Pith. sign in

REVIEW 2 major objections 6 minor 29 references

Advancing Fetal Ultrasound Image Quality Assessment in Low-Resource Settings

T0 review · 2 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 A fetal-ultrasound foundation model adapted with LoRA picks out usable frames in blind-sweep scans better than CNN and Transformer baselines, and a repurposed segmentation model improves the F1 score to 0.771.

desk verdict A fair empirical comparison showing LoRA-finetuned FetalCLIP beats generic ViT on fetal ultrasound frame quality, but the filtered test set keeps the 'deployable' claim from being fully supported. read the letter →

arxiv 2507.22802 v1 pith:DBWN4ZDO submitted 2025-07-30 cs.CV cs.AI

Dongli He , Hu Wang , Mohammad Yaqub This is my paper
classification cs.CVcs.AI
keywords fetalultrasoundimagequalityassessmentblind-sweepparameter-efficientfine-tuningLow-RankAdaptationfoundationmodelsvision-languagepretrainingbiometry
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 tries to establish that a foundation model pretrained on fetal ultrasound images, adapted cheaply with Low-Rank Adaptation (LoRA), can automatically pick out the few frames in a blind-sweep fetal scan that are good enough for abdominal circumference measurement. On the ACOUSLIC-AI dataset, its classification model FetalCLIP_CLS reaches the highest F1 score (0.757) among six CNN and Transformer baselines, and a segmentation model repurposed for classification, FetalCLIP_SEG, does even better (0.771 F1) while using mask supervision. The authors further claim that domain-specific pretraining matters more than scale: FetalCLIP_CLS, trained on 210,000 fetal ultrasound image-caption pairs, beats comparable ViT models pretrained on 400 million and 2 billion generic image-text pairs. If true, this provides a computationally light (about 2.4 million trainable parameters, single-GPU training) way to improve prenatal care where expert sonographers are scarce.

What carries the argument

The load-bearing mechanism is Low-Rank Adaptation (LoRA) applied to a frozen Transformer encoder. LoRA inserts small trainable low-rank matrices into the attention and feed-forward blocks, so the FetalCLIP image encoder (pretrained on over 210,000 fetal ultrasound image-caption pairs with contrastive language-image pretraining) adapts to the IQA task by learning only about 2.4 million parameters. For FetalCLIP_CLS, the adapted encoder's pooled embedding feeds a single linear classification head. For FetalCLIP_SEG, the same encoder feeds a lightweight U-shaped decoder inspired by UNETR, trained with Dice loss on expert masks; predicted masks are converted to binary class labels by thresholding the foreground pixel count at roughly 1% of the image area (about 500 pixels). This shared encoder-plus-LoRA design is what lets both models stay parameter-efficient while outperforming full fine-tuning of larger generic models.

What would settle it

Evaluate the released FetalCLIP_CLS and FetalCLIP_SEG models on original unfiltered blind-sweep sequences from the same dataset, where only 2.6 out of every 100 frames contain usable abdominal structures, and measure precision-recall; if the F1 scores fall below the reported values by a large margin, or if the fixed 1% foreground threshold labels many empty frames as positive, the transferability claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that combining a frozen, domain-specific ultrasound encoder with lightweight adaptation solves fetal ultrasound image quality assessment better than generic pretrained models of any size. FetalCLIP_CLS, built from the frozen FetalCLIP image encoder plus LoRA modules and a linear head, reaches an F1 of 0.757 with accuracy 0.9575, beating all six baselines and, in a follow-up comparison, a ViT pretrained on 2 billion image-text pairs (F1 0.7429). A second result is the repurposing of a segmentation model for the same task: FetalCLIP_SEG uses the same LoRA-adapted encoder with a lightweight U-shaped decoder trained on expert masks, and converts its predicted masks into labels using a foreground threshold of 1% of image area. It outperforms the dedicated classifier on F1 (0.7708) and recall (0.8599), at the cost of precision (0.6988), showing that pixel-level supervision can be turned into a strong frame-quality signal.

Load-bearing premise

For training and testing, all sweeps without any annotated masks are removed, which raises the share of usable frames from 2.6 per 100 to 8.6 per 100; the central premise is that this filtered distribution represents the real blind-sweep scans the model will see in the field, so the reported F1 scores and segmentation threshold will transfer.

Editorial extensions

If this is right

  • FetalCLIP_CLS can be deployed as a frame filter on blind-sweep fetal ultrasound, running on a single GPU with only 2.4M trainable parameters.
  • Transformer-based IQA with LoRA should be preferred over full fine-tuning of CNNs for this task, since every Transformer baseline beat every CNN baseline on F1 and recall.
  • Domain-specific pretraining is more data-efficient than generic scale: 210k domain image-text pairs outperform 400M and 2B generic pairs on this task.
  • FetalCLIP_SEG's mask-thresholding route gives higher F1 and much higher recall (0.8599) than the dedicated classifier, at about 8% lower precision, favoring it when missing usable frames is the worst failure.
  • Mask supervision can substitute for clean classification labels, since segmentation training plus a trivial threshold yields competitive or better classification performance.

Reading between the lines

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

  • A corollary not tested here: the reported numbers are measured at an 8.6% positive-frame rate after dropping mask-free sweeps; on raw sweeps at the original 2.6% rate, the optimal threshold—including FetalCLIP_SEG's 1% rule—would likely need recalibration, so field F1 may differ from 0.757/0.771.
  • The comparison implies a testable scaling law for medical foundation models: plot upstream-domain relevance against downstream F1 while varying pretraining set size; the paper's numbers suggest domain relevance dominates scale by orders of magnitude for fetal ultrasound.
  • A two-stage pipeline combining FetalCLIP_SEG (coarse, high-recall filter) with FetalCLIP_CLS (precision-oriented re-ranker) could recover both high recall and precision; the paper does not test this combination.
  • The hand-set 1% foreground threshold is a free parameter; sweeping it on a held-out set of unfiltered sweeps would quantify how much performance depends on the filtering assumption.
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

2 major / 6 minor

Summary. The paper proposes FetalCLIP_CLS, a binary frame-quality classifier built by adapting the fetal ultrasound foundation model FetalCLIP with Low-Rank Adaptation (LoRA), and FetalCLIP_SEG, a segmentation model repurposed for classification via mask thresholding. The models are evaluated on the ACOUSLIC-AI challenge dataset for identifying frames suitable for fetal abdominal circumference measurement. The authors report FetalCLIP_CLS achieves the best F1 score of 0.757 among classification models and FetalCLIP_SEG further improves F1 to 0.771, concluding that domain-specific pretraining combined with parameter-efficient fine-tuning offers a deployable solution for low-resource settings. The paper includes patient-level splits, five repeated runs, an ablation of fine-tuning strategies, and publicly released code.

Significance. If the central claims hold, the paper makes a useful contribution by demonstrating that a domain-specific vision-language foundation model can be adapted with LoRA for fetal ultrasound IQA, requiring only 2.4M trainable parameters, which is relevant for low-resource deployment. The experimental protocol is largely sound: patient-level data splits prevent leakage, five runs provide variance estimates, the external ACOUSLIC-AI benchmark was not used in FetalCLIP pretraining (so the comparison is not circular), and the code is released. The ablation of linear probing, full fine-tuning, and LoRA is informative. However, the evaluation is conducted on a filtered subset of the challenge data, which materially changes the label distribution and removes all-negative sweeps from the test set; this limits the support for the deployment-oriented claims in the abstract and conclusion.

major comments (2)
  1. [§4.1, Table 1, §4.3] The dataset filtering described in §4.1 is load-bearing for the paper's central claim of a deployable frame filter for blind-sweep ultrasound. By excluding all sweeps without annotated masks, the authors change the positive-frame rate from 2.6% in the original challenge to 8.6% and remove every all-negative sweep from the train, validation, and test sets. Since F1, precision, and recall depend on prevalence, the reported scores in Tables 1 and 2 estimate performance on the filtered distribution, not on the blind-sweep distribution that the abstract and conclusion target. In particular, the false-positive rate on all-negative sweeps is never measured, and the 1% foreground-pixel threshold used by FetalCLIP_SEG (§4.3) is only validated on this filtered distribution. To support the deployment claim, the authors should evaluate on the full unfiltered test frames (including all-negative sweeps) and report sweep-level metrics such as the fraction of sweeps with no retained frames, as well as PR or ROC curves across the original prevalence.
  2. [Table 1 and Table 2] The main quantitative claim that FetalCLIP_CLS outperforms the strongest Transformer baselines is not supported by significance testing. In Table 1, the F1 gap over ViT400M is 0.7570 ± 0.007 versus 0.7506 ± 0.019 across five runs; the difference is smaller than the reported standard deviations. Similarly, in Table 2, the F1 gain of FetalCLIP_SEG over FetalCLIP_CLS (0.7708 ± 0.005 versus 0.7570 ± 0.007) is not tested. The paper should report paired significance tests across the five runs (e.g., Wilcoxon signed-rank or bootstrap) and, ideally, patient-level confidence intervals, before claiming 'consistently outperforms' and 'further improves performance.'
minor comments (6)
  1. [§2, Reference [28]] The related-work text mentions 'Transformer-based methods like TRIQ [28]', but reference [28] is the ManiQA paper, not the TRIQ paper; the citation should be corrected or the text updated.
  2. [Table 1 header] The header contains a typo: '# T rainable' should be '# Trainable'.
  3. [Table 2] The metric label 'DICE' should be 'Dice' for consistency with standard usage.
  4. [§4.4] The name 'ViT400M' is ambiguous; the authors should explicitly state that this is a Vision Transformer pretrained on 400M image-text pairs, and identify the exact checkpoint used.
  5. [Figure 3 caption] The caption states that in the rightmost example 'both the ground truth and prediction contain no mask, indicating a correctly segmented frame'; this should be clarified because an empty prediction for a frame with no ground-truth mask is trivially correct, and it would be more informative to show the percentage of such cases.
  6. [Appendix A.1, Table 3] The row for DenseNet with linear probing reports F1 = 0.3569 ± 0.018 with precision and recall that appear inconsistent with that F1 (0.6113 and 0.2525 yield a harmonic mean of about 0.357, which is correct); no change needed, but a note that precision/recall are consistent would help readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core F1 comparison is measured on a held-out external benchmark with independent FetalCLIP checkpoints.

full rationale

The claimed derivation chain is: (i) take FetalCLIP, a pretrained fetal-ultrasound vision-language encoder; (ii) adapt it with LoRA on ACOUSLIC-AI training frames; (iii) measure F1 on a held-out ACOUSLIC-AI test split. FetalCLIP was pretrained on 210k fetal ultrasound image-caption pairs, not on ACOUSLIC-AI labels or masks, so the adaptation target is external to the pretraining objective. The FetalCLIP checkpoint is cited from prior work with overlapping authorship, but it is an independently trained model rather than a parameter fitted to the IQA labels; the comparison against ImageNet/CLIP/LAION baselines is therefore a genuine transfer-learning experiment. FetalCLIPSEG uses ground-truth masks for training and a fixed 1%-foreground threshold for conversion; the threshold is a decision rule, not a fitted parameter renamed as a prediction, and the reported Dice and F1 are computed on held-out frames. The exclusion of sweeps without annotated masks changes prevalence from 2.6% to 8.6% and is a real external-validity limitation, but it is a dataset-filtering choice, not a circular reduction of the central claim to its inputs. Self-citations to FUSQA are related-work context and do not carry the derivation. No equation or result in the paper is equivalent to its input by construction.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim does not rest on fitted constants in the mathematical sense; the main tuned values are the mask threshold and the LoRA configuration. The axioms are standard transfer-learning and annotation-meaning assumptions.

free parameters (2)
  • Segmentation mask foreground threshold = 1% of image area (~500 pixels)
    Chosen by hand in §4.3 to convert predicted masks to binary labels; no ablation or sensitivity analysis is provided, and FetalCLIP_SEG's classification metrics depend on it.
  • LoRA rank and alpha = Not stated in the paper
    The LoRA configuration determines adapter capacity for both FetalCLIP_CLS and the Transformer baselines; values are left to the code repository, so the claimed parameter efficiency is not fully auditable from the text.
assumptions (3)
  • domain assumption The presence of an expert-annotated abdominal region mask defines the ground truth for image quality suitability.
    The paper equates y=1 (frame suitable for biometric measurement) with the existence of anatomical structures in the ACOUSLIC-AI masks (§3.1, §4.1). A frame could be geometrically suitable yet mask-free due to annotation conventions, or mask-bearing yet of poor imaging quality for AC measurement.
  • ad hoc to paper A segmentation model that perfectly segments the target structure can classify frames without error.
    Stated as the motivating hypothesis for FetalCLIP_SEG in §3.3. It is plausible but not proven; the paper only tests one thresholded instantiation.
  • domain assumption FetalCLIP's frozen image encoder provides transferable representations for the IQA task.
    The entire method rests on the assumption that CLIP-style alignment on 210K ultrasound image-caption pairs yields features that linearly separate good from bad frames after LoRA adaptation (§3.2). This is the standard transfer-learning premise of the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing Fetal Ultrasound Image Quality Assessment in Low-Resource Settings." pith.science (2026). https://pith.science/paper/DBWN4ZDO

@misc{pith2026250722802,
  author       = {Pith},
  title        = {Pith review of: Advancing Fetal Ultrasound Image Quality Assessment in Low-Resource Settings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DBWN4ZDO}},
  note         = {Machine review of arXiv:2507.22802}
}
abstract

Accurate fetal biometric measurements, such as abdominal circumference, play a vital role in prenatal care. However, obtaining high-quality ultrasound images for these measurements heavily depends on the expertise of sonographers, posing a significant challenge in low-income countries due to the scarcity of trained personnel. To address this issue, we leverage FetalCLIP, a vision-language model pretrained on a curated dataset of over 210,000 fetal ultrasound image-caption pairs, to perform automated fetal ultrasound image quality assessment (IQA) on blind-sweep ultrasound data. We introduce FetalCLIP$_{CLS}$, an IQA model adapted from FetalCLIP using Low-Rank Adaptation (LoRA), and evaluate it on the ACOUSLIC-AI dataset against six CNN and Transformer baselines. FetalCLIP$_{CLS}$ achieves the highest F1 score of 0.757. Moreover, we show that an adapted segmentation model, when repurposed for classification, further improves performance, achieving an F1 score of 0.771. Our work demonstrates how parameter-efficient fine-tuning of fetal ultrasound foundation models can enable task-specific adaptations, advancing prenatal care in resource-limited settings. The experimental code is available at: https://github.com/donglihe-hub/FetalCLIP-IQA.

Figures

Figures reproduced from arXiv: 2507.22802 by the authors.

Figure 1
Figure 1. (a) Contrastive pretraining of FetalCLIP, where the model learns to align [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualizations of the feature embeddings extracted from the penultimate [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Each vertical pair displays the ground truth annotation (top) and its [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 11 canonical work pages

  1. [1]

    Scientific Reports 15, 8376 (2025)

    Boumeridja, H., Ammar, M., Alzubaidi, M., Mahmoudi, S., Benamer, L.N., Agus, M., Househ, M., Lekadir, K., Daho, M.E.H.: Enhancing fetal ultra- sound image quality and anatomical plane recognition in low-resource set- tings using super-resolution models. Scientific Reports 15, 8376 (2025). https://doi.org/10.1038/s41598-025-91808-0

    doi:10.1038/s41598-025-91808-0 2025

  2. [2]

    Cengiz, S., Almakky, I., Yaqub, M.: Fusqa: Fetal ultrasound segmentation quality assessment (2023), https://arxiv.org/abs/2303.04418

    work page Pith review arXiv 2023

  3. [3]

    In: 2023 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)

    Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gor- don, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: 2023 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR). p. 2818–2829. IEEE (Jun 2023). https://doi.org/10.1109/cvpr52729.2023.00276

    work page arXiv 2023

  4. [4]

    In: 2009 IEEE Confer- ence on Computer Vision and Pattern Recognition

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Ima- genet: A large-scale hierarchical image database. In: 2009 IEEE Confer- ence on Computer Vision and Pattern Recognition. pp. 248–255 (2009). https://doi.org/10.1109/CVPR.2009.5206848

    arXiv 2009

  5. [5]

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021), https://arxiv.org/abs/2010.11929

    arXiv 2021

  6. [6]

    Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H., Xu, D.: Unetr: Transformers for 3d medical image segmentation (2021), https://arxiv.org/abs/2103.10504

    arXiv 2021

  7. [7]

    org/abs/2106.09685

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models (2021),https://arxiv. org/abs/2106.09685

    arXiv 2021

  8. [8]

    Huang, G., Liu, Z., van der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks (2018),https://arxiv.org/abs/1608.06993

    arXiv 2018

Show all 29 references
  1. [9]

    In: 2014 IEEE Confer- ence on Computer Vision and Pattern Recognition

    Kang, L., Ye, P., Li, Y., Doermann, D.: Convolutional neural net- works for no-reference image quality assessment. In: 2014 IEEE Confer- ence on Computer Vision and Pattern Recognition. pp. 1733–1740 (2014). https://doi.org/10.1109/CVPR.2014.224

    2014 doi

  2. [10]

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows (2021),https: //arxiv.org/abs/2103.14030

    2021 arXiv

  3. [11]

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization (2019),https: //arxiv.org/abs/1711.05101

    2019 arXiv

  4. [12]

    Ma, C., Shi, Z., Lu, Z., Xie, S., Chao, F., Sui, Y.: A survey on image quality as- sessment: Insights, analysis, and future outlook (2025),https://arxiv.org/abs/ 2502.08540

    2025 arXiv

  5. [13]

    Maani,F.,Saeed,N.,Saleem,T.,Farooq,Z.,Alasmawi,H.,Diehl,W.,Mohammad, A., Waring, G., Valappi, S., Bricker, L., Yaqub, M.: Fetalclip: A visual-language foundation model for fetal ultrasound image analysis (2025),https://arxiv.org/ abs/2502.14807

    2025

  6. [14]

    04797 10 D

    Milletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks forvolumetricmedicalimagesegmentation(2016), https://arxiv.org/abs/1606. 04797 10 D. He et al

    2016

  7. [15]

    IEEE Transactions on Image Processing21(12), 4695–4708 (2012)

    Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing21(12), 4695–4708 (2012). https://doi.org/10.1109/TIP.2012.2214050

    2012

  8. [16]

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision (2021),https://arxiv.org/abs/ 2103.00020

    2021 arXiv

  9. [17]

    IEEE Transactions on Image Process- ing 21(8), 3339–3352 (2012)

    Saad, M.A., Bovik, A.C., Charrier, C.: Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE Transactions on Image Process- ing 21(8), 3339–3352 (2012). https://doi.org/10.1109/TIP.2012.2191563

    2012

  10. [18]

    Medical Image Analysis 105, 103640 (2025)

    Sappia, M.S., de Korte, C.L., van Ginneken, B., Ninalga, D., Kondo, S., Kasai, S., Hirasawa, K., Akumu, T., Martín-Isla, C., Lekadir, K., Campello, V.M., Fa- bila, J., Beverdam, A., van Dillen, J., Neff, C., Murphy, K.: Acouslic-ai challenge report: Fetal abdominal circumferen...

    2025

  11. [19]

    JMIR Res Protoc11(9), e37374 (2022)

    Self, A., Chen, Q., Desiraju, B.K., Dhariwal, S., Gleed, A.D., Mishra, D., Thiru- vengadam, R., Chandramohan, V., Craik, R., Wilden, E., Khurana, A., Group, T.C.S., Bhatnagar, S., Papageorghiou, A.T., Noble, J.A.: Developing clinical arti- ficial intelligence for obstetric ult...

    2022 doi

  12. [20]

    Sendra-Balcells, C., Campello, V.M., Torrents-Barrena, J., Ahmed, Y.A., Elattar, M., Botwe, B.O., Nyangulu, P., Stones, W., Ammar, M., Benamer, L.N., Kisembo, H.N., Sereke, S.G., Wanyonyi, S.Z., Temmerman, M., Gratacós, E., Bonet, E., Eixarch, E., Mikolaj, K., Tolsgaard, M.G.,...

    2023 arXiv

  13. [21]

    IEEE Transactions on ImageProcessing 15(2),430–444(2006).https://doi.org/10.1109/TIP.2005.859378

    Sheikh, H., Bovik, A.: Image information and visual quality. IEEE Transactions on ImageProcessing 15(2),430–444(2006).https://doi.org/10.1109/TIP.2005.859378

    2006

  14. [22]

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2015),https://arxiv.org/abs/1409.1556

    2015 arXiv

  15. [23]

    Tan, M., Le, Q.V.: Efficientnet: Rethinking model scaling for convolutional neural networks (2020), https://arxiv.org/abs/1905.11946

    2020 arXiv

  16. [24]

    Touvron, H., Cord, M., Jégou, H.: Deit iii: Revenge of the vit (2022), https: //arxiv.org/abs/2204.07118

    2022 arXiv

  17. [25]

    IEEE Transactions on Image Processing 13(4), 600–612 (2004)

    Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600–612 (2004). https://doi.org/10.1109/TIP.2003.819861

    2004

  18. [26]

    Cochrane Database of Systematic Reviews (2010)

    Whitworth, M., Bricker, L., Neilson, J., Dowswell, T.: Ultrasound for fetal as- sessment in early pregnancy. Cochrane Database of Systematic Reviews (2010). https://doi.org/10.1002/14651858.CD007058.pub3

    2010 doi

  19. [27]

    IEEE Transactions on Cy- bernetics 47(5), 1336–1349 (2017)

    Wu, L., Cheng, J.Z., Li, S., Lei, B., Wang, T., Ni, D.: Fuiqa: Fetal ultrasound image quality assessment with deep convolutional networks. IEEE Transactions on Cy- bernetics 47(5), 1336–1349 (2017). https://doi.org/10.1109/TCYB.2017.2671898

    2017

  20. [28]

    Yang, S., Wu, T., Shi, S., Lao, S., Gong, Y., Cao, M., Wang, J., Yang, Y.: Maniqa: Multi-dimension attention network for no-reference image quality assess- ment (2022), https://arxiv.org/abs/2204.08958

    2022 arXiv

  21. [29]

    Zuiderveld, K.: Contrast limited adaptive histogram equalization, p. 474–485. Aca- demic Press Professional, Inc., USA (1994) Title Suppressed Due to Excessive Length 11 A Ablation Study A.1 Selection of Fine-Tuning Strategies We evaluate three fine-tuning strategies, includin...

    1994

Pith tools

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