Capsule networks in medical image processing

This PhD thesis focuses on the development of Computer Aided Diagnosis (CAD) systems for the automatic classification and segmentation of medical images obtained from different imaging modalities. Medical imaging plays a critical role in healthcare by providing non-invasive tools to visualize the internal structures and functions of the body. However, the analysis and interpretation of medical images by radiologists is often subjective and time consuming. This thesis therefore explores the use of artificial intelligence, specifically deep learning models, to automate the detection and classification of regions of interest in medical images. The thesis includes four studies, each focusing on a different medical imaging modality. The first study involves the classification of benign/malignant masses in mammogram images using classical Convolutional Neural Networks (CNN) and transfer learning models, followed by classification using the capsule network model developed in this thesis. The second study focuses on the segmentation of polyps in colonoscopy images using a modified U-Net model and the performance of a capsule network-based segmentation model after analysis of different parameters to optimize the segmentation performance. In the third study, both classification and segmentation of benign/malignant adrenal lesions in abdominal MR images are investigated. For classification, different regions of interest are extracted, and separate studies are performed using capsule network-based and KSA-based models. For segmentation, a new model is proposed by modifying the classical U-Net model and the effect of different parameters and the capsule network-based segmentation model on the performance is evaluated. Finally, a special capsule network structure is proposed for pneumonia classification in children using X-ray images. All studies are compared with similar studies in the literature or with the latest models and their advantages are shown. Overall, this thesis presents new deep learning models for automatic classification and segmentation of medical images, which can potentially improve the accuracy, efficiency and consistency of medical diagnoses.