Pith. sign in

REVIEW 4 major objections 6 minor 34 references

MS-IQA: A Multi-Scale Feature Fusion Network for PET/CT Image Quality Assessment

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MS-IQA, a dual-branch fusion of ResNet and Swin Transformer features, reports the highest radiologist-correlated PET/CT quality scores on two benchmarks.

desk verdict A useful first PET/CT IQA dataset and a plausible fusion network, but the external-dataset comparison is under-specified and the margins are small. read the letter →

arxiv 2506.20200 v1 pith:EPRCCEV2 submitted 2025-06-25 eess.IV cs.CV

Siqiao Li , Chen Hui , Wei Zhang , Rui Liang , Chenyue Song , Feng Jiang , Haiqi Zhu , Zhixuan Li
show 2 more authors
Hong Huang Xiang Li
This is my paper · ORCID
classification eess.IVcs.CV
keywords PET/CTimagequalityassessmentno-referencemulti-scalefeaturefusionSwinTransformerResNetchannelattentionmedicalimagingdatasetradiologistscores
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a no-reference image quality assessment model for PET/CT should combine low-level distortion cues with high-level anatomical information, and that a multi-scale fusion network can do so better than existing quality assessment methods. This matters because PET/CT image quality directly affects diagnostic confidence, and clinical scans have no reference image to compare against. The authors construct the first PET/CT IQA dataset, PET-CT-IQA-DS, with 2,700 images scored by radiologists, and report that their MS-IQA network outperforms current state-of-the-art methods on that dataset and on the public LDCTIQAC2023 challenge dataset.

What carries the argument

The load-bearing mechanism is a multi-scale feature fusion architecture. An input PET/CT image is passed in parallel through ResNet-50 and Swin Transformer Tiny, and the outputs of four intermediate stages from each backbone are concatenated stage by stage. Each concatenated feature group is refined by an Adaptive Graph Channel Attention (AGCA) block, which treats channels as graph vertices and computes an attention matrix from self-channel attention plus an inter-channel adjacency matrix. The refined features feed two branches, a score branch and a weight branch whose product forms the final prediction, trained with a combined mean-squared-error and pairwise ranking loss. This lets early stages contribute edge and noise detail while later stages contribute organ-level semantics.

What would settle it

Collect a test set of real clinical PET/CT images spanning routine artifacts such as motion, attenuation-correction mismatch, and dose variation, with independent radiologist scores; if MS-IQA's Spearman correlation on this set falls substantially below the 0.9559 it reports on synthetic distortions, the dataset premise fails.

Watch

Extended reading notes

Core claim

The central claim is that PET/CT image quality is jointly determined by low-level artifacts and high-level anatomical visibility, and that a network which fuses features from multiple depths of both a convolutional and a transformer backbone captures this better than models using either alone. On PET-CT-IQA-DS, MS-IQA achieves a Spearman correlation of 0.9559 and a Pearson correlation of 0.9701 with radiologist scores, beating all compared no-reference methods. On LDCTIQAC2023, it achieves 0.9542 and 0.9543, surpassing every participating challenge team. The paper also introduces PET-CT-IQA-DS, the first PET/CT-specific IQA dataset, built from 20 patients with mixed Poisson, Gaussian, and JPEG distortions scored on a 0-4 Likert scale by five radiologists.

Load-bearing premise

The load-bearing premise is that images degraded by synthetic Poisson, Gaussian, and JPEG distortions and scored by five radiologists accurately represent the image-quality failures that matter in real PET/CT scans.

Editorial extensions

If this is right

  • If the reported results hold, no-reference PET/CT quality assessment can match or exceed existing methods on radiologist-labeled data without needing a reference scan.
  • The same architecture transfers to low-dose CT, beating all LDCTIQAC2023 challenge participants, which suggests the multi-scale fusion strategy generalizes across related imaging modalities.
  • Every one of the four ResNet and Swin stages contributes to performance, and removing the middle stage costs the most, indicating that both low-level and high-level features are necessary for quality prediction.
  • The first dedicated PET/CT IQA dataset provides a benchmark on which future automated quality assessment methods can be compared against human scoring.
  • The dual-branch score-weight design and ranking loss act as a regularization mechanism that the paper argues improves generalization beyond a plain regression network.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct extension would be to test MS-IQA on real clinical PET/CT artifacts such as patient motion, attenuation-correction mismatch, and reconstruction-parameter variation; the synthetic distortions in the dataset are not shown to reproduce these.
  • The paper averages five radiologists' scores without reporting inter-rater variability, so treating the full score distribution rather than the mean could reveal whether the model learns genuine consensus or merely averages away disagreement.
  • The feature visualizations suggest that intermediate representations separate noise from anatomy, which could be turned into a quality heatmap showing clinicians which image regions degrade diagnostic value.
  • Because the architecture makes no PET-specific assumptions, applying it to PET/MR or SPECT/CT is a direct way to test whether multi-scale fusion, rather than dataset-specific tuning, is the source of the reported gain.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes MS-IQA, a no-reference image quality assessment network for PET/CT that combines ResNet-50 and Swin Transformer features from four intermediate stages, fuses them through a multi-scale feature fusion module with adaptive graph channel attention, and predicts a quality score via separate score and weight branches. The authors also introduce PET-CT-IQA-DS, a dataset of 2,700 PET/CT images generated by applying Poisson noise, Gaussian noise, and JPEG compression at three levels each to images from 20 patients, with quality labels obtained by averaging ratings from five radiologists. Experiments report SROCC and PLCC improvements over existing NR-IQA methods on PET-CT-IQA-DS and on the public LDCTIQAC2023 challenge dataset, together with ablations and feature-map visualizations.

Significance. If the reported results hold, the paper makes two useful contributions: it provides the first dedicated PET/CT IQA dataset, and it demonstrates that a dual-branch multi-scale architecture can achieve strong correlation with radiologist quality ratings. The authors should be credited for releasing code and data, for using a patient-level train/test split that avoids direct image leakage, and for evaluating on an independent public dataset in addition to their own. The central claim, however, is an empirical superiority claim, and its strength depends on whether the external comparison is apples-to-apples and whether the reported margins are statistically meaningful; both aspects are currently under-supported in the manuscript.

major comments (4)
  1. [§3.1, Table 1] The paper never states the training protocol used for LDCTIQAC2023. It reports that the challenge has a 1000-image training set and a 300-image test set, and Table 1 compares MS-IQA with challenge teams whose results are taken from [19], but it does not say whether MS-IQA was trained on the official 1000-image training split, whether it was fine-tuned from PET-CT-IQA-DS, or whether any model selection used the test set. The claim that MS-IQA 'outperforms all participating teams' is valid only under the first protocol; under other protocols the comparison is not apples-to-apples. This should be stated explicitly in §3.1 or in a protocol paragraph accompanying Table 1.
  2. [Table 1] No uncertainty quantification is reported. The margins over the second-best method on LDCTIQAC2023 are small (SROCC +0.0047, PLCC +0.0052 over agaldran), and on a 300-image test set the sampling error in SROCC is plausibly on the order of 0.01. Without repeated-seed experiments, bootstrap confidence intervals, or a significance test, the reported advantage over the runner-up is not established. The authors should add such statistics for both datasets, and should likewise report variance for the PET-CT-IQA-DS results, where the comparison set is also small.
  3. [§3.1] The test portion of PET-CT-IQA-DS is drawn from only 4 patients (20% of the 20 patients, with 5 source images per patient), so despite containing 540 distorted test images, it represents very limited anatomical diversity. In addition, all distortions are synthetic combinations of Poisson noise, Gaussian noise, and JPEG compression; the paper provides no evidence that these distortions match clinically occurring PET/CT artifacts, and it does not report inter-rater variability among the five radiologists whose averaged scores are used as ground truth. The claim that MS-IQA is an accurate PET/CT IQA method for clinical use therefore needs either additional validation on real distorted scans or a clearly stated scope limitation.
  4. [§2.3, Eq. (6)] The AGCA mechanism, which is a central component of the proposed MSFFM, is underspecified. Equation (6) introduces A2 as an adjacency matrix between channels, but the manuscript never defines how A2 is computed, learned, or initialized, and Eq. (5) does not make clear what object f is when it is 'transposed and replicated' to form A1. Without a precise definition of A2 and the dimensions of the matrices, the module is not reproducible from the description. The authors should provide the exact construction of A2 and the shapes of A0, A1, and A2.
minor comments (6)
  1. [§2.4, Eq. (7)] Equation (7) has unbalanced parentheses; the expression for S should be corrected, and the LeakyReLU/linear notation made consistent with Eq. (8).
  2. [References] References [9] and [10] are identical (Bianco et al., 'On the use of deep learning for blind image quality assessment'); the duplicate should be removed or replaced with the intended citation.
  3. [§2.3, Eq. (5)] The notation in Eq. (5) is confusing: it is unclear why f is repeated C times and how this yields a C×C matrix, and the role of the identity matrix A0 in Eq. (6) should be explained more carefully.
  4. [Abstract and §3.2] The phrase 'various IQA metrics' is not accurate because only SROCC and PLCC are reported; the text should say 'two IQA metrics' or add additional metrics such as Kendall's tau or RMSE.
  5. [§3.3, Fig. 2] The visualization section describes feature maps qualitatively but does not connect them to the quantitative improvements; a brief discussion of what the visualized features reveal about the fusion mechanism would strengthen the presentation.
  6. [§2.5] The ranking loss in Eq. (11) sums over all pairs; the complexity O(n^2) is worth mentioning, and the choice of α = 2.0 and λ1 = λ2 = 0.5 should be justified or shown to be insensitive in an ablation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: MS-IQA is an empirical supervised-learning comparison with external benchmarking, and no prediction reduces by construction to its inputs.

full rationale

The paper's central claims are empirical performance numbers on PET-CT-IQA-DS and the public LDCTIQAC2023 challenge dataset. No derived quantity is defined in terms of the target it is said to predict: the model outputs quality scores from multi-scale features, and these are compared against radiologist-provided ground truth using SROCC and PLCC. The PET-CT-IQA-DS dataset is constructed by the authors and also serves as their main evaluation set, with a patient-level train/test split; this is standard supervised-learning practice and does not make the reported correlation a construction artifact. The only other benchmark, LDCTIQAC2023, is an external dataset with published challenge results taken from reference [19], providing independent grounding. Although the paper does not state the exact training protocol on LDCTIQAC2023 and reports no significance intervals, that is a completeness/rigor concern about the comparison, not a circularity of the kind where a fitted parameter is renamed a prediction or where a claimed derivation is equivalent to its input by construction. No load-bearing step relies on a self-citation chain: the cited AGCA module is from other authors, and the ResNet/Swin backbones are standard pretrained architectures. The loss function (Eq. 10-12) is a standard MSE plus ranking loss; it does not encode the evaluation metric as a fitted target. Therefore, the derivation is self-contained as an empirical architecture study, and no specific circular reduction can be exhibited.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the validity of the radiologist-labeled synthetic dataset, the transferability of ImageNet-pretrained features, and the correct functioning of the borrowed AGCA module. The only hand-set numeric parameters in the loss are alpha (2.0) and lambda1/lambda2 (0.5). No invented entities are introduced.

free parameters (3)
  • alpha (ranking loss scaling factor) = 2.0
    Set by hand in Section 2.5 to control the scaling factor of the ranking loss; no sensitivity analysis is provided, so the reported scores could depend on this choice.
  • lambda1 (MSE loss weight) = 0.5
    Set by hand in Section 2.5 to balance the MSE loss; the paper does not explore other values.
  • lambda2 (ranking loss weight) = 0.5
    Set by hand in Section 2.5 to balance the ranking loss; the paper does not explore other values.
assumptions (4)
  • domain assumption Pretrained ImageNet weights for ResNet-50 and Swin-T provide useful features for medical IQA.
    The paper uses standard backbones without stating whether they are pretrained or trained from scratch; transfer from natural images is assumed to transfer to PET/CT, which the paper does not test.
  • domain assumption The average of five radiologist scores is a reliable ground truth for PET/CT image quality.
    Section 3.1 states the average score of 5 radiologists is the final label, but no inter-rater agreement or variability measure is reported.
  • domain assumption Synthetic mixed distortions (Poisson, Gaussian, JPEG) capture the clinically relevant quality degradation of PET/CT images.
    Section 3.1 constructs the dataset by adding these synthetic distortions; no validation on real clinical artifacts is provided.
  • standard math The AGCA attention mechanism from [22] functions as described in the original paper.
    The paper adopts AGCA as a component and does not re-derive it; the explanation in Section 2.3 contains an inconsistency, so correctness of the borrowed module is taken on faith.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MS-IQA: A Multi-Scale Feature Fusion Network for PET/CT Image Quality Assessment." pith.science (2026). https://pith.science/paper/EPRCCEV2

@misc{pith2026250620200,
  author       = {Pith},
  title        = {Pith review of: MS-IQA: A Multi-Scale Feature Fusion Network for PET/CT Image Quality Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EPRCCEV2}},
  note         = {Machine review of arXiv:2506.20200}
}
read the original abstract

Positron Emission Tomography / Computed Tomography (PET/CT) plays a critical role in medical imaging, combining functional and anatomical information to aid in accurate diagnosis. However, image quality degradation due to noise, compression and other factors could potentially lead to diagnostic uncertainty and increase the risk of misdiagnosis. When evaluating the quality of a PET/CT image, both low-level features like distortions and high-level features like organ anatomical structures affect the diagnostic value of the image. However, existing medical image quality assessment (IQA) methods are unable to account for both feature types simultaneously. In this work, we propose MS-IQA, a novel multi-scale feature fusion network for PET/CT IQA, which utilizes multi-scale features from various intermediate layers of ResNet and Swin Transformer, enhancing its ability of perceiving both local and global information. In addition, a multi-scale feature fusion module is also introduced to effectively combine high-level and low-level information through a dynamically weighted channel attention mechanism. Finally, to fill the blank of PET/CT IQA dataset, we construct PET-CT-IQA-DS, a dataset containing 2,700 varying-quality PET/CT images with quality scores assigned by radiologists. Experiments on our dataset and the publicly available LDCTIQAC2023 dataset demonstrate that our proposed model has achieved superior performance against existing state-of-the-art methods in various IQA metrics. This work provides an accurate and efficient IQA method for PET/CT. Our code and dataset are available at https://github.com/MS-IQA/MS-IQA/.

Figures

Figures reproduced from arXiv: 2506.20200 by the authors.

Figure 1
Figure 1. The architecture of our proposed model: MS-IQA. It consists of 4 modules: [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. For fin, the input of AGCA, a 1 × 1 convolution layer, followed by a sigmoid operation are applied, resulting in f as shown in Eq. 4. f = σ(Conv(fin)) (4) The attention calculation involves three matrices: A0, A1, and A2. First, f is transposed and replicated C times to expand it into a C × C shape, resulting in A1, as shown in Eq. 5. A1 = [f T f T · · · f T ] | {z } C times (5) [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 2
Figure 2. Visualization Samples the weights of two losses and we set λ1=λ2=0.5 in this work. The loss function simultaneously optimizes the model’s ranking ability and regression ability. LMSE = 1 n Xn i=1 (yi − ti) 2 (10) LRanking = 2 n(n − 1) Xn i=1 Xn j=i+1 (σ (α(ti − tj )) − σ (α(yi − yj )))2 (11) LTotal = λ1LMSE + λ2LRanking (12) 3 Experiments 3.1 Dataset and Experiment Details To the best of our knowledge, we are the fi… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

  1. [19]

    Low-dose computed tomography perceptual image quality assessment

    Lee, Wonkyeong, et al. "Low-dose computed tomography perceptual image quality assessment." Medical Image Analysis 99 (2025): 103343

    work page 2025

  2. [1]

    PET/CT imaging artifacts

    Sureshbabu, Waheeda, and Osama Mawlawi. "PET/CT imaging artifacts." Journal of nuclear medicine technology 33.3 (2005): 156-161

    work page 2005

  3. [2]

    An artificial intelligence-driven image quality assessment system for whole-body [18F] FDG PET/CT

    Qi, Chi, et al. "An artificial intelligence-driven image quality assessment system for whole-body [18F] FDG PET/CT." European Journal of Nuclear Medicine and Molecular Imaging 50.5 (2023): 1318-1328

    work page 2023

  4. [3]

    Modern image quality assessment

    Wang, Zhou, and Alan Conrad Bovik. Modern image quality assessment. Diss. Mor- gan & Claypool Publishers, 2006

    work page 2006

  5. [4]

    Review of medical image quality as- sessment

    Chow, Li Sze, and Raveendran Paramesran. "Review of medical image quality as- sessment." Biomedical signal processing and control 27 (2016): 145-154

    work page 2016

  6. [5]

    Making a “completely blind

    Mittal, Anish, Rajiv Soundararajan, and Alan C. Bovik. "Making a “completely blind” image quality analyzer." IEEE Signal processing letters 20.3 (2012): 209-212

    work page 2012

  7. [6]

    No-reference im- age quality assessment in the spatial domain

    Mittal, Anish, Anush Krishna Moorthy, and Alan Conrad Bovik. "No-reference im- age quality assessment in the spatial domain." IEEE Transactions on image pro- cessing 21.12 (2012): 4695-4708

    work page 2012

  8. [7]

    C-DIIVINE: No-reference image quality assessment based on local magnitude and phase statistics of natural scenes

    Zhang, Yi, et al. "C-DIIVINE: No-reference image quality assessment based on local magnitude and phase statistics of natural scenes." Signal processing: image communication 29.7 (2014): 725-747

    work page 2014

Show all 34 references
  1. [8]

    Deep neural networks for no-reference and full-reference image quality assessment

    Bosse, Sebastian, et al. "Deep neural networks for no-reference and full-reference image quality assessment." IEEE Transactions on image processing 27.1 (2017): 206-219

    2017

  2. [9]

    On the use of deep learning for blind image quality assess- ment

    Bianco, Simone, et al. "On the use of deep learning for blind image quality assess- ment." Signal, Image and Video Processing 12 (2018): 355-362. 10 Anonymized Author et al

    2018

  3. [10]

    On the use of deep learning for blind image quality assess- ment

    Bianco, Simone, et al. "On the use of deep learning for blind image quality assess- ment." Signal, Image and Video Processing 12 (2018): 355-362

    2018

  4. [11]

    SSIQA: Multi-task learning for non-reference CT image quality assessment with self-supervised noise level predic- tion

    Pal, Debashish, Bhavik Patel, and Adam Wang. "SSIQA: Multi-task learning for non-reference CT image quality assessment with self-supervised noise level predic- tion." 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). IEEE, 2021

    2021

  5. [12]

    Performance evaluation of image quality metrics for per- ceptual assessment of low-dose computed tomography images

    Lee, Wonkyeong, et al. "Performance evaluation of image quality metrics for per- ceptual assessment of low-dose computed tomography images." Medical Imaging 2022: Image Perception, Observer Performance, and Technology Assessment. Vol. 12035. SPIE, 2022

    2022

  6. [14]

    Gradient-based learning applied to document recognition

    LeCun, Yann, et al. "Gradient-based learning applied to document recognition." Proceedings of the IEEE 86.11 (1998): 2278-2324

    1998

  7. [15]

    Attention is all you need

    Vaswani, Ashish, et al. "Attention is all you need." Advances in neural information processing systems 30 (2017)

    2017

  8. [16]

    No-reference image quality assessment via transformers, relative ranking, and self-consistency

    Golestaneh, S. Alireza, Saba Dadsetan, and Kris M. Kitani. "No-reference image quality assessment via transformers, relative ranking, and self-consistency." Pro- ceedings of the IEEE/CVF winter conference on applications of computer vision. 2022

    2022

  9. [17]

    Deep learning in medical image analysis

    Shen, Dinggang, Guorong Wu, and Heung-Il Suk. "Deep learning in medical image analysis." Annual review of biomedical engineering 19.1 (2017): 221-248

    2017

  10. [18]

    Deep learning-based feature representation for AD/MCI classification

    Suk, Heung-Il, and Dinggang Shen. "Deep learning-based feature representation for AD/MCI classification." Medical Image Computing and Computer-Assisted Inter- vention–MICCAI 2013: 16th International Conference, Nagoya, Japan, September 22-26, 2013, Proceedings, Part II 16. Spr...

    2013

  11. [20]

    Deep residual learning for image recognition

    He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016

    2016

  12. [21]

    Swin transformer: Hierarchical vision transformer using shifted win- dows

    Liu, Ze, et al. "Swin transformer: Hierarchical vision transformer using shifted win- dows." Proceedings of the IEEE/CVF international conference on computer vision. 2021

    2021

  13. [22]

    AGCA: An adaptive graph channel attention module for steel surface defect detection

    Xiang, Xin, et al. "AGCA: An adaptive graph channel attention module for steel surface defect detection." IEEE Transactions on Instrumentation and Measurement 72 (2023): 1-12

    2023

  14. [23]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, Alexey, et al. "An image is worth 16x16 words: Transformers for image recognition at scale." arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  15. [24]

    An empirical comparison of voting classification algorithms: Bagging, boosting, and variants

    Bauer, Eric, and Ron Kohavi. "An empirical comparison of voting classification algorithms: Bagging, boosting, and variants." Machine learning 36 (1999): 105-139

    1999

  16. [25]

    Li, P., Wang, S., Li, T., Lu, J., HuangFu, Y., & Wang, D. (2020). A Large-Scale CT and PET/CT Dataset for Lung Cancer Diagnosis (Lung-PET-CT-Dx) [Data set]

    2020

  17. [26]

    Statistical image reconstruction for polyenergetic X-ray computed tomography

    Elbakri, Idris A., and Jeffrey A. Fessler. "Statistical image reconstruction for polyenergetic X-ray computed tomography." IEEE transactions on medical imaging 21.2 (2002): 89-99

    2002

  18. [27]

    Blindlyassessimagequalityinthewildguidedbyaself-adaptive hyper network

    Su,Shaolin,etal."Blindlyassessimagequalityinthewildguidedbyaself-adaptive hyper network." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020

    2020

  19. [28]

    Musiq: Multi-scale image quality transformer

    Ke, Junjie, et al. "Musiq: Multi-scale image quality transformer." Proceedings of the IEEE/CVF international conference on computer vision. 2021. Title Suppressed Due to Excessive Length 11

    2021

  20. [29]

    Image quality assessment using contrastive learn- ing

    Madhusudana, Pavan C., et al. "Image quality assessment using contrastive learn- ing." IEEE Transactions on Image Processing 31 (2022): 4149-4161

    2022

  21. [30]

    Maniqa: Multi-dimension attention network for no-reference image quality assessment

    Yang, Sidi, et al. "Maniqa: Multi-dimension attention network for no-reference image quality assessment." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022

    2022

  22. [31]

    Topiq: A top-down approach from semantics to distortions for image quality assessment

    Chen, Chaofeng, et al. "Topiq: A top-down approach from semantics to distortions for image quality assessment." IEEE Transactions on Image Processing (2024)

    2024

  23. [32]

    Re-iqa: Unsupervised learning for image quality assessment in the wild

    Saha, Avinab, Sandeep Mishra, and Alan C. Bovik. "Re-iqa: Unsupervised learning for image quality assessment in the wild." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2023

    2023

  24. [33]

    Arniqa: Learning distortion manifold for image quality assessment

    Agnolucci, Lorenzo, et al. "Arniqa: Learning distortion manifold for image quality assessment." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024

    2024

  25. [34]

    Local distortion aware efficient transformer adaptation for image quality assessment

    Xu, Kangmin, et al. "Local distortion aware efficient transformer adaptation for image quality assessment." arXiv preprint arXiv:2308.12001 (2023)

    2023 arXiv

  26. [35]

    Blind image quality assess- ment based on geometric order learning

    Shin, Nyeong-Ho, Seon-Ho Lee, and Chang-Su Kim. "Blind image quality assess- ment based on geometric order learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.