Deep Learning Approach for Dynamic Sparse Sampling for High-Throughput Mass Spectrometry Imaging.

Area of Science:

• Analytical Chemistry
• Computational Imaging
• Biomedical Engineering

Background:

• Compressed sensing methods often struggle with integrating stochastic processes.
• Traditional sampling techniques can be time-consuming for high-resolution imaging.
• Supervised Learning Approach for Dynamic Sampling (SLADS) was developed to address these issues.

Purpose of the Study:

• To introduce an enhanced Supervised Learning Approach for Dynamic Sampling (SLADS), termed Deep Learning Approach for Dynamic Sampling (DLADS).
• To reduce sample acquisition times while maintaining high-fidelity reconstructions in Mass Spectrometry Imaging (MSI).
• To improve the efficiency and applicability of dynamic sampling for molecular imaging.

Main Methods:

• Developed DLADS, an enhancement of SLADS, utilizing deep learning and convolutional layers.
• Adjusted training set creation methodology to mitigate stretching artifacts.
• Applied DLADS to Nanospray Desorption ElectroSpray Ionization (nano-DESI) Mass Spectrometry Imaging (MSI) of mouse tissues.

Main Results:

• Achieved sample acquisition time reductions of approximately 70-80% compared to traditional rectilinear scanning.
• Demonstrated high-fidelity reconstructions for dimensionally asymmetric, high-resolution molecular images.
• Showcased effective generalization capabilities of DLADS with dissimilar training and testing data.

Conclusions:

• DLADS significantly enhances experimental throughput for nano-DESI MSI.
• The deep learning approach eliminates the need for manual feature extraction, leveraging spatial relationships.
• DLADS offers a more efficient and robust solution for dynamic sampling in high-resolution molecular imaging.