Input-Dependent Fisher Information for Local Sensitivity Analysis of Medical Image Classifiers
Researchers have introduced a new framework for analyzing the local sensitivity of medical image classifiers using the input-dependent Fisher Information Matrix (iFIM). This method characterizes how a classifier's predictions change with small input perturbations. By employing a Gram-matrix formulation, the framework avoids computing the full Fisher matrix and instead focuses on the leading iFIM eigenspace to identify high-sensitivity components. These components offer an intrinsic description of local predictive sensitivity, complementing existing attribution-based interpretability tools. AI
IMPACT Provides a principled tool for analyzing local decision sensitivity in medical imaging AI, complementing existing interpretability methods.