LUNA25

Low-dose chest CT of one participant at two rounds of the National Lung Screening Trial, with the nodules annotated on both scans (LUNA25 public training and development set).

10 test pairslungsame patientfixed CT, moving CTaligned once by the benchmarkresearch only

Look at the images

Read before comparing (8)

Truth. The ranked truth is the nodule positions, 3 to 6 points a pair, each on a nodule of 4 mm or more. All are benign, stable over a median follow-up of at least six years by the protocol, and placed by two medical students trained by an experienced radiologist, from each nodule's lobe and CT section. No figure gives how far two placements differ.
Labels. TotalSegmentator's five lobes and airways, made by us on both scans, score the whole lung but are shown, not ranked: a method that registers on that network's features is scored against its own view of the anatomy. The nodules sample the field at a few places and say nothing of the rest of the chest.
Pool. 1,132 of the 2,120 participants have two scans with a nodule annotated on both: 827 share one nodule, 218 two and 87 three or more. A case is the participant's two scans that share the most nodules, the earlier two on a tie, the earlier fixed. Every pair drawn is one year apart.
Alignment. The scans come each in its scanner's frame, the nodules 12 to 2306 mm apart. The moving headers carry one elastix translation on mutual information inside each body; neither the nodules nor the labels are used, and no voxel is resampled.
Baseline. With no registration the 10 test pairs leave a nodule error of 7.1 mm (3.9 to 14.0), a lobe and airway Dice of 0.83 (0.74 to 0.88) and an HD95 of 6.7 mm.
Masks. Each scan's five TotalSegmentator lobes. 5 of the 40 test nodules on the fixed scans and 2 on the moving ones lie outside their scan's mask.
Images. The challenge's .mha files in Hounsfield units, as delivered: low-dose CT from four makers, 0.56 to 0.74 mm in plane, 1.3 to 2.5 mm slices. The -2048 outside the reconstructed circle is kept; set to air, it left ANTs' result on 100069_2000_2001 at 10 mm.
Overlap. Lung250M's ten NLST participants, also L2R_NLST's validation pairs, are left out; three are in LUNA25 (212587, 216160, 216422), none sharing three nodules. Lung250M's 22 NLST training pairs and Learn2Reg NLST's training pairs publish no participant ID; the lung masks of the first 26 scans built match none of Lung250M's development scans.

		Accuracy					Smoothness		Cost
#	Method	TRE mm	Change	Dice *	HD95 mm *	rTRE	SDlogJ	Folds %	Time s	RAM GB	GPU GB	Images
11–2	corrFieldGPUapp	0.71 ±0.23	−6.37	0.950 ±0.005	1.94 ±0.18	0.0012 ±0.0004	0.019	0.00	32	2.7	8.8
21–2	ConvexAdam (MIND)GPUapp	0.63 ±0.17	−6.45	0.956 ±0.005	1.80 ±0.23	0.0011 ±0.0003	0.090	0.60	34	4.6	11.8
3	DemonsCPUapp	2.83 ±1.87	−4.25	0.949 ±0.011	2.17 ±0.48	0.0050 ±0.0034	0.077	0.55	37	6.1
	No registration	7.08 ±2.13		0.826 ±0.035	6.66 ±1.69	0.0122 ±0.0040	0.000	0.00

The order averages each method's ranks on TRE mm, SDlogJ, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 10 pairs. Two neighbours are tied when a paired Wilcoxon test on the shared pairs does not find the upper one better on TRE mm (5%): the places a method could hold are then shown under its own. The best value of each column is in colour. Click a row to see its images, a column title to sort by it. Time and memory measured on Laptop.

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

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
corrField	0.954	0.946	0.952	0.952	0.933	0.951	0.954	0.955	0.947	0.957
ConvexAdam (MIND)	0.962	0.955	0.958	0.957	0.940	0.955	0.961	0.956	0.951	0.962
Demons	0.961	0.943	0.953	0.958	0.934	0.950	0.961	0.958	0.915	0.960
No registration	0.861	0.851	0.864	0.871	0.788	0.775	0.815	0.737	0.812	0.885

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: 10 test pairs drawn with seed 2026 among the 77 outside development, and 10 development pairs. The pool is the 87 participants whose two scans share at least three annotated nodules, so that a pair's landmark error averages at least three points.
Truth: landmarks by experts contours by a program the five lung lobes and the airways, TotalSegmentator (tasks total and lung_vessels) run by us on each scan the dataset's nodule positions, one point a nodule on each scan, linked by the nodule's lesion ID: 3 to 6 a pair
Source: zenodo.org/records/14223624
zenodo.org/records/14673658
doi.org/10.5281/zenodo.15094631
luna25.grand-challenge.org
github.com/DIAGNijmegen/luna25-baseline-public
doi.org/10.7937/TCIA.HMQ8-J677
Licence: images CC BY 4.0, annotations CC BY-NC 4.0

10 pairs for tuning, never tested

100069_2000_2001, 100095_1999_2000, 100438_2000_2001, 111668_2000_2001, 205113_1999_2000, 129831_2000_2001, 211700_2000_2001, 107933_2000_2001, 124014_1999_2000, 131666_2000_2001