Causal responsibility based explainable AI for vibrational spectroscopy applied to oral FTIR and oesophageal Raman...

Area of Science:

• Biomedical applications of vibrational spectroscopy
• Artificial Intelligence in Spectroscopy

Background:

• Deep learning models in vibrational spectroscopy lack transparency.
• Classifications are often provided without identifying key spectral features driving decisions.

Purpose of the Study:

• Introduce Spec-ReX, a novel explainable AI method for vibrational spectroscopy.
• Compare Spec-ReX with SHAP and Grad-CAM using diverse spectral datasets.

Main Methods:

• Developed Spec-ReX, a causal responsibility-based explainable AI approach.
• Evaluated methods on synthetic spectra, biochemical mixtures, and tissue spectra (oral FTIR, oesophageal Raman).
• Assessed performance using metrics like Intersection over Union and attribution concentration.

Main Results:

• Spec-ReX demonstrated superior ground truth localization in definitive in silico experiments.
• Spec-ReX showed robustness in ambiguous in silico and in vitro biochemical experiments.
• Attribution patterns varied and were architecture-dependent in ex vivo datasets.

Conclusions:

• Spec-ReX generates sparse, specific, and model-causal responsibility maps for spectral classifiers.
• Explanations reflect trained model behavior, not biological causality.
• Clinical utility requires further validation through user-centered studies.