MSInet: A Self-Supervised CNN Framework Integrating Global and Local Context for Robust Mass Spectrometry Imaging...

Area of Science:

• Biomedical Imaging
• Computational Biology
• Data Science

Background:

• Mass spectrometry imaging (MSI) provides label-free molecular mapping but faces challenges in spatial segmentation due to data complexity and tissue heterogeneity.
• Existing unsupervised clustering methods often fail to incorporate spatial information, leading to inaccurate and fragmented segmentation results.

Purpose of the Study:

• To introduce MSInet, a self-supervised deep learning framework designed for robust and annotation-free spatial segmentation of mass spectrometry imaging data.
• To improve the accuracy and biological relevance of MSI segmentation by integrating global and local contextual information.

Main Methods:

• Developed MSInet, a convolutional neural network framework employing patch-wise contrastive learning for global relationships and superpixel-guided refinement for local spatial consistency.
• Utilized a dual-consistency training strategy to enhance both global context awareness and local boundary precision.
• Evaluated MSInet on MALDI-MSI of mouse brain, DESI-MSI of renal tumor, and synthetic datasets.

Main Results:

• MSInet significantly outperformed state-of-the-art methods in segmentation accuracy and biological fidelity across diverse MSI datasets.
• Achieved high performance on simulated data (Adjusted Rand Index = 0.89, Normalized Mutual Information = 0.86), showing substantial improvement over baseline methods.
• Accurately delineated complex anatomical structures in brain tissue and distinguished critical regions in renal tumors, aligning closely with histological data.
• Demonstrated robustness to noise inherent in MSI data.

Conclusions:

• MSInet provides a powerful and scalable solution for accurate, biologically meaningful MSI segmentation by effectively integrating global and local contextual modeling.
• The self-supervised, annotation-free nature of MSInet makes it broadly applicable to spatial omics and various biomedical research areas.
• This framework represents a significant advancement in leveraging deep learning for complex biological data analysis.