Overcoming Data Scarcity Challenges in Medical Deep Learning: Innovations and Strategies

Abstract

Medical data's scarcity, multimodal nature, and high dimensionality complicate training robust machine learning (ML) models. This work tackles these challenges with novel techniques for efficient learning processes, dataset development with automatic annotation, and multimodal data fusion. First, it addresses the inefficiency of learning systems when handling small and imbalanced datasets, proposing a new method that enhances data efficiency and reduces sample complexity. Second, it introduces a new dataset and automatic segmentation model for diagnosing musculoskeletal soft tissue tumors (MSTTs), facilitating machine-assisted data annotation and supporting the creation of more accurate ML models. Third, it proposes diagnostic models for MSTTs using single modalities and multimodal fusion strategies, demonstrating that integrating various data types (e.g., magnetic resonance images and clinical data) results in more sensitive diagnostic tools. In summary, this thesis provides comprehensive solutions to critical issues in applying ML to limited medical data, aiming to significantly improve ML models' performance and applicability in clinical practice through efficient learning techniques, novel datasets, and advanced multimodal strategies. First, we address the inefficiency of learning systems in dealing with small and imbalanced datasets prevalent in medical fields. Our novel learning method enhances data efficiency and reduces sample complexity, mitigating the impact of data imbalance on model training. Second, we develop a new dataset and an automatic segmentation model for diagnosing musculoskeletal soft tissue tumors (MSTTs), facilitating machine-assisted data annotation. This dataset includes multi-modal data and aims to support the creation of accurate and efficient ML models. Third, we propose models for MSTT diagnosis using both single modalities and multimodal fusion strategies. We discover that the integration of various data types (i.e., magnetic resonance (MR) images with clinical data) can yield more sensitive diagnostic tools. Our methods provide notable advancements in both the performance and applicability of ML models in clinical practice. In summary, this thesis presents comprehensive solutions to several critical issues in applying ML to limited medical data. By proposing efficient learning techniques, creating a novel dataset for automatic annotation, and developing advanced multimodal fusion strategies, this work aims to significantly improve the use of ML models in medical diagnostics and patient care.