Medical Applications with Disentanglements. First MICCAI Workshop, MAD 2022 Held in Conjunction with MICCAI 2022 Singapore, September 22, 2022 Proceedings

Description

Preface Organization Contents Introduction Applying Disentanglement in the Medical Domain: An Introduction for the MAD Workshop 1 Introduction 2 Generative Models 3 Disentanglement 3.1 Definitions of Disentanglement 3.2 The Different Types of Biases 4 Challenges for Medical Applications 5 Medical Applications 6 Future Directions 6.1 Causality and Disentanglement 6.2 Evaluating Disentanglement 7 Conclusion References GAN-Based Approaches HSIC-InfoGAN: Learning Unsupervised Disentangled Representations by Maximising Approximated Mutual Information 1 Introduction 2 Methodology 2.1 InfoGAN 2.2 Hilbert-Schmidt Independence Criterion (HSIC) 2.3 HSIC-InfoGAN 3 Experiments 3.1 Implementation Details 3.2 Results 3.3 Strategy for Hyperparameter Tuning 4 Discussion References Implicit Embeddings via GAN Inversion for High Resolution Chest Radiographs 1 Introduction 2 Methods 3 Experiments 3.1 Image Compression and Quality of Reconstruction 3.2 Disentanglement in Latent Space 3.3 Guided Image Manipulation 3.4 Proximity Sampling 4 Outlook and Conclusion References Disentangled Representation Learning for Privacy-Preserving Case-Based Explanations 1 Introduction 2 Related Work 2.1 Image Anonymization 2.2 Deep Generative Models 3 Proposed Methodology 3.1 Generative Module 3.2 Identity Module 3.3 Explanatory Module 4 Experiments and Results 4.1 Identity Recognition and Disease Recognition 4.2 Image Anonymization 4.3 Generation of Counterfactual Explanations 5 Conclusions References Autoencoder-Based Approaches Instance-Specific Augmentation of Brain MRIs with Variational Autoencoders 1 Introduction 2 Methods 2.1 Disentangling Shape from Appearance 3 Experiments 3.1 Data 3.2 Training Details 3.3 Augmentation Schemes 3.4 Results and Discussion 4 Conclusion References Low-Rank and Sparse Metamorphic Autoencoders for Unsupervised Pathology Disentanglement 1 Introduction 2 Methods 2.1 Guided Filter Regularized Metamorphic Autoencoder 2.2 Low-Rank and Sparse Image Decomposition for Normal/Abnormal Disentanglement 3 Experiments and Results 4 Discussion and Conclusion References Training -VAE by Aggregating a Learned Gaussian Posterior with a Decoupled Decoder 1 Introduction 2 -VAE 3 The Antagonistic Mechanism of the Reconstruction Loss and KLD Loss in -VAE 3.1 Information Theory Perspective 3.2 Machine Learning Perspective 4 Aggregate a Learned Gaussian Posterior with a Decoupled Decoder 5 Application to Skull Reconstruction and Shape Completion 5.1 Training Curves 5.2 Skull Reconstruction and Skull Shape Completion 6 Discussion and Conclusion A VAE Training Curve (1200 Epochs) under =100 B AE-Based Skull Shape Completion C Matrix Notation for DKL(1) References Normalizing-Flow-Based Approaches Disentangling Factors of Morphological Variation in an Invertible Brain Aging Model 1 Introduction 2 Methods 2.1 Invertible Brain Aging Model – iBAM 2.2 Adding Sex as Another Supervised Factor 2.3 Ordering iBAM's Identity Latent Space Dimensions 3 Experiments and Results 4 Conclusion References Comparision A Study of Representational Properties of Unsupervised Anomaly Detection in Brain MRI 1 Introduction 2 Approaches for Modeling Anomaly 2.1 Selected Methods 2.2 Hierarchy of Properties 3 Experimental Setup 4 Observations 5 Inferences and Discussion 6 Conclusion A Appendix A.1 VAE A.2 FactorVAE A.3 GLOW A.4 SSAE References Author Index