Automated phase maps are an important tool for characterizing samples but the data quality must be evaluated. Common options include the overlay of phases on backscattered electron (BSE) images and phase composition averages and standard deviations. Both these methods have major limitations. We propose two methods of evaluation involving principal component analysis. First, a red–green–blue composite image of the first three principal components, which comprise the majority of the chemical variation, which provides a good reference against which phase maps can be compared. Advantages over a BSE image include discriminating between similar mean atomic number phases and sensitivity across the entire range of mean atomic numbers present in a sample. Second, principal component maps for identified phases, to examine for chemical variation within phases. This ensures the identification of unclassified phases and provides the analyst with information regarding the chemical heterogeneity of phases (e.g., chemical zoning within a mineral or mineral chemistry changing across an alteration zone). Spatial information permits a good understanding of heterogeneity within a phase and allows analytical artifacts to be easily identified. These methods of evaluation were tested on a complex geological sample. K-means clustering and K-nearest neighbor algorithms were used for phase classification, with the evaluation methods demonstrating their limitations.

X-ray fluorescence (XRF) microscopy is an important tool for studying trace metals in biology, enabling simultaneous detection of multiple elements of interest and allowing quantification of metals in organelles without the need for subcellular fractionation. Currently, analysis of XRF images is often done using manually defined regions of interest (ROIs). However, since advances in synchrotron instrumentation have enabled the collection of very large data sets encompassing hundreds of cells, manual approaches are becoming increasingly impractical. We describe here the use of soft clustering to identify cell ROIs based on elemental contents, using data collected over a sample of the malaria parasite Plasmodium falciparum as a test case. Soft clustering was able to successfully classify regions in infected erythrocytes as “parasite,” “food vacuole,” “host,” or “background.” In contrast, hard clustering using the k-means algorithm was found to have difficulty in distinguishing cells from background. While initial tests showed convergence on two or three distinct solutions in 60% of the cells studied, subsequent modifications to the clustering routine improved results to yield 100% consistency in image segmentation. Data extracted using soft cluster ROIs were found to be as accurate as data extracted using manually defined ROIs, and analysis time was considerably improved.