GestroNet: A Framework of Saliency Estimation and Optimal Deep Learning Features Based Gastrointestinal Diseases Detection and Classification

In the last few years, artificial intelligence has shown a lot of promise in the medical domain for the diagnosis and classification of human infections. Several computerized techniques based on artificial intelligence (AI) have been introduced in the literature for gastrointestinal (GIT) diseases such as ulcer, bleeding, polyp, and a few others. Manual diagnosis of these infections is time consuming, expensive, and always requires an expert. As a result, computerized methods that can assist doctors as a second opinion in clinics are widely required. The key challenges of a computerized technique are accurate infected region segmentation because each infected region has a change of shape and location. Moreover, the inaccurate segmentation affects the accurate feature extraction that later impacts the classification accuracy. In this paper, we proposed an automated framework for GIT disease segmentation and classification based on deep saliency maps and Bayesian optimal deep learning feature selection. The proposed framework is made up of a few key steps, from preprocessing to classification. Original images are improved in the preprocessing step by employing a proposed contrast enhancement technique. In the following step, we proposed a deep saliency map for segmenting infected regions. The segmented regions are then used to train a pre-trained fine-tuned model called MobileNet-V2 using transfer learning. The fine-tuned model's hyperparameters were initialized using Bayesian optimization (BO). The average pooling layer is then used to extract features. However, several redundant features are discovered during the analysis phase and must be removed. As a result, we proposed a hybrid whale optimization algorithm for selecting the best features. Finally, the selected features are classified using an extreme learning machine classifier. The experiment was carried out on three datasets: Kvasir 1, Kvasir 2, and CUI Wah. The proposed framework achieved accuracy of 98.20, 98.02, and 99.61% on these three datasets, respectively. When compared to other methods, the proposed framework shows an improvement in accuracy.