New iFIM Method Analyzes Medical Image Classifiers Without Black-Box Guesswork
A principled way to peek inside deep neural nets for medical image classification...
Deep neural networks excel at medical image classification but remain opaque black boxes. Current post-hoc interpretability methods produce heuristic visualizations that only indirectly relate to the classifier’s predictive distribution. To address this, Sourya Sengupta and Mark A. Anastasio introduce a local sensitivity analysis framework built on the input-dependent Fisher Information Matrix (iFIM). The iFIM mathematically characterizes how infinitesimal image perturbations alter the classifier’s output distribution, providing a rigorous alternative to pixel-wise attribution maps.
A key innovation is the Gram-matrix formulation, which recovers the iFIM’s nonzero eigenspace without explicitly constructing the full image-dimensional Fisher matrix—making the method computationally tractable. Projecting an input image onto the leading iFIM eigenvectors yields a high-sensitivity component and an orthogonal low-sensitivity component. Extensive perturbation experiments across multiple classifier architectures (on controlled and clinical medical imaging tasks) confirm that the high-sensitivity component is far more coupled to changes in predictive confidence and classification performance. This positions iFIM as a principled tool for understanding local decision sensitivity and complementing existing interpretability methods in medical AI.
- Uses Gram-matrix formulation to recover nonzero iFIM eigenspectrum without full image-dimensional matrix
- Perturbation experiments show high-sensitivity components strongly correlate with predictive confidence changes
- Evaluated on multiple classifier architectures for both controlled and clinical medical image classification tasks
Why It Matters
Provides a rigorous, model-intrinsic tool to validate medical AI decisions beyond heuristic heatmaps.