Ensemble Model for Medical Diagnostics: A Hybrid Approach to Improve Prediction Accuracy

The wide spread of gastrointestinal tract diseases (GI) and high mortality rate associated with it raises an urgent need for its accurate diagnosis. The traditional methods include endoscopy and CAD but they lack annotated images. Hence, gastrointestinal tract (GIT) disorders classification becomes a very challenging task. To address this critical problem, an accurate and efficient diagnosis method is required for prompt treatment. The current research work aims at presenting an ensemble model based on deep learning approach that efficiently classifies GI tract diseases with comparatively higher accuracy rate. Resnet, densenet and googlenet models are utilized to create an ensemble model. The model takes advantage of the unique characteristics of all of the three models. Two GI tract datasets, Kvasir v1 and Kvasir v2 are utilized for training. A hybrid preprocessing approach is formulated, including scaling, augmentation and guassian noise injection to enhance image quality and further improve performance of classification task. Furthermore, fine tuned pretrained models resnet50, densenet121 and googlenet are combined to create an ensemble model for feature extraction and efficient classification. Stacking and meta learner ensemble technique is employed for creating an ensemble model to identify diseases more precisely. Fine tuned ensemble model achieved the highest accuracy of 99.1 %, and 97.3 % on Kvasir v1 and v2 respectively, while maintaining a minimum testing time of 0.00001 s. It utilizes 5-fold cross-validation and three transfer learning (TL) models as feature extractors. Hence, in terms of accurate predictions, this proposed framework outperforms existing models in terms of performance and accuracy with accuracy of 99.1% on Kvasir v1 and 97.2 % on Kvasir v2.