Core Concepts
Developing a semi-supervised deep causal generative model to enhance medical image analysis by addressing missing labels in clinical data.
Abstract
Abstract: Introduces the need for causal generative models in medical image analysis.
Introduction: Discusses challenges in deploying deep learning models and the importance of causal considerations.
Background: Explains the Structural Causal Model and the do-operation for interventions.
Methodology: Outlines the model structure, loss functions, and training process.
Experiments: Evaluates the model on Morpho-MNIST and MIMIC-CXR datasets, showcasing the effectiveness of counterfactual generation.
Conclusion: Highlights the contributions of the study and suggests future research directions.
Stats
"60k training, 10k test samples" - Morpho-MNIST dataset
"50k training, 20k test samples" - MIMIC-CXR dataset
Quotes
"Introduce a semi-supervised deep causal generative model."
"Our approach uses unlabelled and partially labelled data effectively."
"The key practical contribution of this work is that it enables training causal models on clinical databases where patient data may have missing labels."