| Περίληψη: | Deep Learning (DL) has become dominant in computer vision and its applications are broadly met in several domains. Image and video processing, object detection, object segmentation, and image/video synthesis are some of the most popular applications of DL. DL has attained remarkable results in such fields and has proven to be superior to conventional Machine Learning (ML) methods.
DL’s innovation lies in massive and automatic feature extraction. In contrary with ML, it enables the automatic extraction of millions of potential significant features from any kind of input source. Before the era of Deep Learning, feature extraction was constrained. DL networks are endowed with methods to achieve both feature extraction and feature assessment, thereby reaching to decisions by filtering out some of the irrelevant extracted features.
The advances in computational capabilities over the recent years have catalyzed the implementation of DL, even in tasks where ML has undeniable success already. DL can leverage modern computers and advance further.
In Medical Imaging, the applications of DL have not reached their full potential. Several studies report innovative DL solutions to achieve medical image classification, segmentation, registration and synthesis. However, medical images are a disparate source of information. Unlike conventional images, a medical image contains sensitive information, intricate tissues, bone and organ structures, and ambiguous entities. Subsequently, DL implementations are rigorous and the medical experts are often reluctant to relying their decisions on DL-based frameworks. Despite DL’s progress, it is still acting as a black-box, which raises several concerns among the medical community. Medicine needs explainable solutions, total transparency in decision-making, and accountability. In this context, an increasing amount of research papers are suggesting technical solutions to aid to this matter, which is often met as “interpretability”, or eXplainable Artificial Intelligence (XAI).
Nevertheless, DL is enjoying some praise in medical image detection, segmentation, reconstruction and processing tasks. Those tasks do not involve sophisticated decision-making procedures necessarily, but are a preliminary step to decision-making, which is still offered by the human experts.
From a technical point of view, medical image classification is more naïve compared to detection and segmentation. However, it is more challenging. The reason is that the classification task involves much more than isolating and segmenting objects. It requires some level of reasoning and cognition of the biological, chemical and physical processes to accurately and transparently classify an image, especially when the important high-level features are not visible by the human eye.
The main objective of the thesis is to explore the benefits of DL for various medical imaging tasks, such as image classification and data augmentation. Five sub-studies are analytically conducted, presented and discussed. The main focus of the doctoral thesis is lung cancer, and more specifically, Solitary Pulmonary Nodule (SPN) characterisation, as well as Cardiovascular Diseases, and more specifically, the Coronary Artery Disease (CAD). The study involves data acquired from the PET scanner and the SPECT scanner of the Laboratory of Nuclear Medicine of the University Hospital of Patras.
Firstly, the challenge of SPN malignancy characterisation based-on PET/CT images is addressed. The experiments include classification based on 2D SPN images using transfer learning with Convolutional Neural Networks (CNN), which are the cornerstone of DL methods in imaging tasks. In addition, Generative Adversarial Networks (GANs) are employed to synthesise new SPN representations that serve as additional data and improve the learning horizons of the DL methods. Finally, a 3D DL model is proposed to assess the malignancy of SPNs using volumetric data, which are extracted from multiple 2D PET/CT slices. The results verify that DL models compete with the human eye and show great agreement with the medical experts. In this context, transfer learning and GANs are proved to be remarkable methods for medical image classification and data augmentation tasks.
Next, the study involves experiments for CAD diagnosis based on 2D Polar Maps acquired from a Myocardial Perfusion Imaging (MPI) SPECT system. In this particular study, a multi-input DL framework is designed to handle both image and clinical information for CAD diagnosis. This framework achieved impressive agreement with the human experts.
Finally, the study provides insights on the quality and the usefulness of the extracted image features. To this end, extensive experiments focused on potential biomarker detection and discovery, are conducted. Eleven biomedical imaging datasets are employed to assess DL’s capabilities in classification, feature visualisation, and decision explanation. For those tasks, feature activation maps and other visualisation methods are employed. The experiments revealed that DL methods can mine significant features that could potentially serve as image biomarkers. However, the interpretation of those features’ nature is constrained, because DL models still lack transparency. In addition, it is observed that some features are difficult to reproduce. Those aspects need further attention by the research community in order for DL to evolve into a robust, transparent and reliable strategy for medical image purposes.
The present Doctoral Thesis paves a way towards the development of Medical Decision Support systems for assisting the Nuclear Medicine staff with medical image interpretation. The developed frameworks show significant potential and great agreement ratings with the supervising experts. Finally, they are ready for deployment in real-time routine for further evaluation.
|
|---|
