A Review of Capsule Networks in Medical Image Analysis

Computer-aided diagnosis technologies are gaining increased focus within the medical field due to their role in assisting physicians in their diagnostic decision-making through the ability to recognise patterns in medical images. Such technologies started showing promising results in their ability to match or outperform physicians in certain specialities and improve the quality of medical diagnosis. Convolutional neural networks are one state-of-the-art technique to use for disease detection and diagnosis in medical images. However, capsule networks aim to improve over these by preserving part-whole relationships between an object and its sub-components leading to better interpretability, an important characteristic for applications in the medical domain. In this paper, we review the latest applications of capsule networks in computer-aided diagnosis from medical images and compare their results with those of convolutional neural networks employed for the same tasks. Our findings support the use of Capsule Networks over Convolutional Neural Networks for Computer-Aided Diagnosis due to their superiority in performance but more importantly for their better interpretability and their ability to achieve such performance on small datasets.