Causal Responsibility Based Explainable AI for Vibrational Spectroscopy Applied to Oral FTIR and Oesophageal Raman...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Spectroscopy

Background:

• Deep learning models enhance biomedical vibrational spectroscopy but suffer from a lack of transparency, creating a "black-box" problem.
• This opacity impedes clinical trust and regulatory compliance, which mandate explainability for medical applications.
• Existing post-hoc Explainable AI (XAI) methods often lack rigorous assessment for clinical standards.

Purpose of the Study:

• To introduce Spec-ReX, a novel XAI method grounded in actual causality.
• To identify spectral features most responsible for classifications in biomedical spectroscopy.
• To rigorously evaluate Spec-ReX against other XAI methods across diverse datasets.

Main Methods:

• Developed Spec-ReX utilizing the theory of actual causality.
• Applied Spec-ReX to identify key spectral features driving classifications.
• Assessed Spec-ReX performance on in silico, in vitro, and ex vivo spectral datasets.
• Compared Spec-ReX against commonly used XAI techniques.

Main Results:

• Spec-ReX effectively identifies causal spectral features for classifications.
• The method demonstrates robust performance across increasing data complexity.
• Rigorous assessment provides a nuanced comparison with existing XAI approaches.

Conclusions:

• Spec-ReX offers a causally-principled approach to XAI in biomedical spectroscopy.
• The method enhances transparency and trustworthiness of deep learning models for clinical use.
• This work sets a higher standard for evaluating XAI in medical applications.