Longitudinal-CT

Whole-body contrast CT of one melanoma patient at baseline (fixed) and at follow-up under systemic therapy (moving), with every lesion segmented and linked between the two times (Longitudinal-CT, autoPET IV task 2).

29 test pairswhole bodysame patientfixed CT, moving CTaligned once by the benchmarkresearch only

Look at the images

Read before comparing (10)

Truth. The landmarks are the centres of gravity of the lesions segmented on each image (cog_bl and cog_fu in the patient's CSV), for lesions linked one to one. One radiologist segmented every visible lesion freehand and linked the two times; two radiologists then reviewed both in consensus. The CSV's propagated centres, carried by a registration, are left out.
Links. The annotation software synchronised the two scans by a deformable registration and warned when linked lesions sat apart under it, so the links were checked against a registration. The provider's propagating registration lands a median 4.3 mm from the annotated follow-up centres (1.4 to 9.5 mm between quartiles, up to 40 mm).
Change. Over the 29 test pairs the 276 lesions holding one label at both times changed volume by a median factor of 0.74. A lesion that changes unevenly moves its centre, so the landmark error has a floor above zero, up to the radius of a large lesion. The board measures how close linked lesions come, as lesion tracking asks.
Unmatched. Lesions that disappeared, appeared, merged or split have no one-to-one pair; a deformation cannot match them, and they are left out.
Labels. The organ labels come from TotalSegmentator, run on 2 to 8 overlapping parts of each image to fit in 10 GB; where it also ran on a whole image (19 images), the maps agree at a mean Dice of 0.90 to 0.99. A method registering on that network's features is scored against its own view. Organ overlap is shown, not ranked.
Lesion masks. Shown (dice_lesion), never ranked: their volume changes under therapy. A lesion holds the same label on both masks only when the readers linked it, and dice_lesion is the mean over those lesions.
Alignment. The two scans lie on different grids with unrelated origins and cover 399 to 1610 mm of the body, skull base to mid-thigh as a rule. Each pair gets one elastix translation on mutual information inside each body, written in the moving headers; no voxel is resampled and no lesion or label used.
Baseline. With no registration the 29 test pairs leave a lesion-centre error of 15.0 mm (5.4 to 36.3), 21.6 mm at the 95th percentile, and an organ Dice of 0.43 (0.16 to 0.82).
Volumes. A study can hold two or three acquisition volumes, such as a dedicated head and neck scan; the pair is the first of each, and the other volumes' lesions are left out. The CSV counts a centre from the corner of the first voxel, hence first_index 0.5.
Images. Version 3 of the FDAT record: defaced portal-venous contrast CT in Hounsfield units, 1.5 to 5 mm slices. The paper names Siemens scanners; the demographics list others, such as one test follow-up on a Philips Ingenuity Core.

No method has a result on every pair of this dataset yet.

		Accuracy					Smoothness		Cost
#	Method	TRE mm	Change	Dice *	HD95 mm *	rTRE	SDlogJ	Folds %	Time s	RAM GB	GPU GB	Images
	No registration	15.01 ±3.00		0.430 ±0.057	16.83 ±4.16	0.0136 ±0.0029	0.000	0.00

* shown, not ranked: on this dataset it is checked against the output of a program.

1 result not ranked: an older run, pairs still missing, or a failure on every pair

		Accuracy					Smoothness		Cost
#	Method	TRE mm	Change	Dice *	HD95 mm *	rTRE	SDlogJ	Folds %	Time s	RAM GB	GPU GB	Machine	Images
	ConvexAdam (MIND)GPUapp21/29 pairs	5.19 ±1.55	−9.81	0.810 ±0.033	7.48 ±3.58	0.0055 ±0.0018	0.101	0.76	40	5.9	16.3	Laptop

Dice: how much the organ contours overlap, 1 is perfect. HD95: distance between the contours' surfaces. TRE: distance between matching landmarks. rTRE: the TRE over the length of the image's diagonal, to compare datasets of other sizes and units. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.

Pair by pair

Measure

Dice	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29
ConvexAdam (MIND)	0.845	0.631	0.665	0.780		0.761	0.825	0.848	0.760	0.858	0.842	0.826		0.713	0.771	0.880			0.849			0.817		0.887	0.827	0.840		0.898	0.880
No registration	0.587	0.505	0.315	0.208	0.319	0.370	0.409	0.600	0.401	0.502	0.504	0.373	0.300	0.310	0.412	0.644	0.285	0.457	0.608	0.315	0.159	0.412	0.288	0.608	0.433	0.333	0.350	0.818	0.652

Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.

About the data

Test pairs: The provider's test split (data_split.json) whole: 29 of its 30 patients, 07e1cd7dca left out for want of a lesion linked one to one at both times. 10 development pairs from the provider's training split, the first ten of the reverse seeded order among its 224 patients with such a lesion, so a model trained on that split is tuned where it trained, not where it is scored.
Truth: landmarks by experts contours by a program lesion by experts TotalSegmentator (task total) run by us on each image, in overlapping parts along the body (prepare.py --labels) the centres of gravity of the lesions segmented by hand on each image, for the lesions the readers linked one to one between the two times: 1 to 40 a pair, 274 in the 29 test pairs
Source: doi.org/10.57754/FDAT.qe950-g4h94
doi.org/10.57754/FDAT.75kj1-64747
fdat.uni-tuebingen.de/records/qwsry-7t837
doi.org/10.1038/s41597-026-07466-y
github.com/lab-midas/autoPETCTIV
autopet-iv.grand-challenge.org
Licence: CC BY-NC 4.0

10 pairs for tuning, never tested

a666587afd, f899139df5, 158f3069a4, a3458cbbcc, 069059b7ef, 7e7757b1e1, 34ed066df3, 3def184ad8, d2490f048d, 914b62192c