Artifact Imputation in Federated learning for Robust Medical Image Analysis

다국어 초록 (Multilingual Abstract)

This thesis investigates a unified framework for mask-free artifact imputation in federated learning environments, specially addressing the unique challenges of medical image analysis under heterogeneous and noisy data conditions. Medical imaging workflows are frequently affected by acquisition artifacts, geometric occlusions, and complex real-world corruptions that degrade downstream diagnostic performance. In decentralized settings, where each institution maintains its own data and computational constraints, such degradations interact with the intrinsic non-IID nature of federated learning and lead to severe instability during model training. To address these issues, it is essential to employ inpainting models that are not only accurate but also computationally efficient, as federated clients often differ widely in hardware capability. Motivated by these constraints, this thesis adopts a wavelet-based backbone whose multi-resolution decomposition provides strong representational efficiency with far fewer parameters than conventional CNN, GAN, or transformer-only architectures. Building on this lightweight foundation, we introduce Swin WavePaint, a wavelet–transformer hybrid inpainting model that integrates wavelet-driven multi-resolution token mixing with Swin Transformer modules to enable richer semantic reasoning while maintaining the efficiency required for practical federated deployment. Furthermore, the method supports end-to-end artifact detection and inpainting, as well as federated optimization, without relying on externally provided artifact masks. Comprehensive experiments are conducted across three complementary scenarios: (1) geometric mask imputation on CelebA-HQ dataset, (2) large-scale federated inpainting and classification on PathMNIST under centralized and decentralized configurations, and (3) restoration of nine real-world whole-slide-image corruptions as defined in digital pathology robustness literature. Across all settings, the proposed Swin WavePaint model consistently outperforms the baseline WavePaint architecture in parameter efficiency, Learned Perceptual Image Patch Similarity and downstream classification accuracy. The results demonstrate that artifact imputation is a critical component for robust medical image learning in both centralized and federated environments, and the findings highlight its potential as a fundamental preprocessing paradigm for real-world medical AI.

목차 (Table of Contents)

• CHAPTER 1. Introduction 1
• CHAPTER 2. Related Works 5
• 2.1 Image Inpainting 5
• 2.2 Wavelet-Based Inpainting 6
• 2.3 Federated Learning for Medical Imaging 7
• 2.4 Federated Data Imputation 8
• 2.5 Real-World Digital Pathology Artifacts 9
• CHAPTER 3. Methodology 10
• 3.1 Overall Framework 10
• 3.2 Swin WavePaint 12
• 3.3 Classification Model 16
• 3.4 Federated Learning Framework 17
• 3.5 Loss Functions and Training Strategy 18
• CHAPTER 4. Experimental Analysis 20
• 4.1 Datasets 20
• 4.2 Artifact Generation 21
• 4.3 Model Configurations and Training Protocols 24
• 4.4 Experiment 1: CelebA-HQ Inpainting Benchmark 26
• 4.5 Experiment 2: PathMNIST Inpainting Benchmark 30
• 4.6 Experiment 3: Real-World WSI Artifact Restoration 33
• CHAPTER 5. Conclusion 37
• CHAPTER 6. References 39