µ-RegPro

Transrectal ultrasound (fixed) and T2 MR (moving) of the prostate of one patient (µ-RegPro challenge, MICCAI 2023, holdout test set).

35 test pairsprostatesame patientfixed US, moving MRdelivered alignedresearch only

Look at the images

Read before comparing (7)

Alignment. Each image keeps its own grid in the challenge's common frame, used as delivered; no transform is computed. An elastix translation from the intensities left 20 of the 35 pairs worse than that frame, up to 89 mm, and is not used.
Baseline. With no registration the 35 test pairs start at a landmark error of 24.8 mm (11.4 to 39.1), a gland Dice of 0.22 (0.00 to 0.52) and an HD95 of 25.7 mm.
Landmarks. 2 to 4 a test pair (median 3, 93 in all), 1 to 5 on the training pairs, so a pair's error rests on very few points. The challenge's design document announces at least three a case; 15 of the 35 test pairs hold two.
Truth. The gland was contoured on the MR under the SmartTarget trial's protocol and, on the ultrasound, by an automatic method two research fellows edited by hand. Each landmark was placed by a radiologist or a researcher and agreed by at least one other annotator, and an observer of over ten years' experience checked every case. The challenge gives no figure for the annotators' error.
Bar. It falls short of a main leaderboard's bar: the landmarks number far fewer than ten a pair, against an error of their own nobody measured.
Patients. The 35 test pairs come from 26 patients: six have two or three ultrasound volumes, each paired with the same MR (cases 79 and 80, 86 to 88, 91 and 92, 93 to 95, 97 and 98, 105 to 107), so those pairs' errors are not independent. None of the 46 training patients' MRs is among the test pairs.
Outside numbers. The challenge registered the ultrasound onto the MR and scored otherwise: its leaderboard on these 35 pairs does not compare.

		Accuracy					Smoothness		Cost
#	Method	Dice	Change	HD95 mm	TRE mm	rTRE	SDlogJ	Folds %	Time s	RAM GB	GPU GB	Images
1	ConvexAdam (MIND)GPUapp	0.370 ±0.073	+0.152	22.31 ±2.43	18.58 ±2.94	0.1383 ±0.0219	0.112	0.66	4	2.0	0.5
	No registration	0.218 ±0.058		25.65 ±2.30	24.76 ±2.54	0.1842 ±0.0189	0.000	0.00

The order averages each method's ranks on Dice, HD95 mm, TRE mm, SDlogJ, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 35 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. 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	30	31	32	33	34	35
ConvexAdam (MIND)	0.669	0.216	0.128	0.471	0.472	0.470	0.087	0.132	0.290	0.475	0.750	0.344	0.527	0.184	0.213	0.210	0.621	0.497	0.531	0.404	0.602	0.619	0.619	0.075	0.314	0.451	0.258	0.001	0.646	0.540	0.434	0.511	0.058	0.085	0.045
No registration	0.515	0.051	0.030	0.358	0.277	0.253	0.022	0.032	0.252	0.266	0.523	0.092	0.280	0.105	0.110	0.115	0.512	0.373	0.178	0.099	0.452	0.447	0.446	0.022	0.137	0.271	0.093	0.000	0.322	0.320	0.216	0.408	0.006	0.026	0.023

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 challenge's 35 holdout test pairs, all of them, released with their truth in April 2026. The 10 development pairs are among its 65 training pairs; its 8 validation pairs are not used.
Truth: contours by experts landmarks by experts the prostate gland in both images, contoured or corrected by hand 2 to 4 structures a pair marked in both images (a calcification, a cyst, the apex, ...), each taken at its centre
Source: doi.org/10.5281/zenodo.19414480
zenodo.org/records/7870105
muregpro.github.io
zenodo.org/records/7844908
github.com/muregpro/Baseline-Networks
Licence: CC BY 4.0 (the holdout: the test pairs), CC BY-NC-SA 4.0 (the training pairs: development)

10 pairs for tuning, never tested

case000000, case000001, case000002, case000020, case000060, case000035, case000059, case000047, case000005, case000049