrointestinal Tract Classification Based on Endoscopic Images

This work introduces a set of pre-trained models that can accurately identify GI (gastrointestinal) diseases and disorders using endoscopic pictures. For the purpose of disease classification in the gastrointestinal system, we provide GIT- NET, a weighted average ensemble model. We used an eight- class KVASIR v2 dataset to test the model. Using individual models for categorization increases the likelihood of misclassification because they might not learn all the classes' traits sufficiently. One possible explanation is that different models may excel at learning the traits of certain classes relative to others. We provide an ensemble-model that uses the anticipations of three pre-trained models, namely DenseNet201 (94.54% accuracy), InceptionV3 (88.38% accuracy), &ResNet50 (90.58% accuracy). The two approaches used to aggregate the anticipations of the base learners are model averaging. Based on the results of the model evaluations, the weighted average ensemble achieved a 95.00% success rate, while the model averaging ensemble achieved a 92.96% success rate. Both the model average ensemble and each distinct model are surpassed in performance by the weighted average ensemble. When using an ensemble of base learners, features that were mislearned by discrete base learners can be accurately classified, according to the evaluation results.