Analysing Wireless Capsule Endoscopy Images Using Deep Learning Frameworks to Classify Different GI Tract Diseases

GI Tract related diseases are one of the most prevalent in today's society. Studies have shown that continuous monitoring, early detection, and treatment of these diseases are imperative in improving patients' recovery rate. Wireless Capsule Endoscopy (WCE) is an innovative imaging technology that enables invasive imaging of the GI Tract. Convolutional Neural Networks (CNN) and Image Processing have become very sought-after solutions in the process of developing a Computer Aided Diagnosis (CAD) system for many medical applications. The study aims to design and develop a generalized multiclass CNN classification algorithm to be used in CAD system for diagnosis of various GI tract diseases by analyzing WCE GI tract images with varying tract lining lesions. CNN classification-based solution framework encompassing various network architectures, image processing enhancement techniques and data augmentation methods are proposed. Three histogram stretching based enhancement techniques were introduced to enhance the quality of the raw image prior to performing classification. Data augmentation was performed as well. Different network architectures of self-developed architectures, transfer learning feature extraction, fine tuning and an ensemble of models were developed. The results were analyzed, putting emphasis on the generalization capability of the developed solutions. Results showed that image processing enhancement improved the CNN models' capability in performing accurate classification. In terms of individual network architectures, the transfer learning fine tuning models performed better as compared to the rest of the architectures. CNN networks trained on the dataset with augmentation are more generalized as compared to CNN networks trained on non-augmented data. The final proposed solution for GI tract CAD CNN network is the ensemble model which managed to achieve an overall accuracy of 97.03% when tested and compared to other proposed architectures across 4 phases of result analysis.