Computerized classification of gastrointestinal polyps using stacking ensemble of convolutional neural network

In this paper, we have proposed a classification method of gastrointestinal polyps using the stacking ensemble technique. The ensemble method consisted of three fine-tuned deep convolutional neural network architectures (Xception, ResNet-101, and VGG-19), and the network weights were initialized from the ImageNet dataset. Besides, this paper presented a multi-attribute decision-making technique-based frame selection method utilizing several measures of a suitable frame. The frame selection procedure reduces the processing overhead of the system and attained better classification results. Moreover, this study applied a set of image enhancement operations to remove specular reflection, clipping unnecessary regions, contrast enhancement, and noise reductions. The specified classification method of polyps showed significant improvement in performance metrics on available public datasets. The five-fold cross-validated performance of the study has an accuracy of 98.53 ± 0.62%, recall score of 96.17 ± 0.87%, a precision value of 92.09 ± 4.62%, a specificity score of 98.97 ± 0.36%, and an AUC score of 0.9912. This method can be helpful for endoscopists to make rigid decisions.