Hospital CT Scanners Mismeasure the Ice Hidden in Permafrost, Study Finds

Medical CT scanners, the same machines that hospitals rely on to peer inside the human body, have become a tempting shortcut for climate scientists trying to measure the ice locked inside permafrost. A new study, however, delivers a sobering reality check. Researchers led by Mahya Roustaei of Ghent University, working with colleagues at the University of Montreal and WSP Canada, compared CT-derived estimates of ice content against direct laboratory measurements for 261 permafrost cores collected in Nunavut and Yukon, and found that the technology systematically misestimates how much ice the frozen ground actually contains. The work, published in The Cryosphere, suggests that while medical CT is superb for visualizing the architecture of frozen soil, it should not be trusted for routine quantification of ice.

The stakes are far higher than a technical quibble. Ground ice is the linchpin of permafrost stability. When it thaws, the ground can deform, subside, and collapse, damaging roads, pipelines, and buildings across the Arctic. Accurately knowing how much ice a given volume of soil holds is therefore fundamental to predicting thaw-related ground deformation and to engineering in cold regions. For decades, that knowledge came from visual descriptions, in which field geologists classified ice by hardness, structure, and colour. Those descriptive systems, dating back to the 1960s, proved unreliable and inconsistent, and a more modern cryostratigraphic approach, which classifies ice by its amount and spatial distribution, still depends on subjective visual interpretation. Laboratory measurement of volumetric ice content, based on weighing and drying samples, is accurate but destructive, slow, expensive, and limited in spatial coverage.

Computed tomography promised to break that bottleneck. By scanning frozen cores non-destructively, CT offers detailed three-dimensional images and the tantalizing possibility of computing the fractions of ice, sediment, and gas directly from density values. Medical CT scanners operate at millimetre to sub-millimetre resolution, while micro-CT systems resolve far finer pore-scale structures. But translating greyscale images into reliable estimates of soil components is anything but straightforward, and the new study quantifies exactly how badly the translation can go wrong.

The researchers collected permafrost cores on Bylot Island in Nunavut and near Beaver Creek in Yukon using a portable drill with a 10-centimetre barrel, keeping the samples below zero degrees Celsius from field to laboratory. The cores were subsampled into segments of roughly 5 to 20 centimetres and scanned at the Institut national de la recherche scientifique in Quebec City using a Siemens SOMATOM Sensation 64 medical scanner, with voxel resolutions of 0.18 to 0.24 millimetres in the horizontal plane and 0.4 to 0.6 millimetres vertically. Samples stayed frozen throughout, wrapped in plastic and foam to limit heat exchange, and no phase change was observed during the brief handling periods.

The physics of CT imaging is where the trouble begins. A CT image is built from voxels, three-dimensional pixels stacked from two-dimensional slices, and each voxel records a single averaged density value. A feature must be at least twice the voxel size to be detected at all. When a single voxel contains more than one material, their densities are averaged in what is known as the partial volume effect. In permafrost, pore ice occupying the tiny spaces between fine sediment grains is typically far smaller than a medical CT voxel, so its signal vanishes into the surrounding sediment, and the averaged density falls outside the range assigned to ice. The result is a systematic underestimation of ice content in sediment-rich samples.

Organic matter adds a second, opposite bias. The density of organic material spans a wide range depending on its degree of decomposition, from fibrous peat to fully humified soils, and portions of that range overlap with the density window assigned to ice in standard segmentation schemes. When researchers classify voxels by Hounsfield Unit thresholds, the standardised density scale of CT imaging, organic matter can be misclassified as ice, inflating the apparent ice content. The study found that ice content was underestimated in all low-organic-matter samples and overestimated in nearly all high-organic-matter samples, with the largest positive bias in samples containing more than 20 percent organic matter.

Across all 261 samples, the CT estimates were positively correlated with laboratory values but diverged systematically from the ideal one-to-one line. The mean bias was minus 12.9 percent, with a root mean squared deviation of 25.6 percent, meaning typical errors were enormous relative to the quantities being measured. Residuals depended strongly on composition: for samples with laboratory-measured ice content below about 75 percent, CT generally underestimated the true value, with negative residuals exceeding 60 percent in some cases. Above roughly 75 percent ice content, a tendency toward overestimation emerged. Agreement improved only when ice contents exceeded about 75 percent and organic matter sat in the 10 to 20 percent window, a narrow sweet spot that few natural samples occupy.

The team also tested how sensitive the results were to the choice of Hounsfield Unit thresholds, applying published threshold ranges from previous studies to two representative cores. The outcome was striking: estimated ice content swung from severe underestimation of up to minus 64 percent to marked overestimation of up to plus 31 percent relative to laboratory values, depending solely on which thresholds were used. Because permafrost is so compositionally heterogeneous, unlike relatively uniform sea ice, no single or transferable threshold can be defined for its materials. HU-based segmentation is inherently non-unique, and even calibrated corrections cannot overcome the fundamental limits imposed by scanner resolution and overlapping material densities.

The authors conclude that medical CT cannot reliably quantify volumetric ice content and should not be recommended for that purpose, given the errors, the cost, and the processing effort involved. Its genuine value lies elsewhere: in qualitative visualization of cryostructures, the distinctive patterns of lenticular, layered, reticulate, and suspended ice, sedimentary stratifications, and large-scale heterogeneity within cores, all of which are critical for interpreting how permafrost formed and how it will behave mechanically. Higher-resolution micro-CT systems may reduce partial-volume effects by resolving finer structures, and approaches based on sample-specific calibration or density normalization may partially improve estimates, but whether higher-resolution imaging can deliver accurate quantification remains an open question the authors identify as a key direction for future research.

For a field racing to understand how rapidly thawing Arctic ground will release water, reshape landscapes, and destabilize infrastructure, the message is clear. The scanner that so beautifully reveals the hidden architecture of frozen ground cannot be trusted to count the ice within it, and researchers who need numbers, not pictures, will still have to reach for the oven and the scale. The full dataset of laboratory and CT-derived measurements has been released on Zenodo, giving the community a benchmark for testing the next generation of imaging approaches.

Subject of Research: Accuracy of medical X-ray computed tomography for estimating volumetric ice content in permafrost samples

Article Title: Brief communication: Limitations of medical X-ray computed tomography for estimating ice content in permafrost samples

Article References: Roustaei, M., Darey, J., Mohammadi, Z., & Fortier, D. (2026). Brief communication: Limitations of medical X-ray computed tomography for estimating ice content in permafrost samples. The Cryosphere, 20(9), 5265-5270. https://doi.org/10.5194/tc-20-5265-2026

Image Credits: AI Generated

DOI: 10.5194/tc-20-5265-2026

Keywords: permafrost, ground ice, X-ray computed tomography, cryostratigraphy, Hounsfield Unit segmentation, partial volume effect, volumetric ice content, organic matter, Canadian Arctic, thaw settlement, micro-CT, The Cryosphere