Cracking the clinical code: A scoping review on mechanistic interpretability in medical report generation
Dost Muhammad, Muhammad Salman, Malika Bendechache · Computational and Structural Biotechnology Reports · 2025
Medical report generation (MRG) bridges computer vision and natural language processing by translating clinical images into diagnostic text. Although encoder–decoder models have significantly advanced this domain, their opaque internal processes continue to impede clinical trust and regulatory acceptance. This survey presents a comprehensive methodological review of interpretability techniques applied to MRG, with particular emphasis on mechanistic interpretability. This emerging paradigm shifts the focus from attribution-based methods towards uncovering how specific model components such as neurons, attention heads, and residual connections encode diagnostic logic. We categorise core techniques including activation patching, causal tracing, circuit decomposition, and concept bottleneck probing, and propose a structured taxonomy spanning modalities, architectures, and evaluation strategies. Unlike conventional post hoc attribution methods, mechanistic approaches aim to provide causal, component-level explanations that are predictive and verifiable. We identify major challenges in this area, including cross-modal entanglement, the absence of internal ground-truth benchmarks, and the need for clinically aligned evaluation protocols. To address these limitations, we outline future research directions involving interpretable-by-design architectures, interactive explanation refinement through clinician feedback, and standardised metrics for mechanistic fidelity. By synthesising insights from machine learning, neuroscience, and radiology, this survey establishes a foundation for building MRG systems that are transparent, trustworthy, and clinically useful.