fidopt Interactive view
CT to CT, another reconstruction, dose or contrast phase, synthesis

NLST, sharp-kernel CT to soft-kernel CT

Two reconstructions of one low-dose chest CT acquisition from the National Lung Screening Trial: a sharp kernel (source: Siemens B50f, GE LUNG or BONE) and the maker's soft kernel (target: Siemens B30f, GE STANDARD).

30 test caseslungsource CT, target CT
Look at the images
Read before comparing (10)
  • Pairs. Two original axial series of one study and maker, one sharp and one soft kernel, matching in number of images, model, reconstruction diameter, slice thickness, kVp, mAs, effective mAs and pitch, and covering at least 25 cm. Letter-coded kernels, Philips and Toshiba are left out. A case is a participant's earliest screening with a pair, then its pair with the most images, then the thinnest slices.
  • Alignment. One acquisition reconstructed twice is aligned by construction; nothing is resampled or registered. prepare.py checks one frame of reference, slice positions equal within 0.01 mm and evenly spaced, the same orientation, pixel size and matrix, and each header's kernel. The difference images show noise and edges, no structure out of place.
  • Held out. 117236, 105262 and 122947 hold their reconstructions on different slice positions, 213641 has unevenly spaced slices, and 133841 a rescale slope that leaves fractions of a Hounsfield unit other than 4096/4095's.
  • Values. The series' 16-bit Hounsfield units. A rescale slope of 4096/4095, as on 109666, is rounded to whole units, at most 0.5 HU off. The GE CTs read -3024 HU outside the reconstructed circle, left as is.
  • Mask. The body on the soft-kernel image: above -500 HU, opened by one voxel in the slice to detach the table, its largest connected part, holes filled slice by slice. The table lies outside it; where the reconstructed circle cuts the body (103578, 211515) the mask stops at the circle.
  • Baseline. Inside the mask on the 30 test cases, with KonfAI's criteria at CT's default range, a constant at the target's median scores an MAE of 266 HU (148 to 344), a PSNR of 20.0 dB (18.3 to 24.5) and an SSIM of 0.16 (0.05 to 0.33). The source, the line to beat, scores 58 HU (34 to 222), 34.0 dB (23.2 to 38.8) and 0.85 (0.51 to 0.95).
  • Kernels. The draw is not stratified: the first ten test cases are eight Siemens B50f to B30f, one GE LUNG and one GE BONE. Only on 203067, GE LUNG on 1.25 mm slices, does the source score worse than a constant, 165 against 148 HU. 218480, GE BONE on a large patient streaked by photon starvation, has the largest source error, 222 HU.
  • Geometry. TotalSegmentator's KonfAI port segments the synthetic image and the target; the Dice and HD95 of its structures are shown beside the line, never ranked.
  • Outside numbers. Krishnan et al. (Medical Physics 2024 and 2025) trained kernel conversions on NLST pairs of the same kinds; the first warns that such pairs may be misaligned, which the checks rule out. No published number is scored as here. NLST is public, so a model trained on its kernel pairs may have seen these test participants.
  • Overlap. The pool leaves out by ID (prepare.py's IN_CORPUS) Lung250M's ten, also L2R_NLST's validation pairs, LUNA25's twenty, NLSTNodules' cases, NLSTSeg's sixty and LIDC-annot-NLST501's 501. Lung250M's 22 NLST training pairs and Learn2Reg NLST's training pairs publish no participant ID and cannot be checked.

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

Case by case

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

The first test case of NLST, sharp-kernel CT to soft-kernel CT
The first test caseClick to look at every case.
Test cases
30 test participants drawn with seed 2026, then 30 development participants, the last of the same seeded order. The pool is the 8,081 NLST participants with a sharp and a soft reconstruction of one screening acquisition, once the corpus's other NLST participants and LIDC-annot-NLST501's 501 are left out. Five drawn participants whose pair failed prepare.py's checks are held out.
Truth
target image as released the same low-dose CT reconstructed by the scanner with the maker's soft kernel (Siemens B30f, GE STANDARD), from the acquisition the source was reconstructed from with a sharp kernel (B50f, LUNG or BONE), on the same slices and pixels, in Hounsfield units as delivered
Licence
CC BY 4.0 (NLST's CT images, TCIA)
30 cases for tuning, never tested

116378, 101204, 214844, 133182, 215825, 201156, 214826, 214111, 205275, 216101, 217983, 202252, 202127, 118611, 206074, 103003, 130531, 206525, 130681, 120589, 215528, 110881, 211035, 201489, 115759, 217755, 115606, 216155, 112609, 108073