Explainable AI-Based Feature Selection Approaches for Raman Spectroscopy.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Spectroscopy

Background:

• Raman spectroscopy offers non-invasive tissue analysis for accurate cancer detection.
• Machine learning automates pattern discovery but faces challenges with high-dimensional Raman data.
• Model explainability is crucial for integrating AI in medical diagnostics, necessitating effective feature reduction.

Purpose of the Study:

• To introduce novel, explainable deep learning-based feature selection methods for Raman spectroscopy.
• To compare these new methods against established techniques across multiple datasets and classifiers.
• To address the challenge of feature reduction while minimizing information loss in Raman data.

Main Methods:

• Developed two feature selection methods using explainable deep learning: Convolutional Neural Networks (CNNs) with GradCam and Transformers with attention scores.
• Extracted features using GradCam for CNNs and attention scores for Transformers.
• Evaluated feature performance against established methods using four classifiers and three real-world Raman spectroscopy datasets.

Main Results:

• Explainable deep learning methods achieved comparable accuracy to traditional approaches using only 10% of features.
• CNNs with GradCam and Random Forest showed top performance with 5-20% feature retention.
• LinearSVC with L1 penalization was highly accurate with only 1% of features, while the CNN-GradCam approach demonstrated the highest average accuracy.

Conclusions:

• No single feature selection method is universally optimal for all Raman spectroscopy applications.
• The proposed CNN-GradCam approach shows strong potential for accurate and explainable feature selection.
• Assessing multiple feature selection alternatives is recommended for each specific application to ensure optimal performance.