An explainable three dimension framework to uncover learning patterns: A unified look in variable sulci recognition

3D explanations are essential for medical AI: The paper demonstrates that two-dimensional interpretations are insufficient for complex medical imaging tasks like neuroimaging, requiring three-dimensional global explanations to accurately represent how AI models make decisions about brain structures.

Combining multiple explanation methods improves reliability: By integrating three different explanation techniques (Shape analysis, GradCam, and SHAP), the framework reduced variability between methods and enhanced the "faithfulness and reliability of global explanations," suggesting that no single explanation method alone provides complete trustworthiness.

AI can identify clinically relevant brain regions: The framework successfully identified specific brain areas (posterior temporal, internal parietal regions, cingulate region, and thalamus) associated with the paracingulate sulcus, a structure linked to psychosis, demonstrating that explainable AI can surface medically meaningful insights beyond just making accurate predictions.

Explainability requires accounting for natural human variation: The framework incorporates "cohort-specific variability," meaning responsible AI in medical imaging must recognize that healthy brains naturally differ from person to person, and explanations should reflect this biological reality rather than treating all brains as identical.

Transparency enables validation of AI reasoning: By providing clear, comprehensible explanations of which brain regions the model focused on, clinicians and researchers can verify whether the AI's decision-making process aligns with established medical knowledge, rather than treating the model as a "black box."