TrackRAD2025, two frames of a sagittal cine MR sequence

Two frames of one patient's sagittal cine MR on an MR-linac, the tracking target drawn by people on each: the first labelled frame and the one where the target lies farthest from it (TrackRAD2025).

19 test pairsthorax, abdomen and pelvissame patient, different framefixed MR, moving MR2D, distances in mmdelivered alignedresearch only

Look at the images

Read before comparing (8)

Pair. The fixed frame is the first with a mask, the moving frame the one whose target's centre lies farthest from it: the widest motion the sequence holds, chosen on the masks alone, never on a method's result.
Pool. Only patients whose target moves 5 mm or more enter: 19 of the 30 final-testing and 36 of the 50 training patients. In the others the two frames overlap at 0.8 or more before any registration, so the scores say nothing of targets that stay still.
Baseline. With no registration the target overlaps at a Dice of 0.53 on average on the 19 test pairs (0.00 to 0.84), its centre 5.5 to 27 mm away. On some pairs its outline changes as the organ moves through the 5 to 8 mm slice (C_002: 371 pixels then 670, no overlap), which no in-plane registration brings to a Dice of 1.
Ranked apart. A Dice on one target of a few hundred pixels, on 19 pairs, carries the score.
Truth. Frames without a mask are not used, whether the target left the plane or was not drawn. The second observers' masks given for some patients are not used.
Orientation. Each frame is stored with its rows reversed, so that a viewer shows the patient upright with the front on the left; coordinates run in mm to the back and to the feet. The header's origin is (0, 0, 0) at centre C: positions are not scanner coordinates. The provider's 1 mm pixels and 16-bit values are unchanged.
Timing. The time between the two frames is their index difference over the provider's frame rate, 1.3 to 8 frames a second; where the labelled frames were picked by hand it is not exact.
Overlap. TrackRAD2025_Seg takes its images from these pairs and keeps the split.

		Accuracy			Smoothness		Cost
#	Method	Dice	Change	HD95 mm	SDlogJ	Folds %	Time s	RAM GB	GPU GB	Images
11–4	DemonsCPUapp	0.773 ±0.133	+0.238	7.92 ±3.67	0.051	0.00	0
21–4	greedyCPUapp	0.744 ±0.133	+0.209	8.83 ±3.72	0.083	0.00	0
31–4	ANTs SyNCPUapp	0.726 ±0.147	+0.191	8.88 ±4.04	0.083	0.00	2	0.1
41–4	NiftyRegCPUapp	0.711 ±0.132	+0.176	10.19 ±3.84	0.060	0.13	1	0.1
44–10	vismatch matchers · ALIKED + LightGlueGPU	0.629 ±0.132	+0.094		0.025	0.06	4	2.2	2.1
64–10	vismatch matchers · DISK + LightGlueGPU	0.624 ±0.129	+0.089	11.93 ±3.74	0.009	0.00	3	2.0	3.8
74–10	vismatch matchers · SIFT + LightGlueGPU	0.616 ±0.129	+0.081	12.03 ±3.52	0.008	0.00	4	2.0	0.5
84–10	vismatch matchers · RoMaGPU	0.581 ±0.124	+0.046	12.70 ±3.34	0.006	0.00	8	3.1	6.9
94–10	vismatch matchers · XFeatGPU	0.561 ±0.128	+0.026	13.31 ±3.28	0.006	0.00	4	2.1	0.4
104–10	vismatch matchers · LoFTRGPU	0.546 ±0.123	+0.012	13.66 ±3.26	0.002	0.00	3	2.0	1.2
	No registration	0.535 ±0.126		14.04 ±3.20	0.000	0.00

The order averages each method's ranks on Dice, HD95 mm, SDlogJ, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 19 pairs. Two neighbours are tied when a paired Wilcoxon test on the shared pairs does not find the upper one better on Dice (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.

Dice: how much the organ contours overlap, 1 is perfect. HD95: distance between the contours' surfaces. 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	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19
Demons	0.854	0.735	0.882	0.884	0.751	0.779	0.938	0.907	0.020	0.957	0.842	0.922	0.854	0.860	0.886	0.807	0.873	0.932
greedy	0.909	0.398	0.862	0.899	0.802	0.777	0.939	0.908	0.118	0.947	0.785	0.838	0.866	0.828	0.900	0.587	0.863	0.906
ANTs SyN	0.894	0.848	0.310	0.879	0.770	0.801	0.936	0.906	0.000	0.955	0.828	0.839	0.897	0.832	0.899	0.420	0.862	0.921
NiftyReg	0.877	0.511	0.655	0.663	0.682	0.806	0.942	0.914	0.000	0.943	0.794	0.761	0.864	0.845	0.844	0.685	0.866	0.849
vismatch matchers · ALIKED + LightGlue	0.838	0.486	0.853	0.862	0.509	0.713	0.816	0.454	0.000	0.913	0.590	0.600	0.716	0.826	0.794	0.354	0.842	0.783
vismatch matchers · DISK + LightGlue	0.712	0.507	0.860	0.691	0.615	0.706	0.820	0.458	0.000	0.894	0.586	0.580	0.742	0.827	0.810	0.341	0.840	0.865
vismatch matchers · SIFT + LightGlue	0.753	0.591	0.853	0.521	0.463	0.741	0.837	0.438	0.000	0.904	0.618	0.597	0.751	0.811	0.818	0.372	0.843	0.791
vismatch matchers · RoMa	0.604	0.517	0.523	0.465	0.524	0.706	0.816	0.487	0.000	0.899	0.717	0.532	0.714	0.688	0.799	0.336	0.844	0.863
vismatch matchers · XFeat	0.604	0.521	0.241	0.461	0.463	0.706	0.853	0.459	0.000	0.879	0.590	0.569	0.684	0.828	0.780	0.376	0.836	0.800
vismatch matchers · LoFTR	0.657	0.486	0.259	0.401	0.524	0.706	0.816	0.441	0.000	0.843	0.566	0.532	0.684	0.688	0.797	0.372	0.837	0.769
No registration	0.604	0.486	0.214	0.328	0.463	0.706	0.816	0.441	0.000	0.817	0.566	0.532	0.684	0.763	0.769	0.372	0.837	0.763

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: Test: all 19 patients of the provider's final-testing set, centres A, B, C and X, whose target moves by 5 mm or more; centre D is not released. Development: ten of the 36 such patients among the provider's 50 labelled training patients, in the seeded order. The provider's two sets share no patient, and a patient gives one pair.
Truth: contours by experts The provider's target masks: drawn on every labelled frame by people (students, a medical physicist or radiation oncologists, depending on the centre), then all reviewed and corrected where needed by a medical physics doctoral student with four years of tumour tracking. The target is what the treatment tracks, a tumour as a rule; it is one small structure, not an organ.
Source: huggingface.co/datasets/LMUK-RADONC-PHYS-RES/TrackRAD2025
trackrad2025.grand-challenge.org/data
doi.org/10.1002/mp.17964
doi.org/10.1016/j.media.2026.104134
Licence: CC BY-NC 4.0, provider Hugging Face data licence

10 pairs for tuning, never tested

B_026, C_016, A_023, A_007, A_016, B_002, A_021, A_003, A_006, A_027