Markedly Enhanced Analysis of Mass Spectrometry Images Using Weakly Supervised Machine Learning

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Data Science

Background:

• Machine learning, including supervised and unsupervised algorithms, is widely used for analyzing mass spectrometry imaging (MSI) data, such as time-of-flight secondary ion mass spectrometry (ToF-SIMS).
• Existing methods excel at single-pixel analysis but struggle with weakly supervised problems where labels are only available at the image level, not for individual pixels.
• Unsupervised methods do not leverage available image-level labels, highlighting a gap in MSI data analysis techniques.

Purpose of the Study:

• To introduce a novel method specifically designed for weakly supervised mass spectrometry imaging (MSI) data analysis.
• To adapt a dual-stream multiple instance learning (MIL) approach from computational pathology for MSI applications.
• To demonstrate the capability of the MIL method in revealing spatial-spectral characteristics for distinguishing MSI image classes.

Main Methods:

• A dual-stream multiple instance learning (MIL) framework was adapted and applied to MSI data.
• An information entropy-regularized attention mechanism was employed to identify characteristic pixels within MSI images.
• The identified pixels were utilized to extract characteristic mass spectra for classification.

Main Results:

• The MIL method successfully revealed spatial-spectral characteristics differentiating MSI image classes.
• Proof-of-concept studies using printed ink samples (ToF-SIMS) and a mixed powder sample demonstrated the method's efficacy.
• The approach proved capable of handling weakly supervised MSI data, outperforming methods that cannot utilize image-level labels.

Conclusions:

• The developed dual-stream MIL method offers a powerful new approach for analyzing weakly supervised MSI data.
• This technique enhances the understanding of subtle spatial-spectral features within MSI datasets.
• The findings suggest broad applicability of the MIL method across various MSI applications, improving data interpretation and classification.