SagMSI: A graph convolutional network framework for precise spatial segmentation in mass spectrometry imaging

Abstract

Background

Mass Spectrometry Imaging (MSI) is a label-free imaging technique used in spatial metabolomics to explore the distribution of various metabolites within biological tissues. Spatial segmentation plays a crucial role in the biochemical interpretation of MSI data, yet the inherent complexity of the data—characterized by large size, high dimensionality, and spectral nonlinearity—poses significant analytical challenges in MSI segmentation. Although deep learning approaches based on convolutional neural networks (CNNs) have shown considerable success in spatial segmentation for biomedical imaging, they often struggle to capture the comprehensive structural information of MSI data.

Results

We propose SagMSI, an unsupervised graph convolution network (GCN)-based segmentation strategy that combines spatial-aware graph construction of MSI data with a GCN module within a deep neural network. This approach enables flexible, effective, and precise spatial segmentation. We applied SagMSI to both simulated data and various MSI experimental datasets and compared its performance against three commonly used segmentation methods, including t-SNE + k-means, a pipeline implemented by the R package Cardinal, and a CNN-based segmentation method. Visual comparisons with histological images and quantitative evaluations using the silhouette coefficient and adjusted rand index demonstrate that SagMSI excels in segmenting complex tissues, revealing detailed sub-structures, and delineating distinct boundaries of sub-organs with minimal noise interference. The integration of graph-based neural networks with spatially structural information offers deeper insights into spatial omics.

Significance

The MSI data is modelled by graph structure so as to incorporate the biomolecular profiling and spatial adjacency within neighboring pixels. The GCN framework generates meaningful pixel representations by learning local and global contextual information through the graph-based structure, thus enabling precise segmentation of MSI. The approach demonstrated high flexibility, robustness to noise, and applicability in exploring complex tissue structures and identifying marker ions associated with microregions.