Pith. sign in

REVIEW 4 major objections 5 minor 34 references

NAADA: A Noise-Aware Attention Denoising Autoencoder for Dental Panoramic Radiographs

T0 review · 4 major / 5 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 A denoising autoencoder whose self-attention is guided by an estimated local noise map outperforms Uformer on dental panoramic radiographs, with the largest gains at low signal-to-noise ratios.

desk verdict A plausible new noise-aware attention module, but the mirrored-image split likely leaks training identities into the test set, so the reported PSNR/SSIM win over Uformer is not yet believable. read the letter →

arxiv 2506.19387 v1 pith:CAVGPMG4 submitted 2025-06-24 eess.IV cs.AIcs.CVcs.LG

Khuram Naveed , Bruna Neves de Freitas , Ruben Pauwels This is my paper · ORCID
classification eess.IVcs.AIcs.CVcs.LG
keywords imagedenoisingattentionmechanismnoise-awareself-attentiondentalpanoramicradiographyautoencoderlow-doseX-raynoiseDENTEXdataset
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

This paper argues that convolutional denoising autoencoders fail on panoramic dental radiographs because they favor low-frequency structure and wash out the high-frequency anatomy that matters for diagnosis, and that standard self-attention makes this worse by ignoring features hidden in the noisiest regions. To fix this, it introduces NAADA, a denoising autoencoder whose bottleneck uses a noise-aware self-attention (NASA) module that estimates a local noise map and adds a noise-guided attention term to the usual query-key similarities. On a held-out test set of panoramic radiographs corrupted by a sequential Poisson, Gaussian, speckle, and salt-and-pepper noise model, NAADA reports higher PSNR and SSIM than Uformer, DnCNN, MResDNN, BM3D, and a plain attention DAE, with the largest gains at low signal-to-noise ratios. A blinded dental expert preferred NAADA's output in 90% of 50 cases, which matters because preserving subtle tooth and jaw structure is what makes denoised radiographs diagnostically usable.

What carries the argument

The central object is the noise-aware self-attention (NASA) module placed at the bottleneck of a five-layer convolutional encoder-decoder with skip connections. It estimates a local noise map $\psi$ via a $3\times3$ average-pooling RMS deviation from the local mean, builds a noise-query $Q(\psi)$ with a $1\times1$ convolution, and fuses attention as $$\mathrm{Softmax}\left(\frac{Q(Z)^\top K(Z)+\gamma' Q(\psi)^\top K(Z)}{\sqrt{M'}}\right)^\top V(Z),$$ where $\gamma'$ is a learnable weight. Adding this noise-guided attention term to the standard self-attention map is what shifts the model's focus toward features obscured by high noise rather than only toward cleaner regions.

What would settle it

Collect a set of real panoramic radiographs with paired low-dose (noisy) and routine-dose (cleaner) acquisitions from the same patients, train NAADA and Uformer on those pairs, and compare PSNR, SSIM, and a blinded multi-reader preference test. If NAADA no longer beats Uformer, or if the measured noise statistics of the real images differ sharply from the model in Section III-A, the central claim fails.

Watch

Extended reading notes

Core claim

The central discovery is that making self-attention noise-aware lets a denoising autoencoder recover fine anatomical details that convolutional and standard-attention denoisers lose. Rather than computing attention only from feature similarity to the clean image, NAADA estimates a pixel-wise local noise map, derives a second query from it, and adds the resulting noise-attention map to the standard self-attention map through a learned weight. This reweights the bottleneck features toward structures trapped inside noisy regions. Quantitatively, NAADA reaches 31.86 dB PSNR and 0.8146 SSIM on the test set versus 31.35 dB and 0.8097 for Uformer, and the clinical expert selected NAADA as the best-quality image in 45 of 50 blinded forced-choice comparisons.

Load-bearing premise

The load-bearing premise is that the hand-built sequential mixture of Poisson, Gaussian, speckle, and 5% salt-and-pepper noise in Section III-A matches real panoramic radiograph noise closely enough that a model trained on it will keep its advantage on real clinical images, and that the 925 chosen DENTEX images are clean enough to serve as ground truth.

Editorial extensions

If this is right

  • NAADA outperforms Uformer, the strongest baseline, by 0.51 dB PSNR and 0.0049 SSIM on the test set while being a much lighter model, offering a practical accuracy-versus-compute trade-off for clinical use.
  • Because NASA targets noisy regions rather than only clean correlations, the method's advantage concentrates at low input signal-to-noise ratios, where conventional denoisers blur fine detail or leave artifacts.
  • The blinded clinical result implies the improvement is not just a numerical artifact: a dentist preferred NAADA in 90% of 50 cases when judging anatomical depiction.
  • The architecture is not specific to panoramic radiographs; the paper claims it can extend to other medical imaging modalities with similar mixed-noise characteristics.

Reading between the lines

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

  • Inference beyond the paper: because the noise map responds to any sharp local deviation, NASA also highlights true edges and fine anatomy, so the same mechanism could serve as a saliency or edge-preservation prior in segmentation and lesion-detection pipelines, not only for denoising.
  • Inference beyond the paper: the reported gains are established under the paper's synthetic sequential noise model; transferring to real clinical noise would require re-testing, and the learned fusion weight $\gamma'$ might need recalibration if detector-specific noise differs.
  • Inference beyond the paper: a direct testable extension is to train on real paired low-dose and routine-dose panoramic images; if NAADA's margin over Uformer shrinks on real noise, the synthetic noise model, not the attention mechanism, would be the limiting factor.
  • Inference beyond the paper: the 90% clinical preference came from a single expert, so a multi-reader study with diagnostic tasks such as caries detection or bone-loss assessment would test whether the reported quality gain changes clinical decisions.
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 / 5 minor

Summary. The manuscript proposes NAADA, a denoising autoencoder for panoramic dental radiographs that augments the standard self-attention mechanism with a noise-aware attention path. The noise map is estimated as the local RMS deviation from the local mean in a 3x3 window, and the resulting noise query is fused with the standard attention scores through a learnable weight. The network is trained and evaluated on DENTEX panoramic radiographs that are degraded by a sequential synthetic noise model combining Poisson, Gaussian, speckle, and salt-and-pepper noise. The authors report that NAADA outperforms Uformer, MResDNN, DnCNN, and other baselines in PSNR and SSIM, and that a clinical expert preferred NAADA in 90% of 50 forced-choice comparisons.

Significance. If the reported results are valid, the paper offers a practical improvement for denoising low-dose panoramic dental radiographs, a task where preserving fine anatomical detail is clinically important. The work has several concrete strengths: it uses the public DENTEX dataset, retrains all baseline networks on the same training data, includes an ablation that contrasts NAADA with its non-noise-aware counterpart ADA, and attempts a blinded clinical preference assessment. The noise-aware attention idea is simple and could transfer to other medical imaging modalities. However, the current evidence is conditional on the integrity of the train/test split and on the adequacy of the synthetic noise model; both need to be verified before the central claim can be accepted.

major comments (4)
  1. [Section III-B] The described data split creates a serious risk of train/test leakage through mirrored images. The text states that the 925 selected DENTEX images were doubled by horizontal mirroring to 1850 images, and then a 70/15/15 split is described, but it never states that each mirror pair was kept in the same fold. If the random split is applied to the 1850-image set, roughly 70% of test images have a horizontally mirrored clean twin in the training set. Since a mirror image is pixel-identical after reflection, the network can recognize or memorize the training twin rather than learning to suppress noise. This would invalidate the generalization comparison in Table I and the 90% clinical preference reported in Section IV-C. Please either state explicitly that mirror pairs were never separated across folds, or repeat the experiments with a split performed on the 925 original image identities.
  2. [Section III-A, Eq. (8)] The speckle noise variance sigma_s^2 is never specified. Equation (8) defines multiplicative speckle noise as X''' + X''' * N_sigma_s, where N_sigma_s is zero-mean Gaussian noise of variance sigma_s^2, but the manuscript gives no value or sampling range for sigma_s^2. This is in contrast to the Gaussian variance range [0, 0.35] and the fixed 5% salt-and-pepper fraction. The missing parameter makes the noise generation under-specified and unreproducible, and it directly affects every quantitative result because the amount of speckle changes the difficulty of all test images.
  3. [Table I] The reported confidence intervals do not establish a statistically significant difference between NAADA and Uformer. Table I gives NAADA PSNR 31.86 ± 0.37 dB and SSIM 0.8146 ± 0.0090, while Uformer achieves 31.35 ± 0.35 dB and 0.8097 ± 0.0088; these intervals overlap substantially. The abstract and conclusion claim that NAADA outperforms the state of the art, but with overlapping 95% confidence intervals the difference may be due to chance. Please provide paired statistical tests on per-image differences, such as a Wilcoxon signed-rank test or bootstrap confidence intervals on the mean difference, to support the superiority claim.
  4. [Section IV-C and Section III-A] The external validity of the clinical claim is not established. All training, validation, test, and clinical-preference images are generated with the hand-specified synthetic noise model of Section III-A, and the clinical evaluation in Section IV-C is based on 50 synthetic images scored by a single expert. No experiment is performed on real noisy panoramic radiographs. As a result, the abstract's claim that NAADA ensures 'better image quality and diagnostic accuracy' and the conclusion's claim that the model 'generalizes well across varying noise levels' go beyond the evidence. The authors should either add an evaluation on real noisy radiographs or substantially temper these claims to denoising performance on synthetic noise.
minor comments (5)
  1. [Section III-B] The text says the noisy versions are prepared 'through the process explained above in Section III-B', but the noise model is described in Section III-A; the cross-reference should be corrected.
  2. [Fig. 1 and Section III-C.1] Figure 1 labels the bottleneck as 256 x 60 x 60, while Section III-C.1 states that the final encoder convolutional layer produces 1024 feature maps at 32 x 32 resolution. The figure and text should be reconciled.
  3. [Eq. (13)] The exact tensor shapes in Equation (13) are unclear: the transpose/reshape operator is applied to Q^T K and the softmax result is then transposed before multiplying V. Please clarify the dimensions of Q, K, V, and the noise query, and state whether the softmax is applied over the key dimension or the query dimension.
  4. [Section IV-C] The clinical evaluation reports only the preference rates for NAADA (90%) and ADA (10%); the authors should also report how often each baseline method, including Uformer, was selected by the expert, otherwise the reader cannot determine whether the baselines were ever competitive in the perceptual assessment.
  5. [General] The phrase 'This article is submitted to the IEEE Transactions on Medical Imaging for possible publication' and the arXiv version note are not part of the scientific content and should be removed or placed in the appropriate metadata in the final version.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: NAADA is an architectural modification evaluated on a shared synthetic benchmark, and its reported advantage is not forced by construction.

full rationale

The paper's derivation chain is an architecture proposal plus an empirical comparison. The noise-aware self-attention (NASA) computes a local RMS noise map (Eq. 11) and uses it to modulate standard self-attention scores through a learnable weight (Eq. 13); the network is trained with an MSE loss against the original clean DENTEX patches (Eq. 10). Nothing in this chain is defined in terms of the reported PSNR or SSIM values, and the claimed superiority over Uformer is an empirical outcome on a test set shared by all methods. The synthetic noise model of Section III-A is a simulation assumption rather than a fitted predictor: every method, including Uformer, is trained and evaluated on the same simulated degradation, so the comparison is internally consistent even though its external validity for real clinical noise remains an open question. The clinical preference test in Section IV-C is likewise conducted on images generated with the same noise model, which makes the real-clinical-relevance claim self-referential relative to that simulation, but it does not make the architecture's advantage equivalent to its inputs by construction. The horizontal-mirroring augmentation described in Section III-B could produce train/test leakage if the 70/15/15 split is not stratified by mirror pair, but the manuscript does not state the split procedure, so whether leakage occurred cannot be established from the text; this is a data-integrity concern, not a demonstrable circularity step. No load-bearing self-citations are present. Under the rule that circularity must be exhibited from the paper's own equations or citation chain, no circular step is found.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central evaluation depends on hand-chosen synthetic noise parameters and the assumption that high-quality DENTEX images are noise-free references. The NASA mechanism also assumes noise can be estimated by local RMS, which conflates edges with noise. No code, data list, or real noisy radiograph validation is provided.

free parameters (5)
  • Exposure constant c = 50
    Set in Eq. (3) to convert gray values to intensity and control quantum noise strength; the value is taken from prior literature but conditions all generated data.
  • Poisson scaling factor rho = few hundreds, exact value not stated
    Used in Eq. (4) to simulate low-exposure photon counts; the specific value is not reported, affecting the quantum noise level.
  • Gaussian noise variance sigma_g^2 = uniform in [0, 0.35]
    Selected for clinical realism via consultation in Section III-A, Eq. (7); this range defines the additive noise degradation.
  • Speckle noise variance sigma_s^2 = not specified
    Introduced in Eq. (8) but no value is given, so the exact speckle level is not reproducible.
  • Salt-and-pepper fraction = 5%
    Set in Section III-A, Eq. (9); determines the fraction of pixels corrupted by impulsive noise.
assumptions (4)
  • domain assumption The true noise in panoramic radiographs is well represented by the sequential combination of Poisson, Poisson, Gaussian, speckle, and salt-and-pepper noise with the stated parameter ranges.
    All training and evaluation data are generated under this model, so the method is only validated for this assumed noise distribution.
  • domain assumption The selected DENTEX images are clean, high-quality references sufficient to serve as ground truth.
    Any residual noise or artifacts in the 'clean' images are treated as the target, so the model learns to reproduce them.
  • ad hoc to paper Local RMS deviation from the local mean in a 3x3 window estimates the noise map needed for attention.
    This operational definition of noise cannot distinguish high-frequency texture or edges from actual noise, yet it is the basis of the NASA mechanism in Eq. (11).
  • domain assumption Higher PSNR, SSIM, and a single dentist's preference correspond to improved diagnostic accuracy.
    The abstract and conclusion claim diagnostic benefit, but the study measures image-quality metrics and one expert's visual preference, not diagnostic task performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NAADA: A Noise-Aware Attention Denoising Autoencoder for Dental Panoramic Radiographs." pith.science (2026). https://pith.science/paper/CAVGPMG4

@misc{pith2026250619387,
  author       = {Pith},
  title        = {Pith review of: NAADA: A Noise-Aware Attention Denoising Autoencoder for Dental Panoramic Radiographs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CAVGPMG4}},
  note         = {Machine review of arXiv:2506.19387}
}
read the original abstract

Convolutional denoising autoencoders (DAEs) are powerful tools for image restoration. However, they inherit a key limitation of convolutional neural networks (CNNs): they tend to recover low-frequency features, such as smooth regions, more effectively than high-frequency details. This leads to the loss of fine details, which is particularly problematic in dental radiographs where preserving subtle anatomical structures is crucial. While self-attention mechanisms can help mitigate this issue by emphasizing important features, conventional attention methods often prioritize features corresponding to cleaner regions and may overlook those obscured by noise. To address this limitation, we propose a noise-aware self-attention method, which allows the model to effectively focus on and recover key features even within noisy regions. Building on this approach, we introduce the noise-aware attention-enhanced denoising autoencoder (NAADA) network for enhancing noisy panoramic dental radiographs. Compared with the recent state of the art (and much heavier) methods like Uformer, MResDNN etc., our method improves the reconstruction of fine details, ensuring better image quality and diagnostic accuracy.

Figures

Figures reproduced from arXiv: 2506.19387 by the authors.

Figure 1
Figure 1. An illustration of the architecture of the proposed NAADA network that employs an encoder to extract core features of [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The proposed noise aware self attention (NASA) method that introduces a noise-attention map as a product of the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. An insight into the training and validation (through the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Example of the impact of the proposed noise-aware [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Denoising results of the best comparative methods on a test image at input PSNR of [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visual denoising results of the best comparative methods on an unseen noisy image at input PSNR of [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Visual denoising results of the best comparative methods on an unseen noisy image at input PSNR of [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 25 canonical work pages

  1. [1]

    Panoramic radiology,

    A. G. Farman, “Panoramic radiology,”Getting the most out of Panoramic Radiographic Interpretation. Berlin, Heidelberg: Springer-Verlag, pp. 1– 14, 2007

    work page 2007

  2. [2]

    Panoramic radiographs: an invaluable tool for the study of bone and teeth components in the maxillofacial region,

    R. Fuentes, A. Arias, and E. Borie-Echevarria, “Panoramic radiographs: an invaluable tool for the study of bone and teeth components in the maxillofacial region,”Int. J. Morphol, vol. 39, no. 1, pp. 268–73, 2021

    work page 2021

  3. [3]

    Noise estimation in panoramic x-ray images: An application analysis approach,

    P. M. Goebel, A. N. Belbachir, and M. Truppe, “Noise estimation in panoramic x-ray images: An application analysis approach,” inIEEE/SP 13th Workshop on Statistical Signal Processing, 2005. IEEE, 2005, pp. 996–1001

    work page 2005

  4. [4]

    Noise analysis of a digital radiography system,

    B. A. Arnold and P. Scheibe, “Noise analysis of a digital radiography system,”American journal of roentgenology, vol. 142, no. 3, pp. 609– 613, 1984

    work page 1984

  5. [5]

    Analysis of noise properties in dental images,

    V . Abramova, S. Krivenko, V . Lukin, and O. Krylova, “Analysis of noise properties in dental images,” in2020 IEEE 40th International Conference on Electronics and Nanotechnology (ELNANO). IEEE, 2020, pp. 511–515

    work page 2020

  6. [6]

    Quantitative analysis of metallic artifacts caused by dental metallic restorations: comparison between four cbct scanners,

    G. Omar, Z. Abdelsalam, and W. Hamed, “Quantitative analysis of metallic artifacts caused by dental metallic restorations: comparison between four cbct scanners,”Future Dental Journal, vol. 2, no. 1, pp. 15–21, 2016

    work page 2016

  7. [7]

    Positioning errors and quality assessment in panoramic radiography,

    M. Dhillon, S. M. Raju, S. Verma, D. Tomar, R. S. Mohan, M. Lakhan- pal, and B. Krishnamoorthy, “Positioning errors and quality assessment in panoramic radiography,”Imaging science in dentistry, vol. 42, no. 4, pp. 207–212, 2012

    work page 2012

  8. [8]

    Enhancing the quality of dental radiographic images: A review on panoramic and periapical radiograph enhancement techniques,

    O. E. E.-D. Abdulbadea Altukroni and S. Jabeen, “Enhancing the quality of dental radiographic images: A review on panoramic and periapical radiograph enhancement techniques,” inAdvances in Dentistry & Oral Health. Juniper Publishers, 1999, vol. 16, no. 4, pp. 1–12

    work page 1999

Show all 34 references
  1. [9]

    A novel method based on wiener filter for denoising poisson noise from medical x-ray images,

    V . G¨oreke, “A novel method based on wiener filter for denoising poisson noise from medical x-ray images,”Biomedical Signal Processing and Control, vol. 79, p. 104031, 2023

    2023

  2. [10]

    Application of panoramic dental x-ray images denoising,

    ˙I. C ¸ . Tas ¸, “Application of panoramic dental x-ray images denoising,” International Journal of Innovative Engineering Applications, vol. 7, no. 1, pp. 13–20, 2023

    2023

  3. [11]

    Image denoising by sparse 3-d transform-domain collaborative filtering,

    K. Dabov, A. Foi, V . Katkovnik, and K. Egiazarian, “Image denoising by sparse 3-d transform-domain collaborative filtering,”IEEE Transactions on image processing, vol. 16, no. 8, pp. 2080–2095, 2007

    2007

  4. [12]

    Bm3d outperforms major bench- marks in denoising: an argument in favor,

    B. Goyal, A. Dogra, and A. Sharma, “Bm3d outperforms major bench- marks in denoising: an argument in favor,”Journal of Computer Science, vol. 16, no. 6, pp. 838–847, 2020

    2020

  5. [13]

    Image denoising: Can plain neural networks compete with bm3d?

    H. C. Burger, C. J. Schuler, and S. Harmeling, “Image denoising: Can plain neural networks compete with bm3d?” in2012 IEEE conference on computer vision and pattern recognition. IEEE, 2012, pp. 2392–2399

    2012

  6. [14]

    C. Tian, L. Fei, W. Zheng, Y . Xu, W. Zuo, and C.-W. Lin, “Deep learning THIS ARTICLE IS SUBMITTED TO THE IEEE TRANSACTIONS ON MEDICAL IMAGING FOR POSSIBLE PUBLICATION 10 (a) Noisy at PSNR =18.40dB (b) MedDAE (c) BM3D (d) MResDNN (e) Uformer (f) Prop. NAADA Fig. 8: Visual deno...

    2020

  7. [15]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,

    K. Zhang, W. Zuo, Y . Chen, D. Meng, and L. Zhang, “Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,”IEEE transactions on image processing, vol. 26, no. 7, pp. 3142–3155, 2017

    2017

  8. [16]

    Deep residual learning- based denoiser for medical x-ray images,

    A. Mittal, N. Kaur, A. Gupta, and G. Singh, “Deep residual learning- based denoiser for medical x-ray images,”Evolving Systems, vol. 15, no. 6, pp. 2339–2353, 2024

    2024

  9. [17]

    Inception-residual block based neural network for thermal image denoising,

    S. Hwang, G. Yu, H. T. Nguyen, N. Shahid, D. Sin, J. Kim, and S. Na, “Inception-residual block based neural network for thermal image denoising,”arXiv preprint arXiv:1810.13169, 2018

    2018 arXiv

  10. [18]

    Hierar- chical residual learning for image denoising,

    W. Shi, F. Jiang, S. Zhang, R. Wang, D. Zhao, and H. Zhou, “Hierar- chical residual learning for image denoising,”Signal Processing: Image Communication, vol. 76, pp. 243–251, 2019

    2019

  11. [19]

    Auto-encoders in deep learning—a review with new perspectives,

    S. Chen and W. Guo, “Auto-encoders in deep learning—a review with new perspectives,”Mathematics, vol. 11, no. 8, p. 1777, 2023

    2023

  12. [20]

    Attentive contextual denoising autoencoder for recommendation,

    Y . Jhamb, T. Ebesu, and Y . Fang, “Attentive contextual denoising autoencoder for recommendation,” inProceedings of the 2018 ACM SIGIR international conference on theory of information retrieval, 2018, pp. 27–34

    2018

  13. [21]

    Attention-based convolutional denoising autoencoder for two-lead ecg denoising and arrhythmia classification,

    P. Singh and A. Sharma, “Attention-based convolutional denoising autoencoder for two-lead ecg denoising and arrhythmia classification,” IEEE Transactions on Instrumentation and Measurement, vol. 71, pp. 1–10, 2022

    2022

  14. [22]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,”Advances in Neural Information Processing Systems, 2017

    2017

  15. [23]

    Self-attention generative adversarial networks,

    H. Zhang, I. Goodfellow, D. Metaxas, and A. Odena, “Self-attention generative adversarial networks,” inInternational conference on machine learning. PMLR, 2019, pp. 7354–7363

    2019

  16. [24]

    Attention mechanisms in medical image segmentation: A survey,

    Y . Xie, B. Yang, Q. Guan, J. Zhang, Q. Wu, and Y . Xia, “Attention mechanisms in medical image segmentation: A survey,”arXiv preprint arXiv:2305.17937, 2023

    2023 arXiv

  17. [25]

    A review on the attention mechanism of deep learning,

    Z. Niu, G. Zhong, and H. Yu, “A review on the attention mechanism of deep learning,”Neurocomputing, vol. 452, pp. 48–62, 2021

    2021

  18. [26]

    Extract- ing and composing robust features with denoising autoencoders,

    P. Vincent, H. Larochelle, Y . Bengio, and P.-A. Manzagol, “Extract- ing and composing robust features with denoising autoencoders,” in Proceedings of the 25th international conference on Machine learning, 2008, pp. 1096–1103

    2008

  19. [27]

    Daema: Denoising autoencoder with mask attention,

    S. Tihon, M. U. Javaid, D. Fourure, N. Posocco, and T. Peel, “Daema: Denoising autoencoder with mask attention,” inInternational Confer- ence on Artificial Neural Networks. Springer, 2021, pp. 229–240

    2021

  20. [28]

    Poisson–gaussian noise analysis and estimation for low-dose x-ray images in the nsct domain,

    S. Lee, M. S. Lee, and M. G. Kang, “Poisson–gaussian noise analysis and estimation for low-dose x-ray images in the nsct domain,”Sensors, vol. 18, no. 4, p. 1019, 2018

    2018

  21. [29]

    Modeling quantum noise of phosphors used in medical x-ray imaging detectors,

    N. Kalivas, I. Kandarakis, D. Cavouras, L. Costaridou, C. Nomicos, and G. Panayiotakis, “Modeling quantum noise of phosphors used in medical x-ray imaging detectors,”Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Assoc...

    1999

  22. [30]

    Amorphous and polycrystalline photoconductors for direct conversion flat panel x- ray image sensors,

    S. Kasap, J. B. Frey, G. Belev, O. Tousignant, H. Mani, J. Greenspan, L. Laperriere, O. Bubon, A. Reznik, G. DeCrescenzoet al., “Amorphous and polycrystalline photoconductors for direct conversion flat panel x- ray image sensors,”Sensors, vol. 11, pp. 5112–5157, 2011

    2011

  23. [31]

    Dentex: An abnormal tooth detection with dental enumeration and diagnosis benchmark for panoramic x-rays,

    I. E. Hamamci, S. Er, E. Simsar, A. E. Yuksel, S. Gultekin, S. D. Ozdemir, K. Yang, H. B. Li, S. Pati, B. Stadlingeret al., “Dentex: An abnormal tooth detection with dental enumeration and diagnosis benchmark for panoramic x-rays,”arXiv preprint arXiv:2305.19112, 2023

    2023

  24. [32]

    Adam: A method for stochastic optimization,

    D. P. Kingma, “Adam: A method for stochastic optimization,”arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  25. [33]

    Medical image denoising using convolutional denoising autoencoders,

    L. Gondara, “Medical image denoising using convolutional denoising autoencoders,” in2016 IEEE 16th international conference on data mining workshops (ICDMW). IEEE, 2016, pp. 241–246

    2016

  26. [34]

    Uformer: A general u-shaped transformer for image restoration,

    Z. Wang, X. Cun, J. Bao, W. Zhou, J. Liu, and H. Li, “Uformer: A general u-shaped transformer for image restoration,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 17 683–17 693

    2022

Pith tools

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