MITIGATING DATA SCARCITY CHALLENGES IN MEDICAL IMAGING ANALYSIS:ADVANCED LEARNING APPROACHES WITH EMPHASIS ON HEMOPHILIC ULTRASOUND IMAGES

Medical imaging plays a crucial role in hemophilia research and clinical practice, particularly in assessing joint health and bleeding events. Ultrasound (US) imaging is a fundamental tool in the diagnostic process and is currently used to identify when the joint recess is filled with synovial fluid or blood, a condition known as “recess distention” that, if filled with blood, can potentially lead to pathologies and permanent joint damage. In this context, deep learning (DL) techniques can support image acquisition (possibly at the point-of-care) and enhance the capabilities of computer-aided diagnosis (CAD) systems. However, the lack of labeled training data makes the effective utilization of DL techniques in the medical domain impractical, leading to suboptimal performance in various imaging tasks, such as classification, detection, and segmentation. This thesis investigates the application of advanced DL methods to overcome this challenge and enhance the analysis of medical images in the context of data scarcity, with a particular focus on ultrasound images in hemophilia research. Specifically, this thesis addresses three main challenges: a limited number of total samples, class imbalance, and the adaptation of trained models to different domains (such as knee to elbow transfer). To address the problem of the limited number of total samples, this research investigates the adoption of transfer learning and proposes a new multi-task model to effectively utilize limited labeled data and improve model generalization. Concerning the issue of imbalanced data, the thesis explores anomaly detection techniques that can be trained on normal samples only. However, as demonstrated experimentally, classic unsupervised anomaly detection methods fail in this domain due to the intrinsic variability of musculoskeletal ultrasound images. Therefore, the thesis introduces a new weakly supervised anomaly detection framework that enhances classification and segmentation performance, requiring only the recess location as a weak annotation. To address the third issue, we investigate two domain adaptation frameworks to adapt a model trained on knee images to also identify the distension on elbow images. We first explore test-time adaptation techniques and then introduce a new contrastive feature test-time training approach. By developing and integrating these DL techniques into an existing CAD system, this thesis aims to provide insights into effectively leveraging limited labeled data in medical imaging research, thereby advancing the understanding and management of rare and complex medical conditions.