fidoptICASSP-2024 3D-CBCT challenge, simulated low-dose CBCT to CT
Chest cone-beam CT simulated at a tenth of the clinical dose from a LIDC-IDRI CT (source) and the noiseless reconstruction of the same simulated scan (target), on one grid (ICASSP-2024 3D-CBCT challenge, validation set).
10 test caseslungsource CBCT, target CT
Read before comparing (9)
- Simulation. The organisers simulated cone-beam sinograms from LIDC-IDRI's chest CTs with the ASTRA toolbox, without Monte Carlo physics, with photon, electronic and quantification noise and a simplified scatter and flat-field correction. No cone-beam scanner imaged these patients: no real scatter, beam hardening, motion or detector lag. Graded synthetic and ranked apart: a method that does well here denoises this simulation.
- Case. The source is the provider's FDK reconstruction of the low-dose sinogram, the target its 'over-sampled noiseless reconstruction', both 256 voxels a side of 1.171875 mm, as delivered. The clinical-dose reconstruction and the sinograms are not used.
- Size. The organisers resized each CT to 256 voxels a side whatever its extent: on the twenty cases built first a voxel is really 1.02 to 1.56 mm in the slice and 1.03 to 1.68 mm along the body. A chest reads up to a quarter smaller than it is, and is stretched or squeezed along the body by up to a third.
- Values. Not Hounsfield units: the target is about HU / 1000 + 1, air near 0.02 and water near 1.02, so a method that writes Hounsfield units must convert. truth.dynamic_range is 4.095, CT's 4095 HU on this scale, so that the PSNR and the SSIM read as on the corpus's other CTs.
- Target. The noiseless reconstruction, not the CT: 0.02 to 0.05 from the resized CT in the middle slices. Towards both ends the cone beam sees less of each slice and both images darken; the target is still 0.05 to 0.15 under the CT at the mask's end slices.
- Mask. The body above 0.5 (-500 HU), the table detached and the holes filled, inside the scanner's field of view, a cylinder of 158 mm radius less two voxels, and inside slices 26 to 229, which every projection sees. The 300 mm cube cuts the sides of the larger chests, 0858 among the test cases.
- Baseline. Inside the mask on the ten test cases, at a range of 4.095, a constant at the target's median scores an MAE of 0.320 (0.244 to 0.352), a PSNR of 18.9 dB (18.6 to 20.8) and an SSIM of 0.56 (0.52 to 0.65). The source scores 0.203 (0.168 to 0.254), 24.1 dB (22.2 to 25.8) and 0.41 (0.33 to 0.49): its noise is what a method has to take out.
- Overlap. The challenge's patient n is LIDC-IDRI-(n+1): on the twenty cases built first each target correlates at 0.990 to 0.999 with that patient's CT, against 0.05 to 0.56 with another's. Of the corpus, only LIDCNodules and SAROSBodyCT hold LIDC-IDRI CTs, hence the six held out. A later dataset built on LIDC-IDRI leaves these twenty patients out.
- Outside numbers. None: the challenge ranked the mean squared error on 110 unreleased test patients. Its validation set is public with its targets and the entrants tuned on it, so a network trained for the challenge may have seen these test cases; LIDC-IDRI's CTs are widely used to train networks.
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 | Time s | RAM GB | GPU GB | Machine | Images |
| KonfAI synthesis app · TotalSynth CBCT to CT, KonfAI appGPUapptrained on CC-BY-NC-4.0 dataolder run | 972.5 ±13.7 | -47.60 ±0.10 | -0.000 ±0.000 | 30 | 3.8 | 4.4 | Workstation | ||
| nnsyn · CBCT to CT, thoraxGPUapptrained on CC-BY-NC-4.0 dataolder run | 311.6 ±9.4 | -39.06 ±0.23 | 0.000 ±0.000 | 4.2 | 2.6 | Laptop | |||
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.
Case by case
| 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
The first test caseClick to look at every case.
- Test cases
- 10 test patients drawn with seed 2026 among the 94 of the challenge's validation set (0801 to 0900) left once six are out, then 10 development patients. The six (draw.held_out) are LIDC-IDRI patients whose CT, of which the target is a reconstruction, the test or development cases of LIDCNodules or SAROSBodyCT show.
- Truth
- target image a known deformation the provider's 'ground truth', an over-sampled noiseless cone-beam reconstruction (FDK) of the patient's CT put in the simulated scanner, in the provider's unit (attenuation relative to water: air near 0, water near 1); the source is the FDK reconstruction of the same scan simulated with noise at 10% of the clinical dose, on the same grid, so nothing is registered
- Source
- sites.google.com/view/icassp2024-spgc-3dcbct/home
sites.google.com/view/icassp2024-spgc-3dcbct/data
sites.google.com/view/icassp2024-spgc-3dcbct/evaluation
sites.google.com/view/icassp2024-spgc-3dcbct/results
zenodo.org/records/8377374
doi.org/10.5281/zenodo.8377374
www.cancerimagingarchive.net/collection/lidc-idri
doi.org/10.7937/K9/TCIA.2015.LO9QL9SX - Licence
- CC BY 4.0 (the Zenodo record); the CTs it is made from are LIDC-IDRI's, CC BY 3.0 (TCIA)
10 cases for tuning, never tested
0802, 0848, 0851, 0815, 0878, 0899, 0807, 0829, 0838, 0810