Improving Capsule Network Based Deep Learning Architectures for Classification of Medical Images

Abstract

Subtle details in medical images can be crucial in disease diagnosis. Therefore, these images need to be analyzed accurately and precisely. Capsule Networks (CapsNet) take into account the spatial and hierarchical relationships of the objects in the image by storing information such as orientation, color, texture, etc. in capsules and can perform well with a small amount of data. The use of these advantages of CapsNet in the healthcare field can enable accurate and precise diagnosis of diseases. However, the feature extraction phase of traditional CapsNet is insufficient to capture the details in medical images. The computational cost increases in deepening models to solve this problem. In this thesis, we aim to stabilize the computational cost while increasing the number of extracted features, and by using the advantageous aspects of CapsNet, three new CapsNet models with high reliability and accuracy in the classification of complex medical images are developed. In these models, the classification accuracy is improved by deepening CapsNet layers. In the studies, residual blocks were utilized in order to avoid the vanishing gradient problem of increasing layers. In the first study, the FractalNet module with skip connections such as residual block and residual block is used and a new squash function is proposed. In the second study, the effect of the architecture consisting of parallel lanes instead of the current back-to-back architecture on the classification performance of CapsNet is investigated. The effect of the number of residual blocks on classification is also observed. In the third study, the aim was to reduce the increased number of parameters resulting from the deepened network in order to improve the classification performance of CapsNet. The residual block and the Convolutional Block Attention Module (CBAM) are used instead of the convolution layer of CapsNet. The residual block increased the number of features extracted and CBAM selected the important ones. This reduced the number of features to be processed and the number of parameters of the model. Various medical datasets and the CIFAR10 dataset with complex images were used to evaluate these proposed approaches. These datasets include retinal images, blood cell images, chest X-ray images, skin lesion images, and brain MRI images. The performance values obtained from the experimental studies show that the proposed approaches improve the ability of the existing CapsNet model to classify medical images and make the CapsNet model more stable.