Exploring the Spectrum of Supervision in Medical Image Analysis: From Fully Supervised to Semi-supervised and Unsupervised Approaches

Deep learning techniques have gained tremendous success across various domains with the rise of large datasets and the advancement of computational resources over the past few years. However, when adopting deep learning techniques in medical image analysis, it is a common challenge that limited datasets are available to train the model, and acquiring accurately annotated datasets has become more expensive and labour-intensive compared to other domains. In tackling these obstacles, it is worth exploring deep learning techniques alongside various supervision techniques for medical image analysis - such approaches have the potential to offer promising solutions for many real-world applications in the field. This thesis investigates different levels of supervision techniques used in medical image analysis, including fully supervised, semi-supervised, and unsupervised approaches, with a focus on lung cancer classification as an experimental case study. This study discusses the effectiveness and constraints of different supervision techniques through a comprehensive analysis of various models, algorithms and their applications for classification in lung cancer datasets. Several novel classification approaches, built on top of existing supervision techniques, were proposed, implemented and evaluated using a comprehensive analysis of the lung cancer dataset. The outcome of this research also uncovers significant insights into the suitability of each approach, highlighting that while fully supervised methods demonstrate high performance, semi-supervised and unsupervised techniques offer valuable adaptability and efficiency, particularly in scenarios with limited labelled medical datasets. This research also highlights the potential of combining contrastive learning and generative methods in revolutionising medical image analysis, especially for diseases lacking sufficient datasets with accurate labels. It not only enhances our understanding of the supervision spectrum in lung cancer classification but also sets a foundation for future research in the field, potentially influencing diagnostic processes and aiding in developing robust and adaptable medical image classification systems.