This is the authors' abstract. We don't add key points for this paper.
Mass spectrometry imaging is a powerful tool for investigating the spatial distribution of chemical compounds in a biological sample such as tissue. Two common goals of these experiments are unsupervised segmentation of images into newly discovered homogeneous segments and supervised classification of images into predefined classes. In both cases, the important secondary goals are to characterize the uncertainty associated with the segmentation and with the classification and to characterize the spectral features that define each segment or class. Recent analysis methods have focused on the spatial structure of the data to improve results. However, they either do not address these secondary goals or do this with separate post hoc procedures.We introduce spatial shrunken centroids, a statistical model-based framework for both supervised classification and unsupervised segmentation. It takes as input sets of previously detected, aligned, quantified, and normalized spectral features and expresses both spatial and multivariate nature of the data using probabilistic modeling. It selects informative subsets of spectral features that define each unsupervised segment or supervised class and quantifies and visualizes the uncertainty in spatial segmentations and in tissue classification. In the unsupervised setting, it also guides the choice of an appropriate number of segments. We demonstrate the usefulness of this framework in a supervised human renal cell carcinoma experimental dataset and several unsupervised experimental datasets, including a pig fetus cross-section, three rodent brains, and a controlled image with known ground truth. This framework is available for use within the open-source R package Cardinal as part of a full pipeline for the processing, visualization, and statistical analysis of mass spectrometry imaging experiments. Mass spectrometry imaging is a powerful tool for investigating the spatial distribution of chemical compounds in a biological sample such as tissue. Two common goals of these experiments are unsupervised segmentation of images into newly discovered homogeneous segments and supervised classification of images into predefined classes. In both cases, the important secondary goals are to characterize the uncertainty associated with the segmentation and with the classification and to characterize the spectral features that define each segment or class. Recent analysis methods have focused on the spatial structure of the data to improve results. However, they either do not address these secondary goals or do this with separate post hoc procedures. We introduce spatial shrunken centroids, a statistical model-based framework for both supervised classification and unsupervised segmentation. It takes as input sets of previously detected, aligned, quantified, and normalized spectral features and expresses both spatial and multivariate nature of the data using probabilistic modeling. It selects informative subsets of spectral features that define each unsupervised segment or supervised class and quantifies and visualizes the uncertainty in spatial segmentations and in tissue classification. In the unsupervised setting, it also guides the choice of an appropriate number of segments. We demonstrate the usefulness of this framework in a supervised human renal cell carcinoma experimental dataset and several unsupervised experimental datasets, including a pig fetus cross-section, three rodent brains, and a controlled image with known ground truth. This framework is available for use within the open-source R package Cardinal as part of a full pipeline for the processing, visualization, and statistical analysis of mass spectrometry imaging experiments. Mass spectrometry imaging visualizes the spatial distribution of molecular ions in a sample by repeatedly collecting mass spectra across its surface. Equipped with a computer-controlled sample stage, a mass spectrometer rasters across the sample and collects individual mass spectra from discrete or continuous locations. The leading technologies for performing MS imaging include matrix-assisted laser desorption/ionization (MALDI) and desorption electrospray ionization (DESI) 1The abbreviations used are:DESIdesorption electrospray ionizationMALDImatrix-assisted laser desorption/ionizationMSmass spectrometrym/zmass-to-charge ratioO-PLS-DAorthogonal projection to latent structures discriminant analysisPLS-DAprojection to latent structures discriminant analysisRCCrenal cell carcinomaSAspatially awareSASAspatially aware structurally adaptiveTOFtime-of-flight.. MALDI imaging requires the application of a matrix solution and is typically used to detect large molecules such as peptides and proteins. DESI imaging does not require a matrix and is typically used to detect small molecules such as lipids, metabolites, and drug molecules (19Wiseman J.M. Laughlin B.C. Desorption electrospray ionization (DESI) mass spectrometry: A brief introduction and overview.Current Separations Drug Development. 2007; 22: 11-14Google Scholar). The intensities at a particular m/z value and spatial location can then be plotted as pixels in a false-color image, called an ion image, that displays the spatial distribution of the analyte associated with that m/z value. MS imaging has been shown to be promising in a wide range of biological applications such as molecular histology of tissue, whole body sections, bacterial films, etc. (17Watrous J.D. Alexandrov T. Dorrestein P.C. The evolving field of imaging mass spectrometry and its impact on future biological research.J. Mass Spectrom. 2011; 46: 209-222Crossref PubMed Scopus (102) Google Scholar) and biomedical applications such as cancer diagnosis (7Dill A.L. Eberlin L.S. Costa A.B. Zheng C. Ifa D.R. Cheng L. Masterson T.A. Koch M.O. Vitek O. Cooks R.G. Multivariate statistical identification of human bladder carcinomas using ambient ionization imaging mass spectrometry.Chemistry. 2011; 17: 2897-2902Crossref PubMed Scopus (98) Google Scholar, 8Dill A.L. Eberlin L.S. Zheng C. Costa A.B. Ifa D.R. Cheng L. Masterson T.A. Koch M.O. Vitek O. Cooks R.G. Multivariate statistical differentiation of renal cell carcinomas based on lipidomic analysis by ambient ionization imaging mass spectrometry.Anal. Bioanal. Chem. 2010; 398: 2969-2978Crossref PubMed Scopus (123) Google Scholar). desorption electrospray ionization matrix-assisted laser desorption/ionization mass spectrometry mass-to-charge ratio orthogonal projection to latent structures discriminant analysis projection to latent structures discriminant analysis renal cell carcinoma spatially aware spatially aware structurally adaptive time-of-flight. Computational analysis of MS imaging experiments typically consists of preprocessing, followed by a statistical analysis (1Alexandrov T. MALDI imaging mass spectrometry: Statistical data analysis and current computational challenges.BMC Bioinformatics. 2012; 13: S11Crossref PubMed Google Scholar). The preprocessing ensures that mass spectral intensities are comparable across all mass spectra in the experiment. This is typically done via normalization and (if necessary) baseline reduction. Furthermore, the preprocessing extracts spectral features that correspond to the underlying analytes. This is typically done via peak picking, or m/z binning or resampling. Much progress has already been made in the preprocessing of MS imaging data. Many mature tools for preprocessing mass spectra already exist (21Yang C. He Z. Yu W. Comparison of public peak detection algorithms for MALDI mass spectrometry data analysis.BMC Bioinformatics. 2009; 10: 4Crossref PubMed Scopus (203) Google Scholar). However, a major bottleneck is the downstream statistical analysis of the processed data. This manuscript focuses on the downstream statistical analysis steps, which take place after the detection, quantification, alignment, normalization, and (optionally) identification of the initial set of spectral Two common goals of statistical analysis of MS imaging experiments post are classification supervised pixels to predefined and segmentation unsupervised pixels to newly discovered segments. A number of methods for these already multivariate methods are used for both classification and segmentation. methods including discriminant projection to latent structures discriminant analysis and orthogonal projection to latent structures discriminant analysis have A.L. Eberlin L.S. Zheng C. Costa A.B. Ifa D.R. Cheng L. Masterson T.A. Koch M.O. Vitek O. Cooks R.G. Multivariate statistical differentiation of renal cell carcinomas based on lipidomic analysis by ambient ionization imaging mass spectrometry.Anal. Bioanal. Chem. 2010; 398: 2969-2978Crossref PubMed Scopus (123) Google Scholar, A.L. Eberlin L.S. Costa A.B. Zheng C. Ifa D.R. Cheng L. Masterson T.A. Koch M.O. Vitek O. Cooks R.G. Multivariate statistical identification of human bladder carcinomas using ambient ionization imaging mass spectrometry.Chemistry. 2011; 17: 2897-2902Crossref PubMed Scopus (98) Google Scholar, L.S. J.M. detection of peptides and small in and of by MALDI mass spectrometry 2012; PubMed Scopus Google Scholar, The of Mass to of of PubMed Scopus Google Scholar, T. Z. for imaging mass of in PubMed Scopus Google Scholar, C. A.L. Eberlin L.S. Cooks R.G. Ifa D.R. Mass spectrometry imaging ambient Spectrom. PubMed Scopus Google Scholar). methods such as or by analysis to the of the are used C. MALDI imaging with as a tool for the of human PubMed Scopus Google Scholar, and spectral in MALDI mass spectrometry images by and multivariate Chem. PubMed Scopus Google Scholar, C. and Comparison of for the analysis of mass spectrometry imaging Scholar). The multivariate methods have they are of the spatial structure of the data. each and of spectra from spatially the of the results. they do not the input features to informative the methods to for the spatial structure of MS of methods on the spatial structure to detect from spectra T. for of known and molecules in imaging mass PubMed Scopus Google Scholar, J.M. data analysis pipeline based on a peak for imaging mass spectrometry PubMed Scopus Google Scholar). these methods at the and do not address the goals of image segmentation or image classification. of including spatially aware and spatially aware structurally adaptive by Alexandrov and T. spatial segmentation of large imaging mass spectrometry with spatially aware 2011; PubMed Scopus Google for the spatial structure of the data and improve the of image segmentation (1Alexandrov T. MALDI imaging mass spectrometry: Statistical data analysis and current computational challenges.BMC Bioinformatics. 2012; 13: S11Crossref PubMed Google Scholar). However, to the multivariate analysis they do not subsets of spectral features that define the segments and on post hoc to the features associated with the using the a segment and the ion images (1Alexandrov T. MALDI imaging mass spectrometry: Statistical data analysis and current computational challenges.BMC Bioinformatics. 2012; 13: S11Crossref PubMed Google Scholar, C. and Comparison of for the analysis of mass spectrometry imaging Scholar). the statistical has a of choice for subsets of informative features from multivariate data. such is shrunken by T. of cancer by shrunken of PubMed Scopus Google Scholar, T. by shrunken with applications to Scopus Google which for classification of A has been to in MS imaging experiments using L.S. T.A. of of cancer by PubMed Scopus Google Scholar). However, to the multivariate analysis they do not for the spatial structure of the data. This manuscript a statistical and for unsupervised segmentation and supervised classification of MS It takes as input a set of previously detected, quantified, aligned, and normalized by of It the of both spatially aware by Alexandrov and and statistical by We for unsupervised the spatial probabilistic segmentation. It the of segment for each and to and the uncertainty of segmentation for each statistical by subsets of the spectral such that each each Statistical also of the number of segments. for supervised image classification probabilistic the of predefined tissue class for each and by subsets of spectral features that define each class. We the of supervised classification on a human renal cell carcinoma its for pixels as cancer or tissue. We the of unsupervised segmentation on several biological including a pig fetus and three rodent datasets, its for unsupervised also on a sample of known in to its controlled This framework has been previously in the open-source R package Cardinal Eberlin L.S. C. Vitek O. R package for statistical analysis of mass imaging PubMed Scopus Google as part of a full pipeline for the processing, visualization, and statistical analysis of mass spectrometry imaging experiments. This manuscript the framework this The of this to features of the pig such as major unsupervised analysis of the mass A secondary to spectral features associated with the is an image of the and tissue the of the pig including major such as the and Mass spectra using a ion mass spectrometer with a DESI ion the m/z The images to spectra from the The dataset of mass spectra with spectral The mass spectra normalized to a common ion and peak to the dataset to data and analysis using Cardinal Eberlin L.S. C. Vitek O. R package for statistical analysis of mass imaging PubMed Scopus Google Scholar). a ion image the and and a ion image the use this dataset to demonstrate unsupervised statistical analysis using all the mass spectral to the major The of this to use a controlled sample to the of data and statistical A of a on to a and MS imaging image of the data is shown in The mass spectra using a ion mass spectrometer with a DESI ion the m/z The dataset of mass spectra with spectral Mass spectra normalized to a common ion and peak to the dataset to data and analysis using a ion image the part of the and a ion image the used in the The shown in can be the ground We use this dataset to the of the unsupervised statistical analysis of the mass spectral to the ground truth. The for these is to the of several experiments of data three experiments a rodent The dataset is a image of a T. L. segmentation of imaging mass spectrometry data with image and 2010; PubMed Scopus Google Scholar, T. spatial segmentation of large imaging mass spectrometry with spatially aware 2011; PubMed Scopus Google which is shown in Mass spectra on a mass spectrometer the to m/z The images to and the to m/z range The dataset of mass spectra with spectral The mass spectra normalized to a common ion and baseline using Cardinal used to peak to the dataset to for baseline and normalization, all data and analysis done in The dataset is a shown in This spectra with a of experimental Mass spectra on a ion mass spectrometer with a DESI ion the m/z The images to The dataset of mass spectra with spectral The mass spectra normalized to a common ion and peak to the dataset to data and analysis done in The dataset is a shown in This dataset features a of experimental Mass spectra using an MALDI the to m/z The images to The dataset of mass spectra with spectral The mass spectra normalized to a common ion and baseline then to the dataset to data and analysis done in We use these to characterize the of the of statistical analysis to in data The of this to renal tissue as cancer or In with at of of tissue from human with renal cell carcinoma with each of tissue and tissue A.L. Eberlin L.S. Zheng C. Costa A.B. Ifa D.R. Cheng L. Masterson T.A. Koch M.O. Vitek O. Cooks R.G. Multivariate statistical differentiation of renal cell carcinomas based on lipidomic analysis by ambient ionization imaging mass spectrometry.Anal. Bioanal. Chem. 2010; 398: 2969-2978Crossref PubMed Scopus (123) Google Scholar). images of the tissue tissue as or by a However, the are based on the tissue for each whole tissue, from the class. The mass spectra using a ion mass spectrometer with a DESI ion the m/z The images to spectra from the The dataset of mass spectra with spectral tissue of to mass spectra The mass spectra normalized to a common ion current and to in spectral data and analysis using ion images for which is an ion known to be in tissue, and ion images for which is known to be in tissue A.L. Eberlin L.S. Zheng C. Costa A.B. Ifa D.R. Cheng L. Masterson T.A. Koch M.O. Vitek O. Cooks R.G. Multivariate statistical differentiation of renal cell carcinomas based on lipidomic analysis by ambient ionization imaging mass spectrometry.Anal. Bioanal. Chem. 2010; 398: 2969-2978Crossref PubMed Scopus (123) Google Scholar). to such as the of m/z the of the tissue in sample We use this dataset to demonstrate the of the framework to spectral features important in the the of the biological a with the collects spectra at locations. all the the spectra and the location of a on a sample We do not that the are in the are However, do the location of a on the We that the spectra at these have been that has the as a peak or a m/z We also that the intensities are normalized that spectra are comparable across pixels and across the at a location on sample as In is a of intensities for spectral also that the spectra and the pixels to of supervised or segments unsupervised supervised the class is for from by a and the statistical of the is to each to of these in a supervised based on its for unsupervised the class is and the statistical of the is to these from the spectra in an unsupervised the number of and the number of to class by an unsupervised or a supervised the for a known class or discovered segment as and the as is a of intensities which are the intensities for spectral spectra from all to class and is a of intensities which are the intensities for spectral all In the the spatial shrunken framework for both supervised classification and unsupervised segmentation. supervised the on the known classes. unsupervised the segments are the segments are or by segmentation such as T. spatial segmentation of large imaging mass spectrometry with spatially aware 2011; PubMed Scopus Google Scholar) and are of several is We the important The full algorithms are available in and In mass imaging tissue or class is typically by a also called its that each class is using shrunken centroids, from the of shrunken by T. of cancer by shrunken of PubMed Scopus Google Scholar, T. by shrunken with applications to Scopus Google Scholar). This to the class to the and to the informative spectral features as to the We this We and the class centroids, and use statistical to the the We then the for spectral for class is the of the for The number the to the of the the to the and is the of to a number of to be set to define the intensities of the shrunken for each for each class that is the shrunken for class The shrunken can be as spectra of the the intensities have been the the for a class from the for spectral features that are of the class features that are not from the have intensities set to the for that The shrunken in are for informative The spectral features with are for class spectral features with are from class as with the features with are as the features with a to class the individual or to a to a to define a the spectra from individual pixels and the shrunken We to use the spatially aware by Alexandrov and T. spatial segmentation of large imaging mass spectrometry with spatially aware 2011; PubMed Scopus Google Scholar). We this and it in the the and to the of a or Alexandrov and a spatially aware spectra and which on the spectra from pixels within a of and The that this is as it as with methods that do not for the spatial pixels T. L. segmentation of imaging mass spectrometry data with image and 2010; PubMed Scopus Google Scholar). for a of the spectra is the are spatial of the The of these in either a spatially aware or a spatially aware structurally adaptive In the the are which are of the spectra and on the which with the from the is a choice it that pixels from each are pixels that are In the the are which are adaptive that the spectra are from the This is to and small in which be to by the The is set to the of the of the the spectra in the The can be to the the individual pixels and the shrunken as the using the as in in of the and using adaptive with as in in of the We the in both that they to in in the a and a class the the spectra at that of the has this In the of supervised use this to pixels to the shrunken of known classes. In the of unsupervised image use this to the pixels to discovered segments. also in both supervised and unsupervised this requires the of the for unsupervised the of the number of segments is also The for these and and be in of supervised classification shrunken can be as a of discriminant analysis T. of cancer by shrunken of PubMed Scopus Google Scholar, T. by shrunken with applications to Scopus Google Scholar). In this each of the has a and is as a multivariate are to a common This to a to for individual to a class using a discriminant based on the or from each to each of the shrunken as is the for We typically the as the data are not of the be are using spatial that spectra from the discriminant be as However, demonstrate that the in by class for each for each class A is to the class with the segmentation the is the number of and the segments are or with segmentation We typically use a be that and a each of the a is as to the segment with the using The framework requires the choice of the for unsupervised the number of segments In the of supervised the are and can be by a set of biological on each is as an experimental for experiments with a small number of biological can be each of spatial shrunken for a range of of The of is the that the classification on the This is for the human in In the of unsupervised the individual segments and also the number of segments are However, is a the number of informative features in the by the and the number of segments segments to be by are statistical the segments have and in can informative and this in the of the segments. We the and the number of segments by segmentations for of and and then the and the number of segments. We this using experimental data in and in The full for spatial shrunken for a set of is in for supervised and in for unsupervised segmentation. The classification and segmentation methods are in the R package Cardinal Eberlin L.S. C. Vitek O. R package for statistical analysis of mass imaging PubMed Scopus Google which is available from Cardinal is and and on and The is and for It can large and is by the that the dataset be into the segmentations for the pig dataset and in and and The segmentations for the and in and and The for the human dataset features and in segmentations for the pig fetus dataset are in which from segmentation methods with the segmentation In to the in a segmentation. The is not to a to the of the which also in a the as a and the spatially aware and spatially aware structurally adaptive of Alexandrov and T. spatial segmentation of large imaging mass spectrometry with spatially aware 2011; PubMed Scopus Google which both in segmentations with segments. The is to a segment in both as as the and of the methods require a number of which set to based on the in the Statistical of the of for and the spatial shrunken segmentation with and which segmentations comparable to in and The number of segments for these methods to and in segments in the as in the Statistical of the of for In the segmentations in and are to each in and that the spatial shrunken across of spatial segmentations of the are shown in which the of and methods with the ground truth. The of the number of segments for the pig fetus segmentation from is in an appropriate spatial shrunken with of segments and the from to in of segmentations are that in a comparable number of segments from of of segments. This that all segments have been and the segments the pig fetus dataset in this after We also for the number of segments to as In this The segmentation in to of the number of segments for the segmentation from is in the pig fetus segmentation from the spectral features using the spatial shrunken segmentation are shown in is shown for the and with and ion that each unsupervised or supervised segment is by its of informative features be informative for and features be informative for the spectral features and features the spectral features and the spectral features and features with the and the spectral features associated with for the segmentation from is shown in The rodent of used to the of the to uncertainty in its spatial shrunken segmentation in of segment using to this a of uncertainty in a segmentation. for the three rodent are in shrunken segmentation for each rodent dataset with and the segmentations plotted in using the in Statistical of the of for for an appropriate number of segments. This in segments for both the with three segments for the with and three segments for the with the appropriate as shown in for in and the number of segments to three as to This the of the in this the of the the of the number of segments to segments. and spectral features from the segmentation the segments. segmentations are plotted in the with this at the with this at the with this at the of in the data. of the human dataset using the is in for of the used to the as in shrunken as pixels as cancer or with to the of the tissue. to the dataset and A of spatial shrunken for classification is its of informative features that each as shown in and which use all spatial shrunken the features that each class. the the ion associated with tissue which is known to be in cancer A.L. Eberlin L.S. Zheng C. Costa A.B. Ifa D.R. Cheng L. Masterson T.A. Koch M.O. Vitek O. Cooks R.G. Multivariate statistical differentiation of renal cell carcinomas based on lipidomic analysis by ambient ionization imaging mass spectrometry.Anal. Bioanal. Chem. 2010; 398: 2969-2978Crossref PubMed Scopus (123) Google Scholar). The ion associated with tissue which is known to be in tissue A.L. Eberlin L.S. Zheng C. Costa A.B. Ifa D.R. Cheng L. Masterson T.A. Koch M.O. Vitek O. Cooks R.G. Multivariate statistical differentiation of renal cell carcinomas based on lipidomic analysis by ambient ionization imaging mass spectrometry.Anal. Bioanal. Chem. 2010; 398: 2969-2978Crossref PubMed Scopus (123) Google Scholar). of spatial shrunken is the of of class these with of the in the This can of the individual and in in the tissue an of tissue the and in the tissue an of tissue the are by ions known to be associated with cancer and tissue A.L. Eberlin L.S. Zheng C. Costa A.B. Ifa D.R. Cheng L. Masterson T.A. Koch M.O. Vitek O. Cooks R.G. Multivariate statistical differentiation of renal cell carcinomas based on lipidomic analysis by ambient ionization imaging mass spectrometry.Anal. Bioanal. Chem. 2010; 398: 2969-2978Crossref PubMed Scopus (123) Google Scholar) and the be and and be The manuscript a statistical called spatial shrunken centroids, for both unsupervised segmentation and supervised classification of MS imaging experiments. unsupervised it segmentations of the mass spectra or of It of segment and characterize and the uncertainty in the segmentation. It selects the number of as as subsets of informative features that define each segment to results. supervised spatial shrunken as with used methods such as and However, to the unsupervised it and visualizes the uncertainty of segment and subsets of informative features that define each class. shrunken is to with data after It takes as input a set of previously detected, quantified, aligned, and normalized features and is not for such features from the data spatial shrunken does not require a identification on the underlying analytes. The at the in the and this can be done with or the of the analyte However, spatial shrunken can the of the informative subsets of features in each segment or class can the of that to This framework has been previously in the open-source R package Cardinal Eberlin L.S. C. Vitek O. R package for statistical analysis of mass imaging PubMed Scopus Google Scholar). The in Cardinal can large datasets, as as they can be into and can be used as part of MS imaging data We that the and of the it a tool for biological and We Cooks for the human pig and We for the We Alexandrov for with
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Bemis et al. (2016) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: