Dimensionality reduction for deep learning in infrared microscopy: a comparative computational survey.

Area of Science:

• Biomedical optics
• Computational pathology
• Machine learning in medicine

Background:

• Infrared microscopy offers label-free molecular and spatial information for tissue analysis.
• Classifying disease status using both spatial and molecular data presents a significant challenge.
• Dimensionality reduction of spectral data is a strategy to simplify analysis and improve machine learning accessibility.

Purpose of the Study:

• To compare various dimensionality reduction techniques for infrared microscopy spectral data.
• To evaluate the impact of dimensionality reduction on cancer identification in colon carcinoma.
• To assess the reliance of convolutional neural networks on spatial versus spectral information for disease classification.

Main Methods:

• Application of multiple dimensionality reduction approaches to high-dimensional pixel spectra from infrared microscopy.
• Training convolutional neural networks on both full and dimensionality-reduced spectral data.
• Comparative analysis of classification performance using different spectral data representations.

Main Results:

• Dimensionality reduction resulted in minimal differences in convolutional neural network performance compared to using full spectral data.
• Convolutional neural networks demonstrated a strong tendency to prioritize spatial information over spectral information for disease classification.
• The effectiveness of cancer identification in colon carcinoma was maintained even with reduced spectral complexity.

Conclusions:

• Dimensionality reduction is a viable strategy for simplifying infrared microscopy data for machine learning applications.
• Convolutional neural networks in this context primarily leverage spatial features for accurate disease status classification.
• Future research can explore optimizing the balance between spatial and spectral feature extraction for enhanced diagnostic capabilities.