Source-linked AI summary
RIQE: a NIQE-style reference model for Computed Tomography
Fabio Mattiussi
TL;DR
Photographic NIQE scores do not by themselves establish how well a no-reference quality metric works on CT, especially for choosing among filtered images. This paper releases and tests a reproducible CT-fitted model; it ranks degraded images against their source, but its filtering preferences do not track lesion signal.
Problem
NIQE’s distributed reference model is fitted on photographs, while CT differs in noise texture, intensity values, and image content, so its relation to CT quality must be demonstrated.
Source: §1 · p. 1Method
The paper releases a reproducible NIQE-style model fitted on CT and tests both paired degradation ranking and preferences among filtered images.
Source: §1 · p. 2Results
RIQE ranks degraded images correctly against their source, but its preferences among filtered images do not follow inserted lesion signal.
Source: §1 · p. 2Takeaways & Limitations
RIQE is suited to paired ranking, but under the tested conditions it should not be the sole criterion for choosing, comparing, or tuning denoisers.
Source: §5 · p. 15The reduced-dose images are simulated by projection-domain noise insertion rather than acquired at lower dose, and the corpus covers one collection, two vendors, five protocol cells, and two anatomical regions.
Source: §6 · p. 15
Abstract
from arXiv · showhide
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.
1 Introduction
NIQE’s photographic reference model does not establish how its scores relate to CT image quality. This paper releases a reproducible CT-fitted model and tests whether it ranks degraded images correctly and selects filtered images that preserve lesion signal.
1 Introduction
- CT quality cannot be inferred from NIQE’s photographic reference model because CT noise, intensities, and content differ from photographs.The paper argues that the relationship between scores from that model and CT image quality must be demonstrated rather than assumed.Source: §1 · p. 1
- The paper tests whether a CT-fitted model ranks degraded images against their source and prefers filtered images that preserve diagnostic signal.A wrong filtering preference matters because aggressive filtering can erase low-contrast lesions while the metric improves.Source: §1 · p. 1
- RIQE is released with materials to rebuild it, ranks degraded images correctly against their source, but its filtering preferences do not follow inserted lesion signal.Across tested hyperparameter settings, no single model met both the dose-ranking and filtering criteria.Source: §1 · p. 2
2 Related work
NIQE provides the feature-and-model basis for this work, while prior CT adaptations leave reproducibility and validation questions. Human-scored CT data offer an external perceived-quality check distinct from diagnostic performance.
NIQE
- NIQE fits 36 natural-scene-statistics features per patch at two scales, using activity-selected patches during fitting and all patches during scoring.Its reported patch-size and activity-fraction ranges are based on photographs, and the paper examines why they do not transfer directly to CT.Source: §2 · p. 2
NIQE refitted on CT
- A prior CT refit by Gunawan et al. reports 16 × 16 patches and p = 0.32, but does not publish its model parameters or code.Its HU-to-luminance conversion is unstated, and its tuning criterion assumes the denoiser ordering that the metric is meant to measure.Source: §2 · p. 2
Human-scored CT quality
- LDCTIQAC supplies radiologist quality scores, but its images derive from about 100 slices of seven patients and are not directly comparable with this study’s results.The paper uses the training data as an external perceived-quality check, distinct from diagnostic performance.Source: §2 · p. 2
3 Methods
The study uses public chest and liver CT data, with a patient-level split and simulated reduced-dose reconstructions for a subset. It specifies an HU-to-luminance mapping because that conversion determines the model features.
3.1 Data
- The study uses the public LDCT-and-Projection-data collection’s reconstructed series, covering five full-dose protocol cells across Siemens and GE chest and abdomen scans.The collection version used releases chest and liver data; the head subset remained under controlled access.Source: §3.1 · p. 2
- For 100 Siemens patients, reduced-dose images are reconstructions made after projection-domain noise insertion to 10% routine dose for chest and 25% for abdomen.They use the same settings and slice positions as full-dose images and are not independent low-dose acquisitions; patients were split before fitting, with other results evaluated on 40 held-out patients or external data.Source: §3.1 · p. 2, Figure 1 · p. 3
3.2 Intensity mapping
- The HU-to-luminance mapping uses a −1000 to +1000 HU window and retains the lung parenchyma; 0–255 scaling is kept without 8-bit quantisation.The mapping is part of the model specification because it determines the features.Source: §3.2 · p. 3
3.3 Field-of-view and body masking
The method keeps patches inside the body and field of view, defines NIQE-style features and scoring choices, and builds a patient-separated pristine fitting corpus using declared slice criteria.
3.3 Field-of-view and body masking
- GE slices contain 21.3% of pixels outside the reconstructed field of view, so patches must lie inside that field and at least 90% inside a body mask.The same image masks are used for its degraded and filtered versions, and reduced-dose images use the corresponding full-dose masks so the scored domain stays fixed.Source: §3.3 · pp. 3–4
3.4 Features and model
- The model uses NIQE-style 36-dimensional patch features, fits their mean and covariance, and scores images by a covariance-based distance with a pseudoinverse.At least 72 valid patches are required as a stability choice; the declared constant C also affects how absolute noise amplitude enters the features.Source: §3.4 · p. 4
3.5 Pristine corpus
- The pristine corpus applies declared per-slice inclusion criteria, keeps 24 evenly spaced slices per patient, and splits patients before fitting.The resulting corpus has 4,752 slices from 198 patients; an attempted local-σ artefact criterion was disabled, so motion is not detected directly.Source: §3.5 · p. 4
3.6 Test images
The test bank combines same-slice dose comparisons with synthetic noise, blur, and calibrated filtering experiments on selected patient slices.
3.6 Test images
- Six evenly spaced slices from each test patient received synthetic degradations and filters, yielding 240 slices and an 18,840-image bank.The subsample kept the image bank within the computing budget.Source: §3.6 · p. 5
- 240 chest and 240 abdominal pairs match full-dose slices to reduced-dose slices from the same patients and table positions.Pairs were drawn from 10 patients per body region.Source: §3.6 · p. 5
- Added FBP-like noise increased the median local body standard deviation by 5%, 10%, 20%, 50%, or 100%, using it as a proxy for native noise.The median includes anatomical structure as well as noise.Source: §3.6 · p. 5
- Fixed-level FBP-like noise ranged from 1 to 90 HU across 14 levels, with independent repeats at 5, 10, and 20 HU to estimate the score noise floor.The median absolute score difference between repeated realizations served as the noise floor.Source: §3.6 · p. 5
- Gaussian blur ranged from σ = 0.5 to 3 pixels, while five filters were calibrated to residual changes of 2, 4, 8, 16, 32, or 64 HU.The filters were Gaussian, total variation, bilateral, non-local means, and wavelet BayesShrink; residual measures image change, not noise removal.Source: §3.6 · p. 5
3.7 Inserted lesions
Inserted soft-edged lesions and matched-filter measurements were used to assess filtered-image signal retention under controlled noise and filtering conditions.
3.7 Inserted lesions
- Soft-edged 4, 6, and 10 mm lesions at +10, +15, and +25 HU were placed at up to 16 selected homogeneous soft-tissue sites per size.Sites had local mean 40–80 HU and local standard deviation below the image median; most were in muscle because enhanced liver fell outside the selection range.Source: §3.7 · p. 5
- The substrate was the central kept slice from 20 abdominal test patients, with room for 4 mm lesions in 18 patients, and added FBP-like noise was calibrated to σn = 18.4 HU locally.The local noise level was set to separate full-dose from reduced-dose Siemens abdomen; GE abdomen lacked a reduced-dose series.Source: §3.7 · p. 5
- For each site, 64 matched noise realizations were applied with and without the lesion, and each filter strength was calibrated once on the noisy slice without lesions.The background slice was held fixed across realizations.Source: §3.7 · p. 5
- Retained lesion signal and non-prewhitening matched-filter detectability were both reported because filtering changes the variance in the detectability denominator.The fixed background removes anatomical variability but not anatomy–nonlinear-filter interaction; the observer is not calibrated to human performance, and RIQE scores the lesion-free filtered slice.Source: §3.7 · pp. 5–6
3.8 Photographic baselines
The study specifies photographic refits, revises its validation selection criterion, and scales model divergence against a simulated dose-reduction reference.
3.8 Photographic baselines
- The photographic baseline refit used 119 public-domain or CC0 photographs, comparing RIQE settings with published NIQE parameters rather than the distributed NIQE model.With NIQE parameters, 512 × 512 CT slices yielded up to 25 non-overlapping patches, and scoring required at least two defined patches.Source: §3.8 · p. 6
3.9 Hyperparameter selection
- The search covered 168 settings, but the declared criterion was revised after validation showed it selected a setting with only 44.4% correct reduced-dose abdomen ordering.The revised criterion prioritized at least 95% correct reduced-dose ranking in both chest and abdomen, then filtered-preference rate, relative-noise detection, and stability; both choices were tested.Source: §3.9 · p. 6
3.10 Divergence between models
- Model-divergence contrasts used 200 patient-level repartitions for the null distribution and expressed excess over sampling noise in units of excess from simulated dose reduction.The reference scale compared models fitted on full-dose and reduced-dose reconstructions of the same slices from 80 Siemens fitting patients.Source: §3.10 · pp. 6–7
3.11 Statistics
Uncertainty and significance analyses account for patient-level dependence, while exploratory LDCTIQAC grouping lacks patient-level intervals.
3.11 Statistics
- TCIA 95% confidence intervals used 2,000 patient-level bootstrap resamples, keeping each patient's images together.Sign tests counted patients rather than slices.Source: §3.11 · p. 7
- When every patient was at 100%, the interval collapsed to a point, so the analysis also reported the number of patients with every pair correct and an exact binomial interval.Source: §3.11 · p. 7
- LDCTIQAC analysis was exploratory because patient identifiers were unavailable; inferred image links formed 69 groups, but no patient-level interval was available.Some groups merged adjacent slices, and clustering did not recover patients.Source: §3.11 · p. 7
- Preference analyses counted score changes only above ε = 0.01, 0.05, or 0.1 and compared noise-induced changes with the noise floor.A separate analysis dropped image covariance to distinguish mean-feature shifts from covariance changes.Source: §3.11 · p. 7
4 Results
RIQE is constrained by CT-specific patch statistics, and the tested settings reveal a trade-off between ranking dose degradation and preferring filtered images. On held-out data, it ranks degraded images against their sources well, but its filtering preferences can conflict with lesion signal; protocol-specific models and external checks further qualify its use.
4.1 Constraints that CT imposes on NIQE
- Patch constraints led to testing smaller patches: 96 × 96 left too few patches for per-image covariance, while P = 24 yielded defined scores for all 18,840 test images.At P = 32, 658 images (3.5%) lacked scores; CT’s bone–air interface also made a threshold relative to maximum activity retain only 2–6 patches per slice at p = 0.75.Source: §4.1 · p. 7
4.3 Paired ranking of degraded images
- 240/240 chest pairs and 95.8% [91.2, 99.6] of 240 abdominal pairs ranked simulated reduced-dose images worse than their same-slice full-dose sources.Noise increases of 10% or more were detected in 92.5–100% of images, while blur of one pixel or more was detected in 99–100%.Source: Table 2 · p. 9, §4.3 · p. 9
4.2 A trade-off across the settings tested
- No tested setting combined strong dose ranking with low filtering preference: settings reaching at least 95% correctness in both regions preferred 43.6% or more of filtered images.The selected P = 24, C = 0.01, p = 0.5 model was fitted on 121,213 patches from 3,792 slices; the original-criterion setting ranked reduced-dose abdomen correctly in 28.5% [18.1, 40.9] of 214 scoreable test pairs.Source: Figure 2 · p. 8, §4.2 · p. 9
4.4 Preference for filtered full-dose images
- 41.2% [34.8, 48.3] of 7,200 filtered full-dose cases scored better than their unfiltered source, with preference concentrated on chest and edge-preserving filters.Only 27.8% were improvements larger than 0.05 score units; on abdomen, 21.4% of cases favored filtering, compared with 61.0% on chest.Source: §4.4 · pp. 9–10, Table 3 · p. 10
4.5 Preference and lesion signal
- At 32 HU residual, RIQE preferred bilateral filtering in all 18 lesion-bearing images although only 29% [25, 32] of the 4 mm lesion’s matched-filter signal remained.At that strength, bilateral filtering yielded d′ = 0.33 versus 0.48 after Gaussian filtering and 0.51 without filtering; Gaussian retained 70% of the signal and was never preferred.Source: §4.5 · pp. 10, 12, Table 4 · p. 11
4.6 One model or several
- Protocol-cell models raised validation dose-ranking accuracy for Siemens abdomen to 100% from 95.8% for the single model, but did not reduce filtering preference overall.The collection does not cross anatomy, kernel, and slice thickness, so contrasts cannot be attributed to one factor; per-protocol models narrowed filtering-preference spread across cells from 17–76% to 38–67%.Source: §4.6 · p. 12, Table 5 · p. 13
4.7 Baselines and radiologists’ scores
- With NIQE’s published photographic parameters, the same code ranked every simulated reduced-dose abdominal image better than its full-dose counterpart; RIQE’s radiologist-score correlation was exploratory.On LDCTIQAC, RIQE’s pooled correlation was ρ = −0.24 [−0.32, −0.17] and within inferred groups −0.47 [−0.54, −0.30]; the intervals depend on inferred grouping and do not account for within-patient dependence.Source: §4.7 · pp. 13–14, Table 6 · p. 13
5 Discussion
RIQE can rank a degraded image against its source, but its preferences among filtered images do not track lesion signal. The discussion attributes this mismatch to a reference model that can reward movement toward pooled statistics, and limits RIQE’s recommended use accordingly.
Ranking is not selection
- RIQE’s correct ordering of known degradations does not establish that it selects filtered images that preserve diagnostic signal.Average scores across filters can worsen monotonically with filtering strength even when the score does not follow lesion signal.Source: §5 · p. 14
A mechanism consistent with the data
- The pooled reference may act as a target rather than an ideal because protocols in its corpus differ fourfold in native noise.The authors describe this mechanism as consistent with the data, not demonstrated.Source: §5 · pp. 14–15
Why edge-preserving filters
- Edge-preserving filters may be preferred because active patches emphasize edges, while these filters flatten low-activity texture containing noise and small low-contrast lesions.The authors present this explanation as a hypothesis consistent with the data.Source: §5 · p. 15
Photographic references
- With NIQE’s published parameters, the photographic model in the authors’ code inverts on CT, while the original NIQE model’s behavior was not tested.A plausible interpretation is that added noise moves CT statistics closer to photographic statistics; a similar failure was reported for a NIQE variant on ultrasound.Source: §5 · p. 15
What RIQE is for
- RIQE is suited to paired ranking of an acquisition or reconstruction against a degraded or processed version of the same anatomy, including quality assurance or dose monitoring.It was not validated as a single-image detector, and deep-learning denoiser textures remain untested.Source: §5 · p. 15
Selection criteria
- Validation against dose reductions, rather than synthetic noise alone, exposed a setting that ranked simulated reduced-dose abdomen backwards on unseen patients.The setting had been selected using a criterion declared in advance.Source: §5 · p. 15
Scores are relative
- A RIQE score measures distance from its declared model, not absolute diagnostic quality, and is not comparable with NIQE values reported elsewhere.The model does not know what a lesion is.Source: §5 · p. 15
6 Limitations
The evaluation has important limits in its clinical validation, simulated noise, data coverage, and tested methods. Lesion experiments and all results are bounded by their test conditions and declared intensity mapping.
6 Limitations
- No reader study or diagnostic-performance correlation was performed, and the reduced-dose images were simulated rather than acquired.The simulated reduced-dose test pairs came from one vendor and 10 patients per region.Source: §6 · p. 15
- Only five classical filters and a finite settings grid were tested; deep-learning denoisers were not evaluated.Deep-learning denoisers are identified as the main use case.Source: §6 · p. 15
- The corpus covers two vendors, five protocol cells, and two anatomical regions from one collection; head CT was excluded.Kernel, slice thickness, and anatomy were not crossed, so some contrasts identify combinations of factors.Source: §6 · p. 15
- Lesions were inserted, mostly in muscle rather than liver, and assessed on 18 abdominal slices with a matched filter not calibrated to human readers.The relative ladders and lesion experiment use synthetic noise that approximates, rather than reproduces, the native noise power spectrum.Source: §6 · pp. 15–16
- All results depend on the declared intensity mapping, and motion and streak artefacts are not detected directly.Source: §6 · pp. 15–16
Data and code availability
The release makes the model, corpus manifest, and scripts available, with recorded metadata and a verification procedure for rebuilding the fit. The refit is bit-identical in the recorded environment, while other software versions or numerical backends may differ in their last digits.
Data and code availability
- The code, fitting and evaluation scripts, model, and corpus manifest are publicly available under MIT and CC BY 4.0 licences.The release used here is v1.0, archived on Zenodo; scripts regenerate experiment outputs, tables, and figures from public data.Source: Data and code availability · p. 16
- The model artefact records its intensity mapping, hyperparameters, selection criterion, slice identities and SHA-256 hashes, and code identity.A verification script refits the model from public data and compares it with the released arrays.Source: Data and code availability · p. 16
- The refit is bit-identical in the recorded software environment; other versions or numerical backends may differ in the last digits.The verification script reports maximum relative deviation against a tolerance of 10^-6.Source: Data and code availability · p. 16