fidopt Interactive view
Known deformation, two images of one session

PSMA-PET-CT-Lesions (autoPET III and IV)

Whole-body CT (fixed) and PSMA-PET (moving) of one PET/CT session of a man with suspected or diagnosed prostate cancer, the PET deformed by a known smooth transform (PSMA-PET-CT-Lesions, autoPET III and IV).

30 test pairswhole bodysame patientfixed CT, moving PETdelivered aligned
Look at the images
Read before comparing (9)
  • Truth. The scanner aligns the CT and the PET of one session, so the corpus deforms the PET by a known cubic B-spline, knots 40 mm apart, coefficients within 15 mm. 300 points drawn in the CT's foreground are paired through it, each pair inside both images, and both label maps are warped with the image.
  • Motion. The scanner's alignment is not perfect where the body moves: the PET starts 0.6 to 2.1 minutes after the CT and lasts 11 to 15 minutes. Breathing, bowel and a filling bladder make the images disagree there by an unmeasured amount: a method that matches the PET's own diaphragm or bladder to the CT's is scored off by it, as every method is.
  • Baseline. With no registration the 30 test pairs score a TRE of 4.63 mm (4.40 to 4.84) and an organ Dice of 0.78 (0.74 to 0.81).
  • Labels. Each map is made once on the CT's grid and given to both images, so it scores the deformation, not two readings. The organs are TotalSegmentator's 24 organ classes run on the CT; its 93 other classes are left out. The adrenal glands and the one kidney cyst are small, and their Dice moves the most.
  • Lesions. PSMA-avid tumour lesions drawn on the PET by one reader with 3 years of experience, pre-segmented by a threshold and corrected slice by slice, then validated by experts (version 2, 26 February 2026). On the CT's grid they keep the PET's voxels, 2.7 to 4.1 mm wide. Scored as dice_lesions and hd95_lesions, shown, never ranked; PSMA_206e0f3bf8cb8ed3_1998-11-17 has none.
  • Images. The field is asked on the CT's grid, 0.98 mm in plane; the PET is resampled onto it linearly through the deformation, in Bq/mL as delivered, not SUV. prepare.py sets the GE CTs' -3024 around the reconstructed circle to -1024, so that the foreground, the CT above its Otsu threshold (body and table), is not that circle.
  • Defacing. TCIA's open copy blanks a box over the skull base and neck of each image, separately: near the neck one image shows anatomy where the other is blank. Points where the PET is blank, there or on the table, score only how a method extends its field.
  • Pool. 597 studies of 378 men at LMU University Hospital Munich, the dates shifted, on the Siemens Biograph mCT Flow 20 (251), the GE Discovery 690 (230) and the Siemens Biograph 64-4R TruePoint (116). A case is a study, named after the patient and the study date.
  • Outside numbers. None: the pair is the scanner's and the deformation ours. autoPET III and IV segment these images; Learn2Reg 2026's PSMAReg trains and validates on these 597 studies but registers two sessions of a patient, another problem.

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

AccuracySmoothnessCost
#MethodDiceChangeHD95 mmTRE mmrTRESDlogJFolds %Time sRAM GBGPU GBImages
No registration0.777 ±0.0074.62 ±0.094.63 ±0.040.0042 ±0.00010.0000.00
1 result not ranked: an older run, pairs still missing, or a failure on every pair
AccuracySmoothnessCost
#MethodDiceChangeHD95 mmTRE mmrTRESDlogJFolds %Time sRAM GBGPU GBMachineImages
ConvexAdam (MIND)GPUapp29/30 pairs0.628 ±0.021−0.14911.70 ±0.7813.55 ±0.550.0122 ±0.00050.1685.86587.020.4Laptop

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

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 PSMA-PET-CT-Lesions (autoPET III and IV)
The first test caseClick to look at every case.
Test pairs
30 test studies drawn by scanner with seed 2026, then 28 development studies, from a pool of 318 patients, those outside AutoPSMA_CTPET_Syn, PSMALesionsPET and PSMAFollowUp when it was drawn; the four datasets now share no patient. One study a patient: the one already drawn, else the earliest. The first three development studies are each scanner's first in alphabetical order.
Truth
contours a known deformation landmarks a known deformation lesions a known deformation TotalSegmentator's 24 organ classes run by us on the CT, one map on the CT's grid given to both images, then carried through the known deformation 300 points drawn in the CT's foreground, paired through the known deformation
Licence
CC BY 4.0
28 pairs for tuning, never tested

PSMA_0179419e313f7d8c_2002-05-03, PSMA_024e1c55ab49d220_1997-08-27, PSMA_01a52e26ce5b5e26_1999-10-15, PSMA_53c3a9d0f51745b6_2000-11-16, PSMA_c81bde6455d07a32_2003-11-12, PSMA_7fd33c03d8712a0b_2001-07-18, PSMA_ddf308f3025bd9c8_2001-10-09, PSMA_798b3279d6980a3b_2002-01-25, PSMA_9724be80190a0274_2001-01-31, PSMA_0d2d8c9b29ea6861_2003-03-19, PSMA_75c11a592f49e915_2001-11-09, PSMA_a3d54951d96f5563_1999-03-24, PSMA_c605989163e81d3c_2002-01-08, PSMA_698f9e802d84253c_1997-06-26, PSMA_ae15e8ff542ee5a1_2000-12-21, PSMA_995fbaec49f131ce_1999-03-31, PSMA_41260c3678449a2f_1998-06-10, PSMA_907f1345abedf4fa_1997-05-21, PSMA_06612db91891b34c_2000-11-14, PSMA_834460a6d4330410_2003-01-15, PSMA_f751ed4845fddae7_1999-03-24, PSMA_d44c406a6988dbd6_2001-06-05, PSMA_d7032f7f9ec70157_2001-09-07, PSMA_e7b13a919f814300_2001-12-21, PSMA_614aab262dcced0a_1998-09-08, PSMA_18f866bfa3d793d4_1999-06-25, PSMA_0a3fdc59c5e700d8_2000-11-23, PSMA_572ba4a3592c01f2_1999-01-06