Classification of Celiac disease using ensemble SMOTE-RF approach

Abstract

There are emerging challenges in the medical field as artificial intelligence is being introduced for automated detection of different diseases including celiac disease (CD). Manual interpretation is time time-consuming task performed by endoscopists whereas experience plays a key role in figuring out the abnormal region. The endoscopy images are processed with automated detection as their results are promising in terms of presence and non-presence of the disease. Initially, fastNLmeans denoising enabled pre-processing with Synthetic Minority Oversampling Technique (SMOTE) is utilized for mitigating the effect of the noise and for increasing the minority class images. For effective automatic classification, segmentation is a crucial step in image processing.  Then, the Sobel operator with Multilevel Otsu thresholding is used as a segmentation step to reduce the complexity of the image. Image spatial features are extracted using a hybrid approach of discrete wavelet transform and high-order spectra (HOS). Thereafter, the Balanced random forest classifier is implemented to classify images as normal and abnormal. This model achieves remarkable performance with an accuracy of 89.41%, precision of 94.23%, recall of 89.09%, and F1 score of 91.58%. This approach validates the results with the help of endoscopists to prove its efficacy. These results outperform by imparting a high value of precision and F1 score.