CARDINAL: open-source software for spatially-aware feature-sparse segmentation and classification of mass spectrometry images

摘要

It is useful for interpretation to have a common framework for clustering and classification. Statistical regularization facilitates idenfication and interpretation important masses. Likelihood can help guide parameter choices for the number of clusters and the shrinkage penalty. The t-statistics can be used to rank the top masses for either segmentation or classification; the t-statistics for unimportant masses are shrunken toward 0. While the shrinkage from statistical regularization can be used to select important masses and improve accuracy by reducing the impact of unimportant masses and moderate noise, shrinkage cannot overcome strong systematic noise or excessively weak signal. Further investigation using controlled experiments is recommended to evalute the accuracy of segmented MS images and the limits of what molecules can reliably be detected.