Deep ensemble learning for gastrointestinal diagnosis using endoscopic image classification

Transfer learning is a valuable tool for the effective assistance of gastroenterologists in the powerful diagnosis of medical images with fast convergence. It also intends to minimize the time and estimated effort required for improved gastrointestinal tract (GIT) diagnosis. GIT abnormalities are widely known to be fatal disorders leading to significant mortalities. It includes both upper and lower GIT disorders. The challenges of addressing GIT issues are complex and need significant study. Multiple challenges exist regarding computer-aided diagnosis (CAD) and endoscopy including a lack of annotated images, dark backgrounds, less contrast, noisy backgrounds, and irregular patterns. Deep learning and transfer learning have assisted gastroenterologists in effective diagnosis in various ways. The goal of proposed framework is the effective classification of endoscopic GIT images with enhanced accuracy. The proposed research aims to formulate a transfer learning-based deep ensemble model, accurately classifying GIT disorders for therapeutic purposes. The proposed model is based on weighted voting ensemble of the two state-of-the-art (STA) base models, NasNet-Mobile and EfficientNet. The extraction of regions of interest, specifically the sick portions, have been performed using images captured from endoscopic procedure. Performance evaluation of the proposed model is performed with cross-dataset evaluation. The datasets utilized include the training dataset HyperKvasir and two test datasets, Kvasir v1 and Kvasir v2. However, the dataset alone cannot create a robust model due to the unequal distribution of images across categories, making transfer learning a promising approach for model development. The evaluation of the proposed framework has been conducted by cross-dataset evaluation utilizing accuracy, precision, recall, Area under curve (AUC) score and F1 score performance metrics. The proposed work outperforms much of the existing transfer learning-based models giving 97.83% on Kvasir v1 and 98.45% accuracy on Kvasir v2.