PRISM
The contrast phase of an abdominal CT, read from two numbers: the median intensity of the aorta and of the portal vein.
- Year
- 2024
- Kind
- Research project
- Site
- MD Anderson Cancer Center
- Stack
- TotalSegmentator, PyRadiomics, MATLAB
Background
A CT’s contrast phase is often lost when the scan is de-identified or moved between institutions, and relabelling by hand does not scale. The phase matters. Hepatocellular carcinoma is recognised by how it enhances in the arterial phase, and a study that mixes phases without knowing it compares unlike images.
Method
TotalSegmentator outlines the aorta and the portal vein, and PyRadiomics takes the median intensity inside each, in Hounsfield units. A median barely moves for a calcified wall or a streak of noise. Each scan then becomes one pair of numbers, which passes through three steps; each step settles what it can and hands on the rest.
| Step | What it compares | Settled |
|---|---|---|
| Clear cases | The aorta alone, against each phase’s training median and interquartile range; only a value far from a boundary is decided | 18.0% |
| Cutoffs | Both vessels, against cutoffs fitted on the training split; decided when the two cutoffs point to one phase | 40.1% |
| Neighbours | Both vessels, by the nearest neighbours among the training scans | 41.9% |
Every cutoff is refitted on each training split, from that split’s own medians and interquartile ranges, so the held-out 30 percent never shapes the rule that classifies it.
- From the public repository’s development set and its 1,000 splits; the two intensities only.
- Each cell is 20 by 10 HU; its hatching grades the scans in it, 1 to 2, 3 to 9, 10 and more.
- Bars: each phase’s median and interquartile range, fitted once on all 1,018 for the drawing; each split refits its own.
- Cutoffs at the median of their refits; the hatched strip on each axis spans the 5th to 95th percentile of them.
- The fence is the portal venous median plus 3 interquartile ranges. Its low twin, the arterial median less 3, lies at −85.5 HU, below every scan, and decides none.
- Inside the phantom line the neighbours decide, by the fitted model, whose boundary is not drawn.
- 2 scans lie beyond 600 HU of aorta, off the plane.
Results
On 1,018 development scans (235 arterial, 394 portal venous, 389 pre-contrast), accuracy averaged 96.67 percent over 1,000 random 70/30 splits and ranged from 92.46 to 99.02 percent; the 95 percent interval of the mean is 96.60 to 96.74. By phase, recall was 99.16 percent for pre-contrast, 97.77 for portal venous and 90.76 for arterial.
PRISM was then checked on 827 scans from other institutions: 228 arterial, 529 portal venous and 70 pre-contrast. The public material holds those scans’ features but not the predictions made on them, so this page gives no accuracy for them until it can be reproduced.
Limitations
Every figure above is internal: the splits are all drawn from one development set, and splits of one set flatter a method more than scans from elsewhere do. Until the predictions on the outside scans are public, the internal figures are the only ones this page can stand behind.
Arterial recall is the lowest of the three phases, because a scan taken as contrast leaves the arteries for the portal system falls between the two phases, where both the cutoffs and the neighbours are least sure; an arterial label is the one worth a second look.
Only three phases are modelled, so a delayed or nephrographic scan is forced into one of them and PRISM suits only protocols that use those three. Everything rests on the segmentation: a missed or mislabelled vessel gives a wrong median and a wrong phase, and one of the outside arterial scans has no aorta segmentation at all.
Research project, 2024.