eess.IVSep 30, 2026
SaveRIQE: a NIQE-style reference model for Computed Tomography
Organizations: EOC
Abstract
The Natural Image Quality Evaluator (NIQE) scores an image by its statistical distance from a model fitted on pristine images, and its distributed model is fitted on photographs. We release the Radiology Image Quality Evaluator (RIQE), a NIQE-style model fitted on 3,792 full-dose slices from 158 patients of the public LDCT-and-Projection-data collection, with a declared intensity mapping, a manifest of every slice and a script that reproduces the fit. On 40 held-out patients, RIQE ranks reduced-dose reconstructions, simulated by projection-domain noise insertion, worse than the full-dose reconstruction of the same slice in 240 of 240 chest and 230 of 240 abdominal pairs, and ranks images with 20% more noise worse than their source in 97.5-100% of cases. Its preferences among filtered images, however, do not follow lesion signal. With a 4 mm, +10 HU lesion inserted in noisy abdominal slices, RIQE prefers bilateral filtering to the unfiltered image in every image up to a 32 HU residual, at which 29% of the lesion's matched-filter signal remains and its detectability index falls from 0.51 to 0.33; it never prefers Gaussian smoothing, which at the same 32 HU residual leaves 70% of the signal and a detectability index of 0.48. Fitted on photographs with the parameters published for NIQE, the same code ranks every simulated reduced-dose abdominal image better than its full-dose counterpart. RIQE is suited to ranking a degraded image against its source; under the conditions tested it should not be the sole criterion for selecting, comparing or tuning denoisers.
Figures & tables
Figure 1: Study design. Patients are split once, before any fitting. Hyperparameters are selected on an inner split of the fitting patients (Figure 2 ), the stratification analysis uses the fitting patients (Table 5 ), and all other results are computed on the 40 held-out test patients or on external data. Dashed boxes summarise the main finding of each stage.
| Experiment | Available | Selection | Excluded | Analysed (patients / units) |
|---|---|---|---|---|
| Model fitting | 158 fitting patients, 24 slices each | – | – | 158 / 3,792 slices |
| Hyperparameter search | inner split 118 / 40 patients | 4 evenly spaced slices per validation patient; all 24 for dose pairs | – | 40 / 160 slices, 480 dose pairs |
| Reduced dose (test) | 20 Siemens patients with reduced-dose series | all kept slices | 0 unscoreable | 20 / 480 pairs |
| Noise, blur, filtering (test) | 40 patients, 960 kept slices | 6 evenly spaced slices per patient | 0 unscoreable | 40 / 240 slices, 7,200 filtered |
| Denoisers on reduced dose | 20 Siemens test patients | slices of the test bank with a reduced-dose counterpart | 0 unscoreable | 20 / 120 slices |
| Inserted lesions | 20 abdominal test patients | central slice; sites by rule | 2 without 4 mm sites | 20 / 20 slices |
Table 1: Denominators. On the TCIA data, patients are the unit of the split and of every confidence interval; LDCTIQAC patients cannot be identified, and its intervals resample inferred source-slice groups (Section 3.11 ). Unscoreable: fewer than 72 valid patches in the image domain, the rule of RIQE; the baseline with the NIQE parameters uses its own rule (Section 3.8 ).
Figure 2: Validation results for all 168 hyperparameter settings (inner split of the fitting patients). Each open circle is one setting; the filled circle is the released model and the square the setting chosen by the criterion declared before the search. Settings that rank reduced dose correctly in at least 95% of pairs in both regions prefer 43.6% or more of the filtered images; those with the lowest preference rates, about 21%, rank fewer than half of the abdominal pairs correctly.
| Degradation | Chest (%) | Abdomen (%) |
|---|---|---|
| Reduced dose (10%/25% of routine), same slice | 100.0 [100.0, 100.0] | 95.8 [91.2, 99.6] |
| patients with every pair correct | 10/10 [69.2, 100.0] | 6/10 [26.2, 87.8] |
| Added noise, +5% of native | 85.0 [75.8, 93.3] | 89.2 [81.7, 95.8] |
| Added noise, +10% of native | 92.5 [85.0, 98.3] | 98.3 [95.8, 100.0] |
| Added noise, +20% of native | 97.5 [95.0, 100.0] | 100.0 [100.0, 100.0] |
| Added noise, +50% of native | 100.0 [100.0, 100.0] | 100.0 [100.0, 100.0] |
Table 2: Paired ranking of degraded images on the held-out test split. Percentage of pairs in which the degraded image scores worse than its source (100% is correct), with 95% confidence intervals from a bootstrap over patients. Reduced dose: 240 chest and 240 abdominal pairs from 10 and 10 patients, reconstructions simulated by the data providers through projection-domain noise insertion. Added noise and blur: 240 full-dose slices from 40 patients (120 chest and 120 abdomen, 20 patients each). When every patient is at 100%, the bootstrap interval collapses to [100, 100] and expresses no uncertainty; the second row gives instead the patients in which every pair is correct, with an exact binomial interval (for the rows at 100% with 20 patients: 20/20, [83.2, 100.0]). That is the probability that all pairs of a patient are correct, not the rate of correct pairs.
Figure 3: Change in the RIQE score of the 240 held-out full-dose test images under controlled degradation; positive values are worse. A: FBP-like noise added relative to the native noise of each image. B: Gaussian blur. C: the five filters, calibrated per image to the same residual standard deviation; values below the dashed line mean that the filtered image scores better than the original. D: the lightest filtering levels of C on an expanded scale. Points are medians over images; lines connect successive levels.
| Residual (HU) | ||||||
|---|---|---|---|---|---|---|
| Denoiser | 2 | 4 | 8 | 16 | 32 | 64 |
| Gaussian | 50 (+0.00) | 48 (+0.00) | 39 (+0.02) | 32 (+0.08) | 20 (+0.44) | 0 (+2.70) |
| Wavelet | 56 ( 0.01) | 56 ( 0.02) | 42 (+0.08) | 27 (+0.38) | 9 (+1.15) | 3 (+2.27) |
| TV | 82 ( 0.06) | 70 ( 0.09) | 50 ( 0.01) | 29 (+0.74) | 5 (+3.35) | 2 (+5.35) |
| NLM | 60 ( 0.01) | 57 ( 0.02) | 47 (+0.01) | 37 (+0.56) | 20 (+1.70) | 5 (+2.87) |
| Bilateral | 94 ( 0.05) | 96 ( 0.11) | 85 ( 0.15) | 59 ( 0.04) | 38 (+0.38) | 18 (+1.84) |
Table 3: Preference for filtered full-dose images: percentage of cases in which the filtered image scores better than its unfiltered source (test split, 7200 filtered images from 40 patients), and in parentheses the median score change (positive = worse). A full-dose image still contains noise, so a preference is not by itself an error; whether the preferred filtering removes signal is tested in Table 4 . Overall rate 41.2 [34.8, 48.3] %; counting only changes larger than 0.01, 0.05 and 0.1 score units: 37.7, 27.8 and 19.2%. Strength is the standard deviation of the residual inside the body.
| Residual (HU): preferred (%) / signal retained | ||||||
| Denoiser | 2 | 4 | 8 | 16 | 32 | 64 |
| Gaussian | 0 / 0.98 | 0 / 0.97 | 0 / 0.93 | 0 / 0.86 | 0 / 0.70 | 0 / 0.34 |
| Wavelet | 0 / 0.99 | 6 / 0.98 | 39 / 0.96 | 61 / 0.89 | 0 / 0.69 | 0 / 0.29 |
| TV | 61 / 0.97 | 56 / 0.94 | 6 / 0.87 | 0 / 0.72 | 0 / 0.39 | 0 / 0.05 |
| NLM | 0 / 0.99 | 0 / 0.97 | 0 / 0.88 | 0 / 0.66 | 0 / 0.23 | 0 / 0.06 |
| Bilateral | 100 / 0.98 | 100 / 0.94 | 100 / 0.83 | 100 / 0.63 | 100 / 0.29 | 0 / 0.01 |
Table 4: Metric preference against lesion signal. Substrate: 20 held-out full-dose abdominal slices (18 with room for the 4 mm lesions), one per patient, plus FBP-like noise at the level that separates full-dose from simulated quarter-dose abdominal images (local standard deviation 18.4 HU, global 28.9 HU); 64 noise realisations per condition. Each cell: percentage of images in which RIQE scores the filtered image better than the unfiltered one / median fraction of the matched-filter signal of a 4 mm, +10 HU lesion that survives the filter.
Figure 4: One held-out abdominal slice with FBP-like noise at the level that separates full-dose from simulated quarter-dose abdomen (local standard deviation 18.4 , global 28.9 ) and three lesions inserted by the rule of the experiment (circles, from largest: 10 mm , 6 mm , 4 mm ). Top: a region of the slice; bottom: a 24 mm region centred on the 4 mm lesion. The filters are calibrated to the same 32 residual. Below each column, the RIQE score of the whole slice without lesions, as in the experiment, lower meaning closer to the reference; the panels show the slice with lesions. Display window ; scale bars 20 and 5 mm.
Figure 5: A: percentage of the 18 held-out images in which RIQE scores the filtered image better than the unfiltered one. B: median fraction of the matched-filter signal of a 4 mm, lesion that survives the filter; the dashed line marks half. C: median detectability index of the same lesion after Gaussian and bilateral filtering, the two filters compared in the text, and without filtering (dashed line, shaded band); bars and band are 95% bootstrap intervals over patients. Up to 16 the two filters leave the same ; at 32 , where RIQE prefers the bilateral filter in every image, its is lower.
| Contrast | Patients | Null | ||
|---|---|---|---|---|
| Anchor: full vs reduced dose, same slices | 80 / 80 | 1.84 | 0.25 | 1.00 |
| Abdomen 5 mm: GE STANDARD vs Siemens B30f | 40 / 40 | 0.97 | 0.19 | 0.49 |
| Vendor: all GE vs all Siemens | 78 / 80 | 1.21 | 0.27 | 0.59 |
| Pixel-spacing tertiles, GE abdomen | 13 / 13 | 1.30 | 0.33 | 0.61 |
| Pixel-spacing tertiles, Siemens abdomen | 13 / 13 | 1.43 | 0.32 | 0.69 |
| GE: chest 1.25 mm vs abdomen 5 mm | 38 / 40 | 1.43 | 0.32 | 0.70 |
Table 5: Divergence between sub-models fitted on two groups of patients (fitting split). uses the same functional form as the score. Null: median of 200 patient-level permutations of the pooled groups. expresses the excess over the null in units of the excess between models fitted on the full-dose slices and on the reduced-dose reconstructions of the same slices. All permutation (the minimum attainable with 200 permutations). † Anatomy, kernel and slice thickness are not crossed in this collection; this contrast cannot serve as a clean anatomical control.
| RIQE | Photographs, | Photographs, | |
|---|---|---|---|
| (CT) | RIQE settings | NIQE parameters | |
| Reduced dose ranked worse, chest (%) | 100.0 | 100.0 | 98.3 |
| Reduced dose ranked worse, abdomen (%) | 95.8 | 31.2 | 0.0 |
| Noise +20% of native detected (%) | 98.8 | 70.8 | 3.3 |
| Noise +100% of native detected (%) | 100.0 | 78.8 | 39.6 |
| Filtered full-dose preferred (%) | 41.2 | 31.3 | 16.9 |
Table 6: RIQE against two models fitted with the same code on 119 CC0 natural photographs: one with the RIQE settings, one with the parameters published for NIQE ( , , , no masks). The reference NIQE model fitted on the LIVE photographs is not used. Rows 1–5: held-out test split. Rows 6–7: the LDCTIQAC 2023 training images, decoded as ; 95% intervals resample the inferred source slices; negative is the expected sign (lower score = closer to the reference). Within inferred group: rank correlation among the images of one group inferred from image similarity (the degraded versions of one source slice or of a few adjacent ones), median over groups. LDCTIQAC patients cannot be identified: these rows are exploratory.
Figure 6: Scores of the 1000 LDCTIQAC 2023 training images against the mean score of five radiologists. A: RIQE. B: the same code fitted on CC0 photographs with the RIQE settings. C: the same code fitted on the photographs with the parameters published for NIQE. Grey points are images, jittered horizontally; black points are medians per radiologist score. Lower scores mean closer to the reference, so the expected correlation is negative.