Accurate and interpretable classification of microspectroscopy pixels using artificial neural networks

Area of Science:

• Spectroscopy
• Machine Learning
• Materials Science

Background:

• Pixel-level material classification from microspectroscopy presents challenges in identifying discriminatory spectral features.
• Developing accurate and interpretable models linking spectra to class labels is crucial.

Purpose of the Study:

• To design a supervised classifier using a tandem of Artificial Neural Network (ANN) models for accurate and interpretable pixel-level material classification.
• To integrate classification rule extraction methods to reduce model complexity and enhance interpretability.

Main Methods:

• A supervised classifier was designed using a tandem of ANN models tailored to microspectroscopy data.
• Each ANN model was designed based on the hypothesis that discriminatory features are linear combinations of spectra.
• Classification rule extraction methods were meshed with ANN models for interpretability.

Main Results:

• The proposed method achieved an average accuracy of 85% using the DICE pixel label similarity metric on broadband coherent anti-Stokes Raman scattering (BCARS) microscopy cell images.
• The generated classification rules showed an average similarity of 96% to expert-created rules, measured by the vector cosine metric.
• The tandem ANN and decision rule configuration resulted in accurate and interpretable classification.

Conclusions:

• The developed tandem ANN and decision rule model effectively classifies materials at the pixel level in microspectroscopy data.
• The approach successfully balances high classification accuracy with model interpretability.
• This method offers a promising tool for analyzing complex spectral data in scientific research.