Aariz cephalograms

Two lateral cephalograms, 2D radiographs of the head seen from the side, of two orthodontic patients from one device, with 29 landmarks of bone, teeth and soft tissue placed by orthodontists (Aariz, test split).

10 test pairshead and neckinter-patientfixed XRAY, moving XRAY2D, distances in mmaligned once by the benchmark

Look at the images

Read before comparing (11)

Pool. 1000 patients aged 12 to 62 in orthodontic treatment in Islamabad, imaged with seven devices; the provider splits them into 700 training, 150 validation and 150 test images. Only test images are used: a detector trained on the training split knows that split's truth.
Pairs. A device's test images, in the order of sha256("2026:AarizCeph:<id>"), are paired two by two, each patient in one pair, one ART Plus image left over: 27 ART Plus, 3 ProMax 2D and 10 ProMax with ProTouch pairs.
Unit. Millimetres on the radiograph, from the pixel size the provider gives for each device in cephalogram_machine_mappings.csv; the paper's table 2 prints 1.139 mm for the ProMax 2D, the CSV 0.139. prepare.py gives the ProMax JPEGs that 0.139 mm in place of their 199 dpi (0.128 mm), the size at which their anatomy matches the ART Plus.
Devices. At those sizes sella to nasion measures 61 to 83 mm on every ART Plus, ProMax 2D and ProMax with ProTouch image, and six other distances agree within 3%. The four other devices give medians of 24, 30, 45 and 118 mm: their files are at another scale than the CSV gives, so they are left out.
Images. As delivered, with the cephalostat's ear rods and nasion rest. The ART Plus PNGs show the face and the front of the skull, several with horizontal line artefacts, in three equal 8-bit components a pixel; the ProMax grey JPEGs show the whole head. No mask.
Truth. Two junior orthodontists traced each image on their own, then two seniors corrected the tracings. The corpus takes the senior file, not the mean of both that the paper takes, which keeps half of each correction. 55% of the senior points sit where the juniors' mean put them; the others moved a median of 0.68 mm, 251 of the 29,000 by more than 5 mm.
Precision. On a subset of 100 images read again, the seniors' second reading differs from their first by a mean of 1.21 and 1.35 mm, the juniors' by 1.47 and 1.65 mm: a point is good to a little over a millimetre.
Alignment. Each image keeps its own grid and pixels. The moving image's header carries one translation from elastix mutual information, the best of several starts, no annotation used. Every head faces the same way in the cephalostat: a rigid fit turns them by under half a degree.
Baseline. With no registration the 10 test pairs score a TRE of 12.6 mm (6.2 to 26.4), rTRE 0.040, and a 95th percentile of 19.2 mm on average, against 3.8 mm after an affine fit on the points. That is about ten times the seniors' repeat error.
Coverage. The 29 points lie on the face, the jaws and the skull base: nothing scores the cranial vault, the back of the head, the spine or the image between the points.
Outside numbers. None: the paper and its challenge, CEPHA29 (2023), rank landmark detection, and the pairing is ours.

		Accuracy			Smoothness		Cost
#	Method	TRE mm	Change	rTRE	SDlogJ	Folds %	Time s	RAM GB	GPU GB	Machine	Images
11–9	vismatch matchers · LoFTRGPU	7.87 ±2.42	−4.73	0.0247 ±0.0077	0.038	0.00	5	2.4	21.7	Laptop
11–9	vismatch matchers · RoMaGPU	7.89 ±2.98	−4.72	0.0242 ±0.0074	0.026	0.00	15	3.6	6.9	Laptop
31–9	vismatch matchers · ALIKED + LightGlueGPU	11.59 ±4.98	−1.02	0.0378 ±0.0176	0.067	0.00	4	2.7	2.1	Laptop
41–9	vismatch matchers · SIFT + LightGlueGPU	11.76 ±4.90	−0.85	0.0364 ±0.0135	0.106	0.00	4	2.6	0.6	Laptop
51–9	greedyCPUapp	12.53 ±4.51	−0.07	0.0401 ±0.0159	0.070	0.00	16	1.5		Workstation
61–9	ANTs SyNCPUapp	12.32 ±4.62	−0.29	0.0394 ±0.0162	0.132	0.00	36	0.8		Workstation
71–9	NiftyRegCPUapp	11.93 ±4.28	−0.67	0.0381 ±0.0149	3.219	26.79	19	0.7		Workstation
81–9	vismatch matchers · XFeatGPU	16.92 ±10.38	+4.32	0.0552 ±0.0357	0.583	0.27	4	2.6	0.5	Laptop
91–9	vismatch matchers · DISK + LightGlueGPU	142.31 ±297.00	+129.70	0.4774 ±0.9993	1.164	2.71	4	2.6	3.4	Laptop
	No registration	12.60 ±4.47		0.0403 ±0.0157	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.

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

TRE mm	1	2	3	4	5	6	7	8	9	10
vismatch matchers · LoFTR	12.76	6.37	4.88	5.45	11.15	3.65	8.18	13.29	6.06	6.92
vismatch matchers · RoMa	6.85	5.93	5.20	4.61	11.13	5.12	8.16	9.70	4.35	17.85
vismatch matchers · ALIKED + LightGlue	10.12	17.00	16.17	17.98	24.76	7.67	5.56	3.34	5.82	7.46
vismatch matchers · SIFT + LightGlue	11.32	16.27	11.46	5.66	17.65	4.87	9.29	9.06	5.19	26.80
greedy	12.56	9.58	26.55	20.08	11.05	6.13	8.06	9.77	13.83	7.69
ANTs SyN	12.41	9.59	26.69	19.94	10.65	5.66	8.16	9.32	13.74	7.04
NiftyReg	11.91	9.52	24.30	19.58	10.39	5.18	7.88	9.50	14.48	6.58
vismatch matchers · XFeat	24.36	23.43	53.26	7.64	14.45	15.48	7.14	5.78	4.84	12.84
vismatch matchers · DISK + LightGlue	10.87	14.65	1323.80	19.62	19.57	3.66	12.97	3.96	5.95	8.00
No registration	12.74	9.73	26.39	20.16	11.12	6.22	8.19	9.95	13.87	7.69

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

The first test case of Aariz cephalograms

The first test caseClick to look at every case.

Test pairs: 10 test pairs drawn with seed 2026 among the 30 outside development, the devices in proportion: 7 ART Plus, 1 ProMax 2D and 2 ProMax with ProTouch. The 40 pairs come from the provider's test images of the three devices whose pixel size holds; the 10 development pairs are 7 ART Plus and 3 ProMax with ProTouch.
Truth: landmarks by experts the 29 cephalometric landmarks of the dataset's senior orthodontists' file on each image, 15 of bone, 8 of teeth and 6 of soft tissue: traced by two junior orthodontists each on their own, then reviewed and corrected by two senior orthodontists (Khalid et al. 2025)
Source: doi.org/10.6084/m9.figshare.27986417
doi.org/10.1038/s41597-025-05542-3
arxiv.org/abs/2212.04808
github.com/manwaarkhd/aariz
Licence: CC BY 4.0

10 pairs for tuning, never tested

cks2ip8fr2a4p0yuffc5t82n9_cks2ip8fz2anl0yufe59x6gbe, cl5lg05ul01gq074kb4xm8mml_cl5lg05un01hm074k3dwy9l5q, cks7u9tyg5z7j0ytobkyxce57_cks2ip8fr2a3g0yufgxtf43y5, cks2ip8fx2ahw0yufhn9xb0jn_cks2ip8fz2anv0yufdoup0vpn, cl5lg05ud017a074k37pdhlni_cl5lg05ur01lm074kaznu2ncg, cks2ip8fy2al80yuf9w4x9x27_cks2ip8g52b3u0yuf558g4ihc, cks2ip8fr2a2r0yuf718nerfe_cks7u9tyf5z5g0yto3kvl0waz, cks2ip8g42aze0yuffg2368n3_cks2ip8fy2ak90yuf23jt2wlv, cl5lg05uc016m074k8tltapbp_cl5lg05uf019m074k8q27h7oy, cks2ip8g32aw20yufcvum9zg4_cks2ip8fw2aek0yufd5l89lt6