MEDICAL VOLUMETRIC IMAGE PROCESSING WITH DEEP NEURAL NETWORKS

Medical image processing is a vital step of healthcare systems. The development of modern medical tools and cameras in the last decade has resulted in the mass production of digital medical data. Along with the increase in the human population and the low ratio between medical professionals and patients, the need for an automated and intelligent healthcare system has become vital. Amidst this, Deep learning has emerged as a transformative paradigm in the field of medical image processing, offering unprecedented capabilities in image analysis, interpretation, and diagnostic decision support. This thesis provides a comprehensive overview of the application of deep learning techniques to the medical images presented in our use cases, exploring the advancements, challenges, and future directions in this rapidly evolving domain. The thesis is divided into three parts. The first part discusses the role of Artificial Intelligence (AI) in healthcare systems, along with its history and advantages. The challenges of AI, including trust and explainability, legal concerns, AI fairness, and data sensitivity are also addressed. The limitations of traditional image processing are presented along with the solution in the form of deep learning networks. The second part relates to our first practical medical use case, lung nodule segmentation and classification using UnitoChest, our novel dataset, and other state-of-the-art datasets comprising of DICOM (Digital Imaging and Communications in Medicine) chest images. This part also addresses the process of collecting a new dataset, UnitoChest, and presents its statistical features. We use state-of-the-art 2D and 3D deep learning networks for the segmentation of lung nodules along with proposing a tweaked variant segmentation model. In the third part, we present our second medical use case, malaria parasite type and lifecycle classification.We use various publicly available datasets containing microscopic stained blood slide images and present a comparison between their classification results. Furthermore, we also propose a novel lightweight architecture for malaria classification that can be easily embedded in mobile devices and used in developing countries to assist medical professionals. The findings presented in this thesis contribute to a heightened insight into deep learning networks and highlight the potential of AI in medical and computer vision challenges.