INCEPTION V1-BASED EXPLAINABLE DEEP LEARNING FRAMEWORK FOR GASTROINTESTINAL ENDOSCOPIC IMAGE CLASSIFICATION AND LESION ANALYSIS

Abstract

It is very important for the diagnosis and treatment of GI diseases that they should be done early in order to have good chances of recovery. Although the conventional way of examining the digestive system through visualization using an endoscope is employed, the fact that some abnormalities in the mucosa tend to appear small can be overlooked. The idea of using Inception V1 (GoogleNet V1) for the classification of GI endoscopy images is made as a result of this concern. Using multi-branch architecture of this neural network allows the identification of lesions regardless of their size and property. The three kinds of endoscopy images considered for testing the algorithm include: (1) Dyed Lifted Polyps, (2) Normal Cecum, and (3) Oesophagitis.As a result of the suggested methodology, it provided 92.50% accuracy, 92.68% precision, 92.21% recall rate, and 92.39% F1 score. Moreover, the specificities, AUC-ROC, and Cohen's Kappa values were 95.74%, 0.967, and 0.887 correspondingly. The accuracy for classification was highest for Normal Cecum (95.20%), then for Dyed Lifted Polyps (93.10%), and then for Oesophagitis (89.20%). For training the model, the training and validation accuracy were 96.80% and 92.50% correspondingly and the corresponding loss values were 0.112 and 0.238. Additionally, grad-CAM technique was used in order to understand which parts of the image impact the prediction of the model and understand better the whole process of classification. Thus, based on the overall results of this research, it can be said that Inception V1 is an efficient way of analyzing gastrointestinal endoscopic images.