Fragment-Fusion Transformer: Deep Learning-Based Discretization Method for Continuous Single-Cell Raman Spectral...

Area of Science:

• Biophysics
• Spectroscopy
• Computational Biology

Background:

• Raman spectroscopy is crucial for single-cell biochemical analysis.
• Continuous, high-dimensional Raman spectral data hinders deep learning applications due to lack of discretization.

Purpose of the Study:

• To develop a novel deep learning framework for analyzing continuous Raman spectral data.
• To address the limitations of applying deep learning to discrete sequences.

Main Methods:

• Proposed a fragment-fusion transformer model integrating spectral fragmentation and feature fusion.
• Employed transformer blocks for intrafragment feature extraction and a pyramid structure for interfragment fusion.
• Utilized intrinsic spectral characteristics for data discretization.

Main Results:

• Achieved 94.5% spectral recognition accuracy, outperforming non-fragmented methods by 4%.
• Demonstrated 4.4% higher accuracy than the best-performing CNN model.
• Pyramidal fusion enhanced information gain by 9.24x and information entropy by 13x.

Conclusions:

• The fragment-fusion transformer offers a generalizable framework for continuous spectral data discretization.
• This method enhances the analysis of intrinsic spectral information for biochemical monitoring.
• Paves the way for advanced deep learning applications in Raman spectroscopy.