Recovering Diagnostic Value: Super-Resolution-Aided Echocardiographic Classification in Resource-Constrained Imaging

Annu; Krishan Agyakari Raja Babu; Mohanasankar Sivaprakasam; Om Prabhu

REVIEW 3 major objections 5 minor 24 references

Recovering Diagnostic Value: Super-Resolution-Aided Echocardiographic Classification in Resource-Constrained Imaging

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Super-resolution preprocessing, especially the pixel-accurate SRResNet, measurably recovers classification accuracy lost on poor-quality echocardiograms and can help at test time without retraining.

desk verdict Plausible result, but the experimental evidence is undermined by a likely patient-level split leak and missing error bars; fixable and worth reviewing. read the letter →

arxiv 2507.23027 v1 pith:Y3EDOUI3 submitted 2025-07-30 cs.CV cs.AI

Krishan Agyakari Raja Babu , Om Prabhu , Annu , Mohanasankar Sivaprakasam This is my paper · ORCID

classification cs.CVcs.AI

keywords super-resolutionechocardiographyimageenhancementcardiacclassificationresource-constrainedsettingsultrasounddeeplearning

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 claims that a lightweight super-resolution step, applied to poor-quality 2D echocardiograms, recovers a meaningful share of the classification accuracy that image degradation destroys. Using the CAMUS dataset's expert quality labels, the authors isolate poor frames as a proxy for resource-constrained imaging and test two SR models on view (2CH vs 4CH) and phase (ED vs ES) tasks. They report that SRResNet, an MSE-trained residual network, consistently improves accuracy more than the adversarial SRGAN while costing less compute, and that gains appear even when enhancement is applied only at test time. The intended upshot is that SR can serve as a cheap, model-agnostic preprocessing layer for AI-assisted cardiac diagnosis in low-resource settings.

What carries the argument

The pipeline pairs a quality-stratified classifier with a super-resolution front-end. Expert quality labels split the CAMUS frames into good, medium, and poor subsets, with the 232 poor frames standing in for resource-constrained acquisition. Two SR architectures, SRResNet and SRGAN, are fine-tuned on synthetic pairs formed by 4x bicubic downsampling of the 1,010 good-quality frames, then applied to poor frames to create enhanced inputs. A ResNet-18 classifier trained on each quality subset evaluates the original and enhanced images. The load-bearing mechanism is reconstruction fidelity: SRResNet's MSE-optimized weights preserve anatomical structure, and that fidelity transfers into classification gains, while the sharper-looking but less faithful SRGAN output does not transfer as well.

What would settle it

Take a set of echocardiograms acquired with a low-end or handheld device in a clinic, or simulate speckle noise and realistic ultrasound artifacts, run the same fine-tuned SRResNet and SRGAN preprocessing, and measure view and phase classification accuracy against the original poor frames; if the enhanced frames fail to improve accuracy or only help on bicubic-synthesized degradations, the claim that SR recovers diagnostic value in resource-constrained imaging collapses.

Watch

Extended reading notes

Core claim

The central discovery is that pixel-accurate super-resolution, rather than perceptual sharpness, is what restores diagnostic value in degraded echocardiography. When SR-enhanced poor frames replace or supplement original poor frames, view classification accuracy improves by an average of 7.83% with SRResNet versus 6.6% with SRGAN, and phase classification also improves; in the train-on-SR-enhanced, test-on-good setting, gains reach 14.9% for view and 11.5% for phase. The authors also show that SR applied only at inference time, without retraining the classifier, yields a 12.2% average view-classification improvement for models trained on good-quality data. SRResNet's higher pixel-level fidelity (PSNR 38.98 dB, SSIM 0.9214) tracks its better downstream performance, suggesting that the recovery mechanism is structural reconstruction rather than adversarial realism.

Load-bearing premise

The SR models are trained on pairs made by 4x bicubic downsampling of good frames, and the paper assumes this synthetic corruption stands in for real poor-quality echo acquisition (speckle noise, low contrast, operator artifacts) when it applies the models to native poor frames.

Editorial extensions

If this is right

If SRResNet consistently outperforms SRGAN while using less compute, low-resource deployments should prefer pixel-accurate SR architectures over adversarial ones.
Super-resolution can work as a test-time preprocessing layer for classifiers already trained on good or medium images, since measurable gains appear without retraining.
Using SR-enhanced poor images as training data improves cross-quality generalization, shrinking the accuracy gap between poor and good test domains.
Because medium-quality images already perform robustly across test conditions, the practical bottleneck in resource-constrained settings may be limited to the worst-quality frames, which SR can partially repair.

Reading between the lines

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

The paper's evaluation is limited to view and phase classification; whether the same recovery holds for segmentation or disease prediction is untested, though the mechanism suggests tasks that depend on structural fidelity would benefit most.
Training SR only on bicubic downsampling may not match real handheld-device artifacts such as speckle, clutter, and beam-forming degradation, so a domain-adapted degradation model would likely be needed to realize the gains in actual clinics.
A direct comparison against training the classifier with quality augmentation or on native poor frames would clarify whether the SR front-end adds value beyond simple data augmentation.
The reported improvements are percentage-point differences on small test sets, so the effect sizes carry uncertainty and would benefit from multi-site validation before deployment decisions.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a sensible, well-scoped empirical study of whether SR preprocessing can recover classification accuracy on poor-quality echoes. The result is plausible and the direction is consistent, but the evidence as reported is not yet trustworthy: the split in §4.2 is not stated to be patient-level, and with CAMUS's paired ED/ES and 2CH/4CH frames, a frame-level split likely leaks patients between train and test. That alone undermines the headline 'recovery' numbers.

What's actually new: quality-stratified evaluation on CAMUS for downstream view/phase classification, comparing SRResNet vs SRGAN, including test-time-only SR without retraining. Those are legitimate extensions of prior SR-for-ultrasound work, which focused on perceptual quality or segmentation. The SRResNet-wins-over-SRGAN finding is interesting and consistent with the PSNR/SSIM numbers. The paper is clearly written and the related work is honest. I also appreciate that they report the per-quality transfer tables, which show the 'diagnostic collapse' pattern clearly.

Soft spots, in order of severity. First, the patient-level leakage issue. The paper says an 80/20 split but never says patient-level. With 232 poor frames from at most 116 patients, frame-level splitting puts the same patient's ED/ES (and often 2CH/4CH) in both training and test. A ResNet-18 can memorize patient-specific texture and probe artifacts, inflating the cross-quality gains in Fig. 4 — exactly the numbers that carry the central claim. This needs to be fixed, ideally with patient-level splits and repeated runs. Second, the test sets are tiny: 20% of 232 frames is ~46 images per task, so a few frames swing accuracy by several points. There are no error bars or confidence intervals. Also, the phase-task gains are modest; the strong claim really rests on the view task and the retraining scenario. Third, the SR models are fine-tuned on bicubic-downsampled good frames and applied to native poor frames; that domain gap is a real threat to external validity, though secondary. Fourth, no code or data are released, which makes it hard to check the split and the preprocessing.

On balance, I'd send this to peer review. The question is worth asking, the design is fixable, and the paper makes claims that are specific enough to test. But it needs major revision: patient-level splits, repeated runs, error bars, and ideally code/data release before the numbers mean much.

Referee Report

3 major / 5 minor

Summary. The paper investigates whether super-resolution (SR) preprocessing can recover diagnostic value in poor-quality 2D echocardiograms, motivated by resource-constrained settings. Using the CAMUS dataset, the authors stratify frames by expert-annotated image quality (good, medium, poor) and evaluate two classification tasks: 2CH vs. 4CH view classification and ED vs. ES phase classification. A ResNet-18 classifier is trained on subsets stratified by quality and tested across quality levels. SRGAN and SRResNet are fine-tuned on synthetic low-resolution pairs created by bicubic downsampling of good-quality frames and then applied to the 232 poor-quality frames. The reported results show accuracy gains after SR in both tasks, with SRResNet generally outperforming SRGAN while being computationally cheaper. The paper concludes that SR is a viable lightweight preprocessing tool for AI-assisted echocardiographic interpretation in resource-limited environments.

Significance. If the central result holds, the paper offers a practical, model-agnostic preprocessing step that could improve automated echocardiography in low-resource settings, where poor image quality is common. Strengths of the study include the use of clinically validated quality metadata instead of a resolution proxy, the evaluation of two tasks of differing difficulty, the use of a public dataset, and the explicit comparison of two SR architectures including computational cost. The empirical setup is straightforward and the paper does not rely on any fitted derivation, so circularity is not a concern. However, the current evaluation design does not yet support the strength of the claims: the train/test split is not demonstrated to be patient-level, test sets are small, and no uncertainty quantification or repeated runs are reported. These issues are load-bearing because the central claim is a measured accuracy improvement, which can be inflated by patient leakage and is sensitive to a handful of test images.

major comments (3)

[§4.2] The 80/20 train/test split is not stated to be patient-level, and this is the most serious threat to internal validity. In CAMUS, each patient contributes four frames (2CH-ED, 2CH-ES, 4CH-ED, 4CH-ES), and Table 1 shows identical poor counts for ED and ES within each view, which is consistent with per-sequence quality annotations. The 232 poor frames therefore come from at most 116 patients (69 with a poor 2CH sequence and 47 with a poor 4CH sequence). If the split is performed at the frame level, then ED and ES frames (and often 2CH and 4CH frames) from the same patient appear in both training and test sets. A ResNet-18 can exploit patient-specific texture, noise, or probe artifacts as a shortcut, inflating the accuracies in Table 2 and, more importantly, the cross-quality comparisons in Figure 4 that carry the central claim. The authors must repeat the evaluation with a patient-level split (or report results under both split strategies) and report the number of patients in each fold.
[§4.2, Table 2, Figure 4] The reported accuracies are based on a single train/test split with test sets of only about 46 images (20% of 232). For the view task, that is roughly 28 2CH and 19 4CH test frames; for the phase task, roughly 23 ED and 23 ES frames. A single test image corresponds to about 2.2 percentage points of accuracy, so differences of a few points, such as the 1.3% and 3.7% gains reported in Section 5 item 3, are within the noise of the experiment. No confidence intervals, bootstrap estimates, or repeated runs with different seeds are provided. The phrase "significant gains" in the abstract and Section 5 is not supported by statistical evidence. The authors should provide confidence intervals or repeated-run variability, and ideally report raw counts of correct/incorrect predictions for each cell in Table 2 and each bar in Figure 4.
[§4.3] The SR models are trained on synthetic low-resolution pairs created by bicubic downsampling of good-quality frames and then applied to native poor-quality frames. The assumption that real resource-constrained acquisition degradation is well modeled by bicubic downsampling is not tested. Poor-quality echocardiograms exhibit speckle, low contrast, and operator-dependent artifacts, which may not be captured by this synthetic degradation. Additionally, the paper does not compare SR against a simple baseline of bicubic upsampling of the poor-quality frames to the same target size, or against other lightweight enhancement methods such as histogram equalization or denoising. Without such a baseline, the reported gains could be attributed simply to increasing spatial resolution rather than to learned SR reconstruction. A controlled experiment on native poor frames, or at least a synthetic degradation that better mimics ultrasound physics, is needed to support the claim that the method transfers to real resource-constrained settings.

minor comments (5)

[Figure 4] The bars in Figure 4 are labeled only as "percentage improvement" with no numeric values, error bars, or indication of which baseline each improvement is computed against; the authors should add numeric labels or a companion table reporting exact values and the number of test samples.
[§4.3] The evaluation of SR quality in Figure 3 reports PSNR and SSIM values (32.70 dB / 0.7164 for SRGAN, 38.98 dB / 0.9214 for SRResNet) but does not state what the reference image is for these metrics; if the reference is the original poor-quality frame, the comparison is not meaningful across different image sizes, and if it is a good-quality frame, that should be stated explicitly.
[§5, Observation 1] The statement that medium-quality images "strike an effective balance between noise and structural fidelity—making them an ideal candidate for RCS" is an interpretive claim that goes beyond the presented accuracy numbers; it would benefit from a statistical comparison or at least explicit acknowledgment that the differences may not be significant given the test-set sizes.
[§4.2] The training details report a fixed batch size, learning rate, and 10 epochs for the classifier, but no validation-based early stopping or selection procedure is described; because the test set is used only once, the authors should clarify how hyperparameters were chosen to avoid implicit test-set tuning.
[Abstract] The abstract and Section 5 use the word "significant" repeatedly, but no significance tests are reported; the authors should either add statistical tests or replace "significant" with "measured" or "observed."

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported SR gains are empirical measurements, with SR models trained on disjoint good-quality synthetic pairs and evaluated on native poor-quality frames.

full rationale

The central claim is empirical rather than derivational. The SR modules are fine-tuned on 1,010 good-quality frames synthetically degraded by bicubic downsampling (Section 4.3), then applied to the 232 native poor-quality frames from the clinically annotated D_poor stratum. Classification accuracies in Table 2 and Figure 4 are measured by evaluating a fixed ResNet-18 on original versus SR-enhanced inputs; no target quantity is introduced as a fitted parameter. The PSNR/SSIM values in Figure 3 are reported reconstruction metrics, not quantities used to derive the accuracy gains. The paper relies on standard external sources (CAMUS, ResNet-18, SRGAN/SRResNet) and contains no self-citations, no imported uniqueness theorem, and no ansatz smuggled in via citation. The most notable concern raised by the reviewer, that the 80/20 split in Section 4.2 is not explicitly stated to be patient-level, is a data-leakage and internal-validity issue rather than a circularity: it does not make any reported prediction equivalent to the model's inputs by construction. Accordingly, the derivation chain is self-contained and no circular step is present.

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

The paper introduces no new theoretical entities or fitted physical constants. Its central claim depends on experimental hyperparameters and domain assumptions about degradation and quality labels rather than a derivation.

free parameters (5)

SR upsampling factor r = 4
Chosen by hand in Section 4.3; the enhancement scale affects how much detail SR can add and therefore the measured classification gains.
SR fine-tuning epochs = SRResNet 4000, SRGAN 8000
Training durations in Section 4.3 are arbitrary; longer training may improve SR fidelity and inflate downstream gains.
Classifier training epochs = 10
ResNet-18 fine-tuning for 10 epochs (Section 4.2) is a hand-chosen stopping point that affects baseline and post-SR accuracy.
Batch size and learning rate = 16, 1e-4
Optimizer settings in Section 4.2 are standard but arbitrary; no sensitivity analysis is provided.
Train/test split ratio = 80/20
The split in Section 4.2 determines the (small) test set size; no repeats or cross-validation are performed.

assumptions (4)

domain assumption CAMUS expert quality labels (good/medium/poor) are a valid proxy for image quality in resource-constrained settings.
Section 4.1 treats D_poor as representative of RCS; if the CAMUS quality ratings do not reflect real RCS acquisition conditions, the whole evaluation setup is invalid.
domain assumption Bicubic downsampling by a factor of 4 approximates the degradation present in poor-quality echocardiograms.
Section 4.3 constructs synthetic LR pairs from good frames via bicubic downsampling; the fine-tuned SR models are then applied to native poor frames. This transfer is assumed, not validated.
domain assumption ImageNet-pretrained ResNet-18 features transfer to echocardiographic view and phase classification.
Section 4.2 fine-tunes an ImageNet-pretrained ResNet-18; no ablation of the pretraining choice is given.
domain assumption The two classification tasks (2CH vs 4CH, ED vs ES) are meaningful proxies for diagnostic value.
The paper equates classification accuracy on these tasks with 'recovering diagnostic value'; clinical utility is asserted, not measured.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Recovering Diagnostic Value: Super-Resolution-Aided Echocardiographic Classification in Resource-Constrained Imaging." pith.science (2026). https://pith.science/paper/Y3EDOUI3

@misc{pith2026250723027,
  author       = {Pith},
  title        = {Pith review of: Recovering Diagnostic Value: Super-Resolution-Aided Echocardiographic Classification in Resource-Constrained Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y3EDOUI3}},
  note         = {Machine review of arXiv:2507.23027}
}

read the original abstract

Automated cardiac interpretation in resource-constrained settings (RCS) is often hindered by poor-quality echocardiographic imaging, limiting the effectiveness of downstream diagnostic models. While super-resolution (SR) techniques have shown promise in enhancing magnetic resonance imaging (MRI) and computed tomography (CT) scans, their application to echocardiography-a widely accessible but noise-prone modality-remains underexplored. In this work, we investigate the potential of deep learning-based SR to improve classification accuracy on low-quality 2D echocardiograms. Using the publicly available CAMUS dataset, we stratify samples by image quality and evaluate two clinically relevant tasks of varying complexity: a relatively simple Two-Chamber vs. Four-Chamber (2CH vs. 4CH) view classification and a more complex End-Diastole vs. End-Systole (ED vs. ES) phase classification. We apply two widely used SR models-Super-Resolution Generative Adversarial Network (SRGAN) and Super-Resolution Residual Network (SRResNet), to enhance poor-quality images and observe significant gains in performance metric-particularly with SRResNet, which also offers computational efficiency. Our findings demonstrate that SR can effectively recover diagnostic value in degraded echo scans, making it a viable tool for AI-assisted care in RCS, achieving more with less.

Figures

Figures reproduced from arXiv: 2507.23027 by the authors.

**Figure 2.** Representative echocardiographic frames from the CAMUS dataset across [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** SR-enhanced outputs for a poor-quality 2CH ED image: (c) SRGAN — [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Percentage improvement in classification performance after integrating [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 20 canonical work pages

[1]

In: Proceedings of the International Conference on Advanced In- telligent Systems and Informatics 2016

Abdel-Nasser, M., Omer, O.A.: Ultrasound image enhancement using a deep learn- ing architecture. In: Proceedings of the International Conference on Advanced In- telligent Systems and Informatics 2016. pp. 639–649. Springer International Pub- lishing, Cham (2017).https://doi.org/10.1007/978-3-319-48308-5_61

work page doi:10.1007/978-3-319-48308-5_61 2017
[2]

Biomedical Signal Processing and Control78, 104003 (2022)

Afrakhteh, S., Jalilian, H., Iacca, G., Demi, L.: Temporal super-resolution of echocardiography using a novel high-precision non-polynomial interpolation. Biomedical Signal Processing and Control78, 104003 (2022). https://doi.org/ https://doi.org/10.1016/j.bspc.2022.104003

arXiv 2022
[3]

In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 1122–1131 (2017).https://doi.org/10. 1109/CVPRW.2017.150

work page 2017
[4]

Journal of Clinical Medicine 10(7) (2021)

Akkus, Z., Aly, Y.H., Attia, I.Z., et al.: Artificial intelligence (ai)-empowered echocardiography interpretation: A state-of-the-art review. Journal of Clinical Medicine 10(7) (2021). https://doi.org/10.3390/jcm10071391

work page doi:10.3390/jcm10071391 2021
[5]

Ashrafian, P., Yazdani, M., Heidari, M., Shahriari, D., Hacihaliloglu, I.: Vision- language synthetic data enhances echocardiography downstream tasks (2024), https://arxiv.org/abs/2403.19880

work page Pith review arXiv 2024
[6]

Tropical Medicine & International Health21(3), 294–311 (2015)

Becker, D.M., Tafoya, C.A., Becker, S.L., et al.: The use of portable ultrasound devices in low- and middle-income countries: a systematic review of the literature. Tropical Medicine & International Health21(3), 294–311 (2015). https://doi. org/https://doi.org/10.1111/tmi.12657

work page doi:10.1111/tmi.12657 2015
[7]

Medical & Biological Engineering & Computing61(10), 2511– 2526 (Oct 2023).https://doi.org/10.1007/s11517-023-02818-x

Cammarasana, S., Nicolardi, P., Patanè, G.: Super-resolution of 2d ultrasound images and videos. Medical & Biological Engineering & Computing61(10), 2511– 2526 (Oct 2023).https://doi.org/10.1007/s11517-023-02818-x

work page doi:10.1007/s11517-023-02818-x 2023
[8]

Circulation 136(22), 2178–2188 (2017)

Chamsi-Pasha, M.A., Sengupta, P.P., Zoghbi, W.A.: Handheld echocardiography: current state and future perspectives. Circulation 136(22), 2178–2188 (2017). https://doi.org/10.1161/CIRCULATIONAHA.117.026622

work page doi:10.1161/circulationaha.117.026622 2017

Show all 24 references

[9]

In: Bioinformatics and Biomedical Engineering

Fernández-Rodríguez, A., López-Rubio, E., Torres-Salomón, P., et al.: Enhanc- ing echocardiography quality with diffusion neural models. In: Bioinformatics and Biomedical Engineering. pp. 169–181. Springer Nature Switzerland, Cham (2024). https://doi.org/10.1007/978-3-031-64636-2_13

2024 doi
[10]

Frija, G., Blažić, I., Frush, D.P., et al.: How to improve access to medical imaging in low- and middle-income countries ? eClinicalMedicine38, 101034 (2021).https: //doi.org/https://doi.org/10.1016/j.eclinm.2021.101034

2021
[11]

IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control63(1), 6–19 (2016).https: //doi.org/10.1109/TUFFC.2015.2493881

Gifani,P.,Behnam,H.,Haddadi,F.,etal.:Temporalsuperresolutionenhancement of echocardiographic images based on sparse representation. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control63(1), 6–19 (2016).https: //doi.org/10.1109/TUFFC.2015.2493881

2016
[12]

Echocardiog- raphy 41(10), e15939 (Oct 2024).https://doi.org/10.1111/echo.15939, pMID: 39367770

Hamza, I., Pellikka, P.A., Abdulla, A., Ahmad, M.: Artificial intelligence echocar- diography in resource-limited regions: Applications and challenges. Echocardiog- raphy 41(10), e15939 (Oct 2024).https://doi.org/10.1111/echo.15939, pMID: 39367770

2024 doi
[13]

In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770–778 (2016).https://doi.org/10.1109/CVPR.2016.90

2016 doi
[14]

JMA Journal8(1), 141–150 (Jan 2025)

Hirata, Y., Kusunose, K.: Ai in echocardiography: State-of-the-art automated mea- surement techniques and clinical applications. JMA Journal8(1), 141–150 (Jan 2025). https://doi.org/10.31662/jmaj.2024-0180, epub 2024 Dec 6 10 K.A. Raja Babu et al

2025 doi
[15]

In: 2015 International Conference on Technologies for Sustainable Development (ICTSD)

Isaac, J.S., Kulkarni, R.: Super resolution techniques for medical image processing. In: 2015 International Conference on Technologies for Sustainable Development (ICTSD). pp. 1–6 (2015).https://doi.org/10.1109/ICTSD.2015.7095900

2015
[16]

IEEE Transac- tions on Medical Imaging 38(9), 2198–2210 (2019)

Leclerc, S., Smistad, E., Pedrosa, J., Ostvik, A., et al.: Deep learning for segmen- tation using an open large-scale dataset in 2d echocardiography. IEEE Transac- tions on Medical Imaging 38(9), 2198–2210 (2019). https://doi.org/10.1109/ TMI.2019.2900516

2019
[17]

Ledig, C., Theis, L., Huszar, F., et al.: Photo-realistic single image super-resolution using a generative adversarial network (2017), https://arxiv.org/abs/1609. 04802

2017
[18]

In: 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Soci- ety (EMBC)

Li, S., Zhang, M., Li, Y., Sumi, C.: Segmentation with speckle reduction and super- resolution by deep leaning for human ultrasonic echo image. In: 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Soci- ety (EMBC). pp. 3663–3667 (2021).ht...

2021
[19]

Heart, Lung and Circulation28(9), 1427– 1435 (2019)

Marangou, J., Beaton, A., Aliku, T.O., et al.: Echocardiography in indigenous populations and resource poor settings. Heart, Lung and Circulation28(9), 1427– 1435 (2019). https://doi.org/10.1016/j.hlc.2019.05.176

2019 doi
[20]

Michelis, K.C., Narotsky, D.L., Choi, B.G.: Cardiovascular Imaging in Global Health Radiology, pp. 207–224. Springer International Publishing, Cham (2019). https://doi.org/10.1007/978-1-4614-0604-4_18

2019 doi
[21]

Biomedical Engineering Letters14(3), 465–496 (May 2024)

Shin, M., Seo, M., Lee, K., Yoon, K.: Super-resolution techniques for biomedical applications and challenges. Biomedical Engineering Letters14(3), 465–496 (May 2024). https://doi.org/10.1007/s13534-024-00365-4, epub 2024 Mar 19

2024 doi
[22]

Journal of Cardiovas- cular Translational Research 7(8), 701–710 (2014)

Vedanthan, R., Choi, B.G., Baber, U., et al.: Bioimaging and subclinical car- diovascular disease in low- and middle-income countries. Journal of Cardiovas- cular Translational Research 7(8), 701–710 (2014). https://doi.org/10.1007/ s12265-014-9588-y

2014
[23]

Netherlands Heart Journal 31(2), 47–51 (Feb 2023).https://doi.org/10.1007/ s12471-022-01741-4, epub 2022 Dec 12

Wierda, E., Bouma, B.J., van den Brink, R.B.A.: Handheld echocardiography in patients with cardiovascular disease: to use or not to use, that is the question. Netherlands Heart Journal 31(2), 47–51 (Feb 2023).https://doi.org/10.1007/ s12471-022-01741-4, epub 2022 Dec 12

2023
[24]

Xochicale, M., Thwaites, L., Yacoub, S., et al.: A machine learning case study for ai-empowered echocardiography of intensive care unit patients in low- and middle- income countries (2023),https://arxiv.org/abs/2212.14510

2023 arXiv

Pith tools

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

[1] [1]

In: Proceedings of the International Conference on Advanced In- telligent Systems and Informatics 2016

Abdel-Nasser, M., Omer, O.A.: Ultrasound image enhancement using a deep learn- ing architecture. In: Proceedings of the International Conference on Advanced In- telligent Systems and Informatics 2016. pp. 639–649. Springer International Pub- lishing, Cham (2017).https://doi.org/10.1007/978-3-319-48308-5_61

work page doi:10.1007/978-3-319-48308-5_61 2017

[2] [2]

Biomedical Signal Processing and Control78, 104003 (2022)

Afrakhteh, S., Jalilian, H., Iacca, G., Demi, L.: Temporal super-resolution of echocardiography using a novel high-precision non-polynomial interpolation. Biomedical Signal Processing and Control78, 104003 (2022). https://doi.org/ https://doi.org/10.1016/j.bspc.2022.104003

arXiv 2022

[3] [3]

In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 1122–1131 (2017).https://doi.org/10. 1109/CVPRW.2017.150

work page 2017

[4] [4]

Journal of Clinical Medicine 10(7) (2021)

Akkus, Z., Aly, Y.H., Attia, I.Z., et al.: Artificial intelligence (ai)-empowered echocardiography interpretation: A state-of-the-art review. Journal of Clinical Medicine 10(7) (2021). https://doi.org/10.3390/jcm10071391

work page doi:10.3390/jcm10071391 2021

[5] [5]

Ashrafian, P., Yazdani, M., Heidari, M., Shahriari, D., Hacihaliloglu, I.: Vision- language synthetic data enhances echocardiography downstream tasks (2024), https://arxiv.org/abs/2403.19880

work page Pith review arXiv 2024

[6] [6]

Tropical Medicine & International Health21(3), 294–311 (2015)

Becker, D.M., Tafoya, C.A., Becker, S.L., et al.: The use of portable ultrasound devices in low- and middle-income countries: a systematic review of the literature. Tropical Medicine & International Health21(3), 294–311 (2015). https://doi. org/https://doi.org/10.1111/tmi.12657

work page doi:10.1111/tmi.12657 2015

[7] [7]

Medical & Biological Engineering & Computing61(10), 2511– 2526 (Oct 2023).https://doi.org/10.1007/s11517-023-02818-x

Cammarasana, S., Nicolardi, P., Patanè, G.: Super-resolution of 2d ultrasound images and videos. Medical & Biological Engineering & Computing61(10), 2511– 2526 (Oct 2023).https://doi.org/10.1007/s11517-023-02818-x

work page doi:10.1007/s11517-023-02818-x 2023

[8] [8]

Circulation 136(22), 2178–2188 (2017)

Chamsi-Pasha, M.A., Sengupta, P.P., Zoghbi, W.A.: Handheld echocardiography: current state and future perspectives. Circulation 136(22), 2178–2188 (2017). https://doi.org/10.1161/CIRCULATIONAHA.117.026622

work page doi:10.1161/circulationaha.117.026622 2017