Spatial-Spectral Deep Learning for Prostate Cancer Tissue Classification in Infrared Spectroscopy.

Area of Science:

• Biomedical Engineering
• Spectroscopy
• Artificial Intelligence

Background:

• Hyperspectral imaging from infrared (IR) spectroscopy generates rich data for deep learning.
• Convolutional neural networks may exhibit spatial bias in processing this data.
• Prostate cancer tissue classification is a key application area.

Purpose of the Study:

• Compare deep learning classifiers for IR spectroscopy data.
• Evaluate the impact of spectral dimension compression (bottleneck) on performance.
• Investigate the relationship between model architecture, spatial bias, and classification accuracy.

Main Methods:

• Applied various deep learning models, including modified Vision Transformers, to IR hyperspectral images.
• Tested the effect of a spectral bottleneck (16 features) on model performance.
• Analyzed the correlation between model spatial receptive field and classification outcomes.

Main Results:

• Highest classification performance was achieved by a modified Vision Transformer model.
• Model spatial receptive field strongly correlated with classification success.
• Limited correlation found between spectral information and deep learning performance; a 16-feature bottleneck had negligible impact.

Conclusions:

• Tissue classification relies on a limited set of spectral features, not broad spectral information.
• Spatial features are more critical than spectral depth for current deep learning classification tasks.
• Current success in tissue classification may be an inadequate benchmark for developing deep learning models that leverage spectral data.