HaN-Seg, T1 MR to CT

T1-weighted MR (source) and planning CT (target) of the head and neck of one radiotherapy patient at the Institute of Oncology Ljubljana, the MR registered onto the CT by the corpus (HaN-Seg, public set 1).

10 test caseshead and necksource MR, target CTresearch only

Look at the images

Read before comparing (12)

Case. The planning CT and the T1-weighted MR of one patient, taken 0 to 11 days apart in the same thermoplastic mask, each on its own grid and in its own frame: the deposit does not register them.
Registration. The CT is never resampled: it is cut to the box the MR reaches, and the MR is resampled into it once, linearly, through a stored transform. The start is a translation matching the two bodies' areas slab by slab along the head-foot axis; then IMPACT-Reg 1.8.7's ConvexAdam_Composite (MIND features, an affine stage on mutual information, then coarse and fine deformable stages) with regularisation 10 instead of the preset's 1.25.
Choice. The recipe was chosen on HaNSeg's 20 patients, none of them here: Elastix_IMPACT_Static failed the check below on 16; ConvexAdam_Composite at regularisation 1.25, 2.5, 5, 10 and 20 failed it on 7, 5, 2, 1 and 1 (a mean Dice more than 0.02 below the start's or the rigid's, or a bone, the brain or the cord losing more than 0.25).
Check. TotalSegmentator (total_mr and vertebrae_mr) on both images, matched by name, inside the CT's body where the MR holds data: over the 5 to 28 structures both hold (spinal cord, brain, vertebrae one by one, discs, clavicles, lungs), the median mean Dice is 0.47 after the start, 0.65 after an elastix rigid on mutual information and 0.66 after the transform kept (0.63 to 0.72). Four test cases (case_03, case_16, case_19, case_20) were looked at in three planes.
Exception. ConvexAdam failed the check on case_18 (vertebra C7), case_20 and case_35 (a mean Dice 0.02 to 0.03 below the rigid's): they keep an elastix rigid. case_04 and case_01 stay 0.013 and 0.005 below their rigid.
Storage. The transforms hold each displacement sampled every 3 mm in tenths of a mm (23 MB): it departs from the full field by 0.16 to 0.43 mm at the 99th percentile, 2.9 mm at most. Folding: 0.005% of the region on case_16 and 0.001% on case_17, none elsewhere.
Residual. The neck bends otherwise, the jaw and tongue move and the swallowing structures change between the scans; TotalSegmentator's MR labels are themselves coarse. What the registration leaves misaligned counts as error for every method, and no method reaches an MAE of 0.
Mask. The CT's body (above -500 HU, the couch and the thermoplastic mask cut away, holes filled slice by slice) where the MR holds data through the transform: inside its grid and off the deposit's blank voxels, whole slices and corners at 0. The MR's field bounds it: the head and neck, on case_19 the neck and upper chest. On a slice where the nostrils or an ear canal open to the outside, their air stays out.
Values. The CT in Hounsfield units from -1000 to 3000, as the provider clipped it; dental fillings and their streaks are in it (case_03, case_20). The MR in the scanner's units; the deposit stores case_01's, case_06's and case_20's from 32768 up, and prepare.py takes 32768 off them, so that every MR's blank is 0.
Baseline. Inside the mask on the 10 test cases, with KonfAI's criteria at CT's range of 4095, a constant at each CT's median there scores an MAE of 164 (142 to 185) HU, a PSNR of 22.0 (20.7 to 23.1) dB and an SSIM of 0.37 (0.32 to 0.44).
Overlap. HaN-Seg's public set 1 also gives HaNSeg its planning CTs; its 20 test and development patients are held out here, so no CT is both a method's input there and a target here. No other descriptor names HaN-Seg, Ljubljana or Slovenia.
Licence. No derivatives: the cases are never uploaded, and the transforms, fields computed from these CC BY-NC-ND images, are kept on the reference machine and not distributed, so the set builds there only. prepare.py pins each transform's sha256.

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

Under a licence that is not open

Methods whose code or weights are released for non-commercial use, or under share-alike or custom terms: shown for reference, ranked nowhere.

		Accuracy					Cost
#	Method	MAE	PSNR dB	SSIM	Anatomy Dice	Anatomy HD95 mm	Time s	RAM GB	GPU GB	Machine	Images
	KonfAI synthesis app · TotalSynth MR to CT, KonfAI appGPUapptrained on CC-BY-NC-4.0 dataolder run	104.0 ±16.9	26.10 ±1.16	0.642 ±0.078	0.441 ±0.089	13.89 ±4.03	26	3.7	4.4	Workstation
	nnsyn · MR to CT, head and neckGPUapptrained on CC-BY-NC-4.0 dataolder run	101.6 ±18.3	26.13 ±1.29	0.679 ±0.069	0.505 ±0.092	11.02 ±4.18	34	4.6	2.5	Workstation

MAE: mean absolute error of a synthetic image against its target, inside the body, in the image's units. PSNR: the same by signal-to-noise ratio, higher is better. SSIM: the same by structural similarity, 1 is perfect. Anatomy Dice and HD95: TotalSegmentator run on the synthetic and on the target image, the structures compared; shown, not ranked.

Case by case

Measure

MAE	1	2	3	4	5	6	7	8	9	10

Click a score to look at that result: the target image, and over it the method's synthetic image. Underlined scores have their image; for the others you see the case without the result.

About the data

Test cases: 10 test and 10 development patients drawn with seed 2026 from the 22 of public set 1 that HaNSeg holds in neither its test nor its development cases (draw.held_out), case_19 included: HaNSeg gives methods those 20 planning CTs, which would be targets here. case_23 and case_40 are left over.
Truth: target image by a program the patient's planning CT in Hounsfield units, cut to the box of its grid the MR reaches; the MR registered onto it by the corpus with IMPACT-Reg 1.8.7 (ConvexAdam_Composite, regularisation 10), three patients rigidly
Source: zenodo.org/records/7442914
doi.org/10.1002/mp.16197
lit.fe.uni-lj.si/en/research/resources/HaN-Seg
Licence: CC BY-NC-ND 4.0

10 cases for tuning, never tested

case_18, case_33, case_37, case_06, case_36, case_01, case_17, case_30, case_02, case_31