Nodule volume and slice thickness: 3D demo

The nodule is one slightly lobulated synthetic shape, the same at every diameter, and its true volume is known. A scan cuts space into voxels: 0.7 × 0.7 mm in the slice plane and the chosen thickness along the scan axis. A voxel on the edge of the nodule is only partly filled, and its value is the average of nodule and lung. This is the partial-volume effect. The segmentation here keeps a voxel when at least half of it is nodule, and the measured volume is the number of kept voxels times the voxel volume.

With thin slices the edge voxels are small, so the kept voxels follow the surface closely. With thick slices a single voxel can span most of the nodule’s height, and whether it is kept depends on where the slice boundaries fall. The slice position slider moves those boundaries by a share of one slice, as happens when a patient lies a few millimetres differently on the table. The 3D view shows the kept slices as blue slabs inside the true surface, drawn as a white dashed outline; drag it to see the staircase from the side.

Growth of a nodule is judged from the change in volume between two scans, and the British Thoracic Society guideline on pulmonary nodules (2015) defines significant growth as a change of 25% or more. If the measurement itself can move by more than that, an unchanged nodule can be reported as growing and a growing one as stable. That is why volumetry needs thin slices, the same protocol at baseline and follow-up, and a stated measurement variability for the nodule size in question. See medical imaging AI.

Picture a follow-up programme for small lung nodules. A patient has a baseline CT and a follow-up some months later, at a different site with a different protocol. Software segments the nodule on both scans and reports the volume, and the report compares the two numbers.

The demo’s default settings are the thick-slice scan: an 8 mm nodule with a true volume of 254 mm³, 5 mm slices and slice position 0%. One slice is kept, the measured volume is 189 mm³ and the measurement error is −26%. The 3D view shows why: a single slab whose flat top and bottom cut off the rounded ends of the nodule.

Now move the slice position slider and leave everything else alone. The nodule does not change, but at 45% two slices are kept and the measured volume is 299 mm³, an error of +18%. Across all positions the measured volume runs from −26% to +18% of the true volume, and the largest gap between two scans reads +58%. A change of that size between baseline and follow-up would be reported as clear growth.

Press Thin 1 mm. Nine slices are kept, the measured volume is 250 mm³ and the error is −2%. Across all slice positions it runs from −4% to −1%, and the largest gap between two scans is 3%. With Standard 2.5 mm the gap is 10%. Size matters as well: at 5 mm slices the gap is 27% for a 6 mm nodule and 7% for a 12 mm nodule, and a 4 mm nodule is not segmented at all at some slice positions.

The correct practice is to treat volume as a measurement with an error. Use thin slices for volumetry, keep slice thickness and reconstruction the same at baseline and follow-up, and record them with the result. Know the variability to expect for the nodule size and protocol in use, and compare any reported change with it before calling growth. A model or software validated on thin-slice data should be validated again before it is used on thick slices.

• Are baseline and follow-up acquired and reconstructed with the same slice thickness and kernel?
• Is the slice thickness thin enough for the smallest nodule you intend to measure?
• Do you know the repeat-scan variability of the volume for each nodule size and protocol in use?
• Is a reported change compared with that variability before it is called growth or shrinkage?
• Was the segmentation software or model validated on the slice thicknesses used in practice?
• Are slice thickness, pixel size and reconstruction recorded next to every volume?

Related: our medical imaging AI service, the wiki page on medical AI validation, and the demos on Dice score vs clinical error and scanner shift.

Why does the same lung nodule measure differently on two CT scans?

Because the measured volume depends on the scan as well as on the nodule. Voxels on the nodule’s edge are only partly filled, and with thicker slices or a small shift in slice position a different set of voxels is counted. In this demo an unchanged 8 mm nodule measures between −26% and +18% of its true volume on 5 mm slices, depending only on where the slices fall.

How does CT slice thickness affect nodule volume measurement?

Thicker slices make each edge voxel larger, so the measurement becomes coarser and more sensitive to slice position. In this demo the largest gap between two scans of the same 8 mm nodule is 3% at 1 mm slices, 10% at 2.5 mm and 58% at 5 mm. The effect is stronger the smaller the nodule is relative to the slice.

How much volume change counts as real growth in a lung nodule?

The British Thoracic Society guideline on pulmonary nodules (2015) defines significant growth as a volume change of 25% or more. That criterion assumes the measurement is repeatable to within less than 25%. The demo shows a case where it is not: on 5 mm slices two scans of an unchanged 8 mm nodule can differ by 58%. This is general information, not medical advice.

What is the partial-volume effect in CT?

A voxel that contains both nodule and surrounding lung shows the average of the two, so the edge of a structure is blurred over one voxel. A segmentation has to decide whether each edge voxel counts, and that decision changes the measured size. The larger the voxel relative to the structure, the larger the effect.

Can an AI segmentation model fix thick-slice volume errors?

Only partly. A model can estimate the boundary inside a voxel better than a fixed cut-off, but the information lost by averaging over a thick slice is not in the image. The reliable fix is at acquisition: thin slices and the same protocol at every time point, with the model validated on the protocols it will meet.