MSIMG: A Density-Aware Multi-Channel Image Representation Method for Mass Spectrometry.

Area of Science:

• Analytical Chemistry
• Bioinformatics
• Computer Vision

Background:

• Extracting key features from complex mass spectrometry (MS) data for phenotype classification is challenging.
• Conventional methods like peak lists or grid-based imaging lead to information loss and reduced signal integrity.
• This compromises the performance of deep learning models used in clinical diagnostics.

Purpose of the Study:

• To introduce MSIMG, a novel data representation framework for mass spectrometry data.
• To improve information fidelity and signal integrity in MS data for enhanced deep learning applications.
• To develop a more effective data representation for robust clinical diagnostic models.

Main Methods:

• Developed MSIMG, a framework inspired by computer vision object detection.
• Implemented a data-driven, density-peak-centric patch selection strategy using density map estimation and non-maximum suppression.
• Transformed raw MS data into a multi-channel image representation with higher information fidelity.

Main Results:

• MSIMG significantly outperformed traditional peak list and grid-based MetImage approaches on two public clinical MS datasets.
• The MSIMG framework demonstrated a more information-dense and discriminative data representation.
• Content-aware patch selection proved crucial for improving deep learning model performance.

Conclusions:

• Data representation critically impacts deep learning model performance in MS data analysis.
• The MSIMG framework offers a superior paradigm for representing MS data for deep learning.
• Applying computer vision strategies to analytical chemistry data shows great potential for advancing clinical diagnostics.