Performance Evaluation of Basic Capsule Network Architecture in Classification of Biomedical Images

Abstract

In order to diagnose and treat diseases, a variety of imaging techniques are used, including X-ray, computed tomography (CT), mammography, ultrasound, and magnetic resonance imaging (MRI). Correct analysis of these medical images is required for early disease detection and application of the appropriate treatment. In image analysis, the identification of the relevant area, as well as information such as its size, location, and direction, are critical in determining the best treatment methods. The convolutional neural network (CNN) architecture is one of the most widely used deep learning architectures in medical image analysis. However, it was stated that CNN was insufficient to measure the relationship between these features while extracting image features, and it could not hide features such as pose (position, direction, size), deformation, and texture. The Basic Capsule Network (CapsNet) Architecture was proposed to overcome CNN’s disadvantage and increase success. In this study, MedMNIST dataset collection consisting of medical images was used. The RetinaMNIST, BreastMNIST, and OrganMNIST-A datasets included in MedMNIST were used to evaluate the classification performance of the CapsNet architecture. Capsnet succeeded in these with accuracy rates of 54%, 83%, and 89%, respectively. CapsNet has been shown to produce comparable results to advanced CNN models.