Lars Chen: Counterfactual Inference with Deep Structural Causal Modeling for 3D Clinical Neuroimaging

BCCN Berlin / Technische Universität Berlin

Abstract

Understanding neurological disease processes through imaging often involves posing counterfactual questions, such as, ``How would this patient's brain scan appear if they had received a different treatment?" These hypothetical scenarios, called counterfactuals, are critical for guiding clinical decisions but challenging to infer since only one outcome is observable. Tools from causal inference, such as structural causal models (SCMs), enable counterfactual inference using observational data by modeling relationships between variables using explicit functional mechanisms. However, they face scalability issues when applied to complex, high-dimensional data. Deep SCMs (DSCMs) enable efficient counterfactual inference, e.g. for imaging applications, by approximating SCMs with deep learning components. While demonstrating promising results for modeling neurological diseases, DSCMs have been limited to 2D brain images, preventing direct volumetric evaluation. Moreover, validation of DSCMs has been constrained to synthetic toy datasets, limiting their clinical applicability.

This thesis addresses these challenges by generating realistic 3D brain images, using a pre-trained diffusion model, conditioned on covariates sampled from a ground-truth SCM that emulates a disease scenario. Next, the DSCM framework is extended from 2D to 3D inference and is designed to align with the ground-truth SCM. Finally, the trained DSCM is validated through volumetric analysis to assess how well causal relationships are reflected in the generated images. The results demonstrate that the DSCM can generate plausible 3D counterfactual brain images that preserve high-level features. Volumetric analysis shows the interventions follow intended causal relationships but are constrained by low-resolution outputs and discrepancies in volumetric labels. Despite these limitations, this thesis establishes a foundation for robust verification of brain counterfactuals and advances the field toward clinically applicable models.

 

 

Additional information:

Master thesis defense

 

Organized by:

Prof. Dr. Kerstin Ritter & Prof. Dr. Klaus-Robert Müller