SNet: A novel convolutional neural network architecture for advanced endoscopic image classification of gastrointestinal disorders

With the intent of assisting gastroenterologists from all over the world, the proposed work aims to eliminate the effort required to achieve accurate diagnoses. Statistically, gastrointestinal diseases often result in fatal disorders, contributing to a significant number of fatalities. The upper gastrointestinal tract (GIT) includes the stomach, esophagus, and duodenum, while the lower one comprises a section of the small intestine, namely the ileum, as well as the large intestine, including the colon. The challenges associated with GIT tract issues are apparently complex. Therefore, multiple challenges exist regarding CAD (Computer-aided diagnosis) and endoscopy, including a lack of annotated images, a dark background, poor contrast, and an irregular pattern. The objective of this research is to develop a robust deep network, called SNet, that offers a solution to complex classification problems. Firstly, the endoscopic images undergo preprocessing before being subjected to feature extraction. This step involves image resizing along with the augmentation step. The proposed convolutional neural network (CNN) model is comprised of six blocks placed at different layers. To enable the exhaustive evaluation of proposed framework across different datasets, the model has undergone training on a very complex HyperKvasir dataset, and later tested on Kvasir v1 and v2 datasets. This facilitates cross-dataset system evaluation, resulting in an efficient system for an unseen image diagnosis. To avoid the problem of "curse of dimensionality", the most discriminant feature information is selected based on proposed minimum redundancy maximum relevance (MRMR) algorithm. The proposed architecture has been evaluated using a range of performance metrics, such as accuracy, sensitivity, specificity, and Area under curve (AUC). With classification accuracy as the main metric, the achieved accuracy is 98.45% on the Kvasir v1, and 97.83% on the Kvasir v2 datasets.