Modern vision and vision–language models (VLMs) achieve remarkable perceptual performance, yet their internal representations often misalign with human-understandable concepts, clinical reasoning, or the causal structure of data. Such misalignment limits trust, generalization, and safety – particularly in high-stakes domains such as medical imaging. This thesis proposes a comprehensive framework for model alignment, developing methods that bring model representations closer to clinically meaningful, causally grounded, and task-adaptive concepts across diverse visual tasks.

First, I introduce biomarker-grounded alignment for lung ultrasound (LUS), where domain-informed interpretable biomarkers serve as anchors to structure deep model representations. I develop methods that disentangle anatomical, morphological, and artifact-level biomarkers and demonstrate that these aligned representations improve interpretability while matching or exceeding fully supervised baselines across diagnostic and severity scoring tasks.

Second, I propose a causal feature selection framework based on Markov blanket discovery to identify minimal yet causally relevant feature subsets across medical and non-medical datasets. By uncovering natural experiments embedded in observational data, the method reveals features that are inherently robust, reduces spurious correlations, and provides theoretical and empirical evidence for improved generalization and interpretability.