Scan check

These are general checks that apply to any scan. They do not find organs, so they cannot tell whether contrast was timed well or whether the patient moved. Exam-specific checks will add those.

Voxel size. Voxels up to 1.5 mm pass, up to 5 mm get a warning, and thicker ones fail. Orientation. The angle between each image axis and the nearest body axis, from the header; above 1° the scan is tilted (oblique), which some tools ignore. Missing values. Any voxel that is not a number or infinite fails. Distinct values. 16 or fewer different values look like a segmentation and fail; up to 256 look like an 8-bit export and get a warning. Hounsfield units (CT). Air should read about −1000 HU; the check takes the median of values between −1100 and −950 HU (lung, at about −870 HU, is left out) and passes it between −1030 and −960 HU. A scan whose lowest value fills a large share of the image above −950 HU has been clipped, as has one whose highest value is below 2000 HU and holds at least 0.05% of the voxels; a scan with no values below −200 is not in Hounsfield units. Value range (MRI). Values at the limits of a 16-bit integer mean the data overflowed, and many negative values mean it is not a magnitude image. Empty slices. Slices that hold a single value. More than 10% fails. Duplicate slices. A slice stored exactly the same as the one before it, which noise makes impossible in a real acquisition.

Sudden changes between slices. For each slice the tool takes the mean tissue value (−200 to 300 HU in CT, above the background in MRI) and the centre of the anatomy. Neighbouring slices are flagged when the mean jumps by more than 25 HU in CT (10% of the median signal in MRI) and well beyond the scan's usual change, or when the anatomy shifts by more than 5 mm without a matching change in area. Anatomy at the image edge. On up to 64 slices the tool keeps the largest connected region above −500 HU in CT (above an Otsu threshold in MRI) and warns when it covers more than 5% of an image edge on more than a quarter of them.

Noise (CT) is a global noise index: the most common standard deviation in windows of about 5 mm over soft tissue (0 to 100 HU) on up to 30 slices (Christianson et al., Radiology 2015). It warns above 40 HU. Signal to noise (MRI) is the median tissue signal over the background noise. The noise is the most common value in the background at the outer edge of the image, which for magnitude images equals the Rayleigh noise level and ignores dark tissue there. It warns below 10 and is skipped when the background has been blanked. Very dense material (CT). Voxels above 3000 HU are metal or undiluted contrast in the vein it was injected into; the check warns from 3 mL.

Signal and background are optional points you choose. Around each, the tool takes the voxels within 1.5 mm, then grows a region from the point through connected voxels within 3 robust standard deviations of their median, up to the chosen radius, so it stops at the edge of the tissue. It reports the median and the robust standard deviation (1.4826 × the median absolute deviation) of each region. SNR is the signal median over the noise in the signal region, and CNR is the signal minus the background, over that same noise, as the coronary CT page does with the aorta and fat. In MRI, where coil combination and parallel imaging make noise vary across the image, noise measured inside the tissue is the more reliable choice. In CT, SNR depends on where 0 HU sits (water), so CNR is the more meaningful number.

The thresholds are starting points, not validated limits. Noise in particular depends on dose, reconstruction kernel and slice thickness, so compare scans of the same protocol.