Gastrointestinal Disease Diagnosis with Hybrid Model of Capsules and CNNs

• Published in: IEEE International Conference on Electro Information Technology pp. 143 - 146
• Main Authors: Sarsengeldin, Merey, Imatayeva, Sanim, Abeuov, Nurmukhamed, Naukhanov, Myrzakhan, Erdogan, Abdullah Said, Jha, Debesh, Bagci, Ulas
• Format: Conference Proceeding
• Language: English
• Published: IEEE 18.05.2023
• Subjects: Benchmark testing; Capsule Networks; Computational modeling; Computer Aided Diagnosis; Dynamic Routing; En-doscopy; Endoscopes; Heuristic algorithms; Pathology; Polyp Detection; Routing; Scalability
• ISSN: 2154-0373
• DOI: 10.1109/eIT57321.2023.10187250

The Gastrointestinal (GI) tract is responsible for different types of cancer-related mortality worldwide. Regular screening is recommended to detect abnormalities in the GI tract early. However, studies have shown a large number of miss-rates of early GI precursors. This is mostly due to the shortage of experienced physicians and the overall clinical burden. A computer-aided diagnosis system can play a significant role in identifying abnormalities and assisting gastroenterologists during the examination. The main objective of this work is to develop a deep learning-based model for gastrointestinal tract findings classification (pathological findings, anatomical landmarks, polyp removal cases, therapeutic interventions, and the quality of mucosal views) using VGG16 and Capsule Networks. We ex-periment with two commonly available GI endoscopy datasets (Kvasir and HyperKvasir) to achieve this goal. We proposed VGG16+CapsNets-based architecture for the classification of GI abnormalities and findings. For the Kvasir dataset (5 classes), we obtained Matthew's correlation coefficient (MCC) of 89.00%. Similarly, for the HyperKvasir dataset (23 classes), we obtained an MCC of 83.00%. Overall our obtained results are good with the highly imbalanced dataset. Our experimental results on the retrospective dataset showed that the proposed model could act as a benchmark for GI endoscopy image classification tasks.