Invisible Yet Detected: PelFANet with Attention-Guided Anatomical Fusion for Pelvic Fracture Diagnosis

Ashraful Islam; Naomi Yagi; Rashedur Rahman; Saadia Binte Alam; Sefatul Wasi; Siam Tahsin Bhuiyan; Syoji Kobashi

REVIEW 4 major objections 5 minor 31 references

Invisible Yet Detected: PelFANet with Attention-Guided Anatomical Fusion for Pelvic Fracture Diagnosis

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read PelFANet claims a dual-stream network trained only on visible fractures detects CT-confirmed pelvic fractures that are invisible on X-rays.

desk verdict PelFANet is a plausible incremental step on a clinically interesting task, but the invisible-fracture claim rests on 35 images and an unspecified segmentation/classification split alignment that could leak test information. read the letter →

arxiv 2509.13873 v3 pith:SVNUW236 submitted 2025-09-17 cs.CV

Siam Tahsin Bhuiyan , Rashedur Rahman , Sefatul Wasi , Naomi Yagi , Syoji Kobashi , Ashraful Islam , Saadia Binte Alam This is my paper

classification cs.CV

keywords pelvicfractureinvisibledetectiondual-streamattentionnetworksegmentation-guidedclassificationCBAMX-rayCT-confirmedfractures

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 sets out to show that pelvic fractures invisible on plain X-rays can still be detected by a network that sees both the raw radiograph and a bone-segmentation mask. PelFANet combines these two inputs through attention-guided fusion blocks, and although it is trained only on visible fractures, it reports 82.29% accuracy and 0.8688 AUC on a CT-confirmed invisible-fracture subset, beating the prior best baseline. If the result holds, it would mean subtle fracture cues are present in the X-ray pixel data and recoverable with the right architecture, which matters because missed pelvic fractures are a known source of delayed treatment and mortality.

What carries the argument

The core unit is the Fused Attention Block (FABlock): at each of eight blocks, feature maps from the raw-X-ray stream and the bone-segmentation stream are independently convolved, concatenated, passed through a Convolutional Block Attention Module (CBAM) that applies channel then spatial attention, projected with a 1×1 convolution, and added back into both streams through residual connections. This exchange-refine-redistribute cycle is what lets the model combine global context from the raw image with localized anatomical structure from a U-Net-generated bone mask, and it is the mechanism the paper credits for detecting fractures that show no visible radiographic sign.

What would settle it

Retrain the segmentation model only on the training folds, regenerate all test masks with that model, and re-run PelFANet's 5-fold evaluation; if the INVIS AUC falls from 0.8688 toward the 0.8002 baseline, the reported transfer was inflated by mask leakage.

Watch

Extended reading notes

Core claim

The paper claims that a dual-stream attention network, PelFANet, trained only on radiographically visible pelvic fractures, can generalize to CT-confirmed fractures that are invisible on X-rays. Feeding each pelvic X-ray alongside an automatically segmented bone image, the network iteratively fuses the two inputs through eight attention blocks and reaches 82.29% accuracy and 0.8688 AUC on the 35-case invisible-fracture subset, improving on the previous best AUC of 0.8002, while maintaining 88.68% accuracy and 0.9334 AUC on visible fractures. The authors attribute this transfer to the fusion mechanism, which lets global context cues compensate for the absence of a visible cortical break.

Load-bearing premise

The argument depends on the two inputs being independent of the test set: the paper never states whether the bone-segmentation model that produced the masks was trained on the same X-ray images later used to evaluate the classifier, and if it was, the classifier could be receiving test information through the second input.

Editorial extensions

If this is right

Trained only on VIS cases, PelFANet's INVIS results imply that invisible fractures leave detectable traces in the X-ray pixel distribution, contradicting the assumption that no signal exists when no cortical break is visible.
The dual-input design preserves non-local cues such as joint spacing, alignment, and soft-tissue contour that pure bone-segmentation models discard.
An automated second look could flag subtle pelvic fractures for radiologists, potentially reducing the up-to-20% initial miss rate the paper cites.
The architecture's transfer from chest X-ray pretraining to pelvic X-rays suggests pretraining on paired radiograph-mask data can generalize across anatomies.

Reading between the lines

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

Editorial inference: The 35-case INVIS set is small enough that the 0.0686 AUC gap over the prior baseline may carry wide confidence intervals; a larger invisible-fracture cohort would tell whether the improvement is stable.
Editorial inference: Because the paper does not ablate the two input streams on INVIS, the cleanest test of the fusion mechanism is a single-stream raw-X-ray ablation; if that also reaches 0.8688 AUC, the segmentation stream is not the active ingredient for invisible fractures.
Editorial inference: The same train-on-visible, test-on-CT-confirmed-invisible protocol could transfer to other skeletal sites where radiographs miss fractures, such as the scaphoid or femoral neck, given paired X-ray and CT data.
Editorial inference: If INVIS detection relies on non-local alignment cues such as pubic symphysis widening, the method could complement, not replace, radiologist reading by flagging suspicious cases for CT.

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.

Desk Editor's Note

read the letter

Quick take: this is a credible application of known components to a genuinely hard clinical problem, and the reported INVIS improvement over the authors' prior DRR20 baseline is worth a careful look. But the headline result rests on 35 curated images from a private set, and the paper never specifies whether the segmentation masks given to PelFANet at test time come from a model that saw those images during segmentation training. If the 2-fold segmentation CV and the reshuffled 5-fold classification CV are not nested, the second input leaks test-image information and the 0.8688 vs 0.8002 gap could be an artifact. That is a load-bearing process question, not a stylistic one.

What is genuinely new: the FABlock fusion design, applied specifically to invisible fractures, and the finding that an anatomy-aware dual-stream network trained only on visible cases transfers to CT-confirmed invisible cases at all. The architecture and training are described clearly enough to reproduce the setup, and the segmentation scores (IoU 90.28, Dice 92.78) are solid. The paper is also honest about what is missing: no ablations of FABlock or CBAM, no from-scratch training, no code or data release.

Soft spots, in proportion: first, the variance values are implausibly low. For 228 images across 5 folds, an accuracy variance of 0.11% means fold accuracies are within a few tenths of a percentage point of each other. With random splits that almost never happens. Either they are reporting standard error, or the folds are highly stratified, or there is an error in the computation. The INVIS variance on 35 images is even more suspicious. Second, the comparison in Table 2 comes entirely from the authors' prior paper on the same private dataset; there is no independent replication. Third, and most important, the leakage risk from the segmentation/classification split alignment is real and unaddressed. If the mask for a test-fold image was generated by a U-Net that trained on that same image, the classifier gets image-specific anatomical information that a deployment-time model would not provide.

Who this is for: someone working in medical imaging, especially fracture detection, will want to know about this result and the leakage risk. It is not a methods breakthrough and the empirical support is thin, but the clinical question is relevant and the experimental setup is falsifiable. If the fold-nesting question is resolved and the variance numbers explained, the INVIS result becomes interesting. As is, I would not rely on the quantitative claim.

Recommendation: send it to peer review, but with the explicit requirement that the authors clarify how the segmentation folds align with the classification folds, report per-fold results, and ideally release code or masks or run a leak-free nested evaluation. The paper deserves referee time because the question matters and the flaw is addressable, not because the current evidence is convincing.

Referee Report

4 major / 5 minor

Summary. PelFANet is a dual-stream convolutional network for pelvic fracture classification that processes a raw pelvic X-ray together with a bone-segmentation mask. The two streams are fused through stacked Fused Attention Blocks (FABlocks) using CBAM attention. The model is pretrained on the COVID-QU-Ex chest X-ray dataset, fine-tuned on a private 228-image VIS subset of the AMERI pelvis dataset, and evaluated both on VIS and on a 35-image INVIS subset of CT-confirmed fractures that are not visible in X-rays. The authors report VIS accuracy 88.68% and AUC 0.9334, and INVIS accuracy 82.29% and AUC 0.8688, outperforming a prior DRR20 baseline (INVIS AUC 0.8002). The central claim is that anatomy-aware dual-stream fusion enables detection of fractures that are invisible in the training distribution.

Significance. The invisible-fracture evaluation is clinically motivated and, if the reported result is trustworthy, would be a meaningful step for computer-aided pelvic radiograph diagnosis. The architecture is clearly described and the idea of training only on visible fractures and testing on CT-confirmed invisible fractures is a strong, falsifiable benchmark. However, the current evidence is not sufficient: the main metric table contains internally impossible numbers, the fold-alignment between the segmentation and classification stages is unspecified, no ablations support the attribution of the gain to the proposed FABlock/CBAM fusion, and the comparison relies on small samples and baselines from the authors' own prior work without significance testing. These issues must be resolved before the central claim can be accepted.

major comments (4)

[Sections 3.2/3.3] The paper does not state whether the bone mask supplied to PelFANet for a test-fold image was produced by a segmentation model that saw that image during segmentation training. The segmentation U-Net is trained with 2-fold cross-validation on the same 228 VIS images, while PelFANet is evaluated with a reshuffled 5-fold cross-validation. If the mask for a VIS test-fold image comes from a segmentation model trained on that exact image, the classifier receives test-time information through the second input, inflating VIS metrics. The INVIS subset is not part of segmentation training, so this particular leak would not directly affect the INVIS transfer numbers, but it undermines the VIS results and the integrity of the pipeline. Please specify how masks were generated (out-of-fold vs. retrained-on-all-data) and, preferably, use out-of-fold masks for all classifier evaluation.
[Table 1] Several reported values cannot be produced by the described 5-fold cross-validation protocol. With 228 VIS images, accuracy must be an integer multiple of 1/228 (≈0.439%); 88.68% would correspond to 202.2 of 228 correct. With 35 INVIS images, accuracy must be an integer multiple of 1/35 (≈2.86%); 82.29% would correspond to 28.8 of 35 correct. The reported INVIS variance of 0.01% is also impossible because a single-image error changes accuracy by 2.86 percentage points. In addition, the F1 values in Table 1 are not consistent with the reported precision and recall under the standard binary F1 formula (e.g., VIS precision 92.49% and recall 92.21% give F1 ≈ 92.35%, not 84.71%). These inconsistencies suggest either an averaging/rounding convention that is not described or an error in the computation; the authors must clarify the protocol and provide a corrected table.
[Section 4.3 (and Section 4.2)] The central attribution claim is that the dual-stream anatomy-aware fusion with FABlocks and CBAM is what enables the improvement, especially on INVIS. No ablation is provided: there is no single-stream variant, no raw-image-only or segmentation-only baseline, no removal of CBAM, no variation of the number of FABlocks, and no replacement of the FABlock with simple concatenation. The paper itself lists these ablations as planned future work. Without them, the gain over the prior ResNet baselines could come from any component of the pipeline (e.g., the two-stage design, pretraining, augmentation, or the specific backbone). This is load-bearing for the claimed contribution and must be addressed.
[Table 2 / Section 4.3] The INVIS evaluation contains only 35 images (23 fractures, 12 normal). The reported AUC gap of 0.8688 vs. 0.8002 is not accompanied by confidence intervals or a significance test; with 35 samples, such a gap may be within sampling variability. The baseline numbers are taken from the authors' own prior paper [21] and are not re-implemented or independently verified. Please provide CIs (e.g., bootstrap or DeLong for AUC), per-fold results, and ideally an external or public dataset validation before claiming a statistically reliable advantage.

minor comments (5)

[Section 2.2 / References] The text says 'The implementation follows the [25] segmentation library', but reference [25] is a COVID-19 imaging paper; the Segmentation Models PyTorch library appears to be reference [26]. Please correct this citation.
[Table 2] The Table 2 entry for ImageNet F1 is written as '80,00%' (comma instead of decimal point); it should be '80.00%'.
[Section 3.3] The sentence about augmenting 'the dataset' to 108,575 images is ambiguous. The number is approximately 80% of COVID-QU-Ex (33,920 × 0.8 × 4), so this likely refers to pretraining data, not the 228-image fine-tuning set. Please clarify the data flow.
[Table 1] The variance notation is inconsistent: accuracy variance is listed as a percentage (e.g., 0.11%), while AUC variance is listed without a percent sign (0.10 and 0.04). State whether these are standard deviations, standard errors, or confidence intervals, and use a consistent notation.
[Table 1 / Section 4.2] The same specificity value (78.33%) is reported for both VIS and INVIS. This may be a coincidence, but it is suspicious and should be checked.

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional or fitted-input circularity; only a minor, non-load-bearing reliance on the authors' prior baseline [21] in the comparison table.

full rationale

PelFANet's central claim is an empirical benchmark, not a derivation. The INVIS generalization result is obtained by evaluating a classifier trained only on the VIS subset against a held-out INVIS subset that is not used in training, so the reported accuracy/AUC is not equal by construction to any fitted value or training objective. The bone-segmentation masks are inputs to the classifier, not predictions derived from the fracture label, and no equation in the paper defines the classification output in terms of the target variable. The only notable self-citation is reference [21], used as the source of the DRR20 baseline numbers and as motivation that invisible fractures are challenging. That citation is not load-bearing for the central architecture claim: the comparison is against previously published numbers, and the generalization result is assessed independently on the INVIS subset. The unspecified alignment between the 2-fold segmentation cross-validation and the reshuffled 5-fold classifier cross-validation is a process-level correctness risk that could inflate metrics, but it is not a circularity in the derivation chain because it is not a definitional or fitted-input substitution. Consequently, no circular step meeting the required evidence standard is present; the minor self-citation warrants a low score rather than zero.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No invented physical entities; the ledger records the parameters and domain assumptions the reported performance depends on. The most consequential assumption is that segmentation masks are generated without test-fold leakage, because the segmentation and classification cross-validation schemes are both defined on the same 228-image dataset and never explicitly aligned.

free parameters (3)

Number of FABlocks = 8
Architecture depth chosen by hand; no ablation varies 8, 4, 12, or tests without fusion, so the claimed benefit of iterative fusion is not isolated.
Per-class augmentation multipliers on VIS = 2x fracture, 6x normal
Chosen to address class imbalance (168 vs 60) and changes the effective class prior the classifier sees; no sensitivity analysis is reported.
Fine-tuning epochs / batch size / learning rate = 30 epochs, batch size 8, LR 1e-4 (SGD)
Standard hyperparameters selected by hand; the reported fold variance could depend on them, but no search or sensitivity is reported.

assumptions (5)

domain assumption Invisible fractures confirmed by 3D-CT share detectable radiographic cues with visible fractures in pelvic X-rays.
Section 1 and Section 4.2 invoke transfer from VIS training to INVIS testing; if this assumption is false, the INVIS result cannot be a generalization effect.
domain assumption The bone segmentation masks used as PelFANet's second input are generated without the segmentation model having seen the test-fold images.
Section 3.2 describes 2-fold CV segmentation on the same 228-image AMERI set, then Section 3.3 describes 5-fold CV classification on the same set; the paper never states how the two fold schemes are aligned.
domain assumption Pretraining on chest X-rays (COVID-QU-Ex) with paired lung masks transfers to pelvic X-ray fracture classification.
Section 3.3 chooses COVID-QU-Ex because it provides paired masks; this is a strong domain-transfer assumption with no from-scratch comparison (stated as future work).
domain assumption The AMERI VIS/INVIS labels are correct and the curated subsets are representative.
Section 3.1 relies on radiologist confirmation for VIS and 3D-CT confirmation for INVIS; no inter-observer or label-quality analysis is provided.
standard math CBAM (channel and spatial attention) improves feature selection as reported in [22].
PelFANet's FABlock (Eq. 3-4) depends on CBAM's effectiveness; the paper does not re-derive or ablate it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Invisible Yet Detected: PelFANet with Attention-Guided Anatomical Fusion for Pelvic Fracture Diagnosis." pith.science (2026). https://pith.science/paper/SVNUW236

@misc{pith2026250913873,
  author       = {Pith},
  title        = {Pith review of: Invisible Yet Detected: PelFANet with Attention-Guided Anatomical Fusion for Pelvic Fracture Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SVNUW236}},
  note         = {Machine review of arXiv:2509.13873}
}

read the original abstract

Pelvic fractures pose significant diagnostic challenges, particularly in cases where fracture signs are subtle or invisible on standard radiographs. To address this, we introduce PelFANet, a dual-stream attention network that fuses raw pelvic X-rays with segmented bone images to improve fracture classification. The network employs Fused Attention Blocks (FABlocks) to iteratively exchange and refine features from both inputs, capturing global context and localized anatomical detail. Trained in a two-stage pipeline with a segmentation-guided approach, PelFANet demonstrates superior performance over conventional methods. On the AMERI dataset, it achieves 88.68% accuracy and 0.9334 AUC on visible fractures, while generalizing effectively to invisible fracture cases with 82.29% accuracy and 0.8688 AUC, despite not being trained on them. These results highlight the clinical potential of anatomy-aware dual-input architectures for robust fracture detection, especially in scenarios with subtle radiographic presentations.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 1 linked inside Pith

[21]

Rahman, R., Yagi, N., Hayashi, K., Maruo, A., Muratsu, H., Kobashi, S.: Enhancing fracture diagnosis in pelvic X -rays by deep convolutional neural network with synthesized images from 3D-CT. Sci. Rep. 14(1), 8004 (2024)

2024
[1]

Bone Joint Surg

Tile, F.: Pelvic ring fractures: Should they be fixed? J. Bone Joint Surg. Br. 70(1), 1 –12 (1988)

1988
[2]

Pape, T., Pohlemann, H., Gänsslen, M.: Mortality associated with pelvic fractures: Evidence from the German Trauma Registry. J. Trauma 53(6), 1079–1086 (2002)

2002
[3]

Eastridge, B.M., Starr, J.E., Minei, L., O’Keefe, M., Holcomb, R.B.: Death within the first hour of trauma: Analysis of early mortality. J. Trauma 63(3), 463–468 (2007)

2007
[4]

Flint, M., Cryer, T., Russell, M., Rodriguez, J., Smith, D.: High mortality in elderly trauma patients with pelvic fractures. J. Trauma 69(4), 1034–1040 (2010)

2010
[5]

Sarwar, A., Rehman, C.H., Yousaf, M.M.: Missed injuries in pelvic trauma: Frequency and clinical impact. J. Trauma 71(5), 1213–1219 (2011)

2011
[6]

Harris, G.R., Gray, T.S., MacKenzie, L.A.: Limitations of radiographs in detecting pelvic fractures. Emerg. Radiol. 13(3), 130–134 (2006)

2006
[7]

Blackmore, J.A., Mann, C.J., Moore, E.R.: Pelvic fractures missed by emergency physi- cians. Ann. Emerg. Med. 38(6), 678–685 (2001). 10 Bhuiyan et al

2001

Show all 31 references

[8]

Injury 44(12), 1914–1918 (2013)

Lee, S.H., Lee, H.J., Kim, M.J.: Delayed diagnosis of pelvic fractures in trauma: Clinical consequences. Injury 44(12), 1914–1918 (2013)

1914
[9]

Kassem, M.A., Naguib, S.M., Hamza, H.M., Fouda, M.M., Saleh, M.K., Hosny, K.M.: Ex- plainable Transfer Learning-Based Deep Learning Model for Pelvis Fracture Detection. Int. J. Intell. Syst. 2023(1), 3281998 (2023)

2023
[10]

Tanzi, L., Vezzetti, E., Moreno, R., Aprato, A., Audisio, A., Massè, A.: Hierarchical fracture classification of proximal femur X-Ray images using a multistage Deep Learning approach. Eur. J. Radiol. 133, 109373 (2020)

2020
[11]

et al.: Application of a deep learning algorithm for detection and visualization of hip fractures on plain pelvic radiographs

Cheng, C.T. et al.: Application of a deep learning algorithm for detection and visualization of hip fractures on plain pelvic radiographs. Eur. Radiol. 29, 5469–5477 (2019)

2019
[12]

Biocybern

Mahapatra, A., Sethi, S., Paul, R.: A modified U-Net for segmentation and classification of colorectal cancer. Biocybern. Biomed. Eng. 40(3), 1296–1310 (2020)

2020
[13]

Xie, W., Wang, J., Xu, Y., Zhang, L., Wang, H.: Liver cancer classification using virtual staining and segmentation. IEEE J. Biomed. Health Inform. 25(8), 2938–2949 (2021)

2021
[14]

In: Proc

Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: UNet++: A nested U -Net ar- chitecture for medical image segmentation. In: Proc. DLMIA (2018)

2018
[15]

arXiv preprint arXiv:2105.05537 (2021)

Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin -Unet: Unet- like pure transformer for medical image segmentation. arXiv preprint arXiv:2105.05537 (2021)

2021 arXiv
[16]

In: MICCAI 2024, LNCS, vol

Jaus, A., Seibold, C., Reiß, S., Heine, L., Schily, A., Kim, M., Bahnsen, F.H., Herrmann, K., Stiefelhagen, R., Kleesiek, J.: Anatomy-guided Pathology Segmentation. In: MICCAI 2024, LNCS, vol. 15008, pp. 3–13. Springer, Switzerland (2024)

2024
[17]

IEEE Access 10, 13476–13488 (2022)

Lee, H., Kwon, S., Yoon, Y., Ham, B.: Pelvic X-ray segmentation using attention and Swin U-Nets. IEEE Access 10, 13476–13488 (2022)

2022
[18]

Lee, S.H., Lee, H.J., Cho, S.Y., Kim, H.S.: Segmentation-guided pelvic ring fracture classi- fication using deep learning. Comput. Biol. Med. 144, 105360 (2022)

2022
[19]

-T., Ho, H.C., Lee, C.H., Chang, Y.T., Chou, Y.C., Chen, W.C., Lee, Y

Cheng, C. -T., Ho, H.C., Lee, C.H., Chang, Y.T., Chou, Y.C., Chen, W.C., Lee, Y. -H., Chang, Y.-H., Hsu, Y.-C.: A scalable physician -level deep learning algorithm detects uni- versal trauma on pelvic radiographs. Nat. Commun. 12, 1066 (2021)

2021
[20]

In: Computer Vision – ECCV 2020

Chen, H., Xie, J., Cao, Z., Niu, Y., Zhang, L., Li, S., Yu, J., Li, S.: Anatomy-aware Siamese network: Exploiting semantic asymmetry for accurate pelvic fracture detection in X-ray im- ages. In: Computer Vision – ECCV 2020. LNCS, vol. 12367, pp. 341–357. Springer (2020)

2020
[22]

In: European Conference on Computer Vision (ECCV), pp

Woo, S., Park, J., Lee, J.Y., Kweon, I.S.: Cbam: Convolutional block attention module. In: European Conference on Computer Vision (ECCV), pp. 3–19 (2018)

2018
[23]

In: Medical Image Computing and Computer -Assisted Intervention – MICCAI 2015

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer -Assisted Intervention – MICCAI 2015. LNCS, vol. 9351, pp. 234–241. Springer (2015)

2015
[24]

Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: SegFormer: Simple and efficient design for semantic segmentation with transformers. Adv. Neural Inf. Process. Syst. 34, 12077–12090 (2021)

2021
[25]

Tahir, A.M., Chowdhury, M.E.H., Khandakar, A., Qiblawey, Y., Khurshid, U., Kiranyaz, S., Ibtehaz, N., Rahman, M.S., Al-Madeed, S., Mahmud, S., Ezeddin, M., Hameed, K., Ha- mid, T.: COVID-19 Infection Localization and Severity Grading from Chest X-ray Images. Comput. Biol. Med....

2021
[26]

GitHub repository, https://github.com/qub- vel/segmentation_models.pytorch, last accessed 2025/06/01 PelFANet 11

Iakubovskii, P.: Segmentation Models Pytorch. GitHub repository, https://github.com/qub- vel/segmentation_models.pytorch, last accessed 2025/06/01 PelFANet 11

2025
[27]

Kaggle (2021)

Tahir, A.M., Chowdhury, M.E.H., Qiblawey, Y., Khandakar, A., Rahman, T., Kiranyaz, S., Khurshid, U., Ibtehaz, N., Mahmud, S., Ezeddin, M.: COVID-QU-Ex. Kaggle (2021)

2021
[28]

Com- put

Rahman, T., Khandakar, A., Qiblawey, Y., Tahir, A., Kiranyaz, S., Kashem, S.A., Islam, M., Al Maadeed, S., Zughaier, S., Khan, M., Chowdhury, M.E.H.: Exploring the Effect of Image Enhancement Techniques on COVID-19 Detection using Chest X-rays Images. Com- put. Biol. Med. 1043...

2021
[29]

Health Inf

Degerli, A., Ahishali, M., Yamac, M., Kiranyaz, S., Chowdhury, M.E.H., Hameed, K., Ha- mid, T., Mazhar, R., Gabbouj, M.: Covid -19 infection map generation and detection from chest X-ray images. Health Inf. Sci. Syst. 9, 15 (2021)

2021
[30]

Chowdhury, M.E.H., Rahman, T., Khandakar, A., Mazhar, R., Kadir, M.A., Mahbub, Z.B., Islam, K.R., Khan, M.S., Iqbal, A., Emadi, N.A., Reaz, M.B.I., Islam, M.T.: Can AI Help in Screening Viral and COVID-19 Pneumonia? IEEE Access 8, 132665-132676 (2020)

2020
[31]

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

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

2016

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [21]

Rahman, R., Yagi, N., Hayashi, K., Maruo, A., Muratsu, H., Kobashi, S.: Enhancing fracture diagnosis in pelvic X -rays by deep convolutional neural network with synthesized images from 3D-CT. Sci. Rep. 14(1), 8004 (2024)

2024

[2] [1]

Bone Joint Surg

Tile, F.: Pelvic ring fractures: Should they be fixed? J. Bone Joint Surg. Br. 70(1), 1 –12 (1988)

1988

[3] [2]

Pape, T., Pohlemann, H., Gänsslen, M.: Mortality associated with pelvic fractures: Evidence from the German Trauma Registry. J. Trauma 53(6), 1079–1086 (2002)

2002

[4] [3]

Eastridge, B.M., Starr, J.E., Minei, L., O’Keefe, M., Holcomb, R.B.: Death within the first hour of trauma: Analysis of early mortality. J. Trauma 63(3), 463–468 (2007)

2007

[5] [4]

Flint, M., Cryer, T., Russell, M., Rodriguez, J., Smith, D.: High mortality in elderly trauma patients with pelvic fractures. J. Trauma 69(4), 1034–1040 (2010)

2010

[6] [5]

Sarwar, A., Rehman, C.H., Yousaf, M.M.: Missed injuries in pelvic trauma: Frequency and clinical impact. J. Trauma 71(5), 1213–1219 (2011)

2011

[7] [6]

Harris, G.R., Gray, T.S., MacKenzie, L.A.: Limitations of radiographs in detecting pelvic fractures. Emerg. Radiol. 13(3), 130–134 (2006)

2006

[8] [7]

Blackmore, J.A., Mann, C.J., Moore, E.R.: Pelvic fractures missed by emergency physi- cians. Ann. Emerg. Med. 38(6), 678–685 (2001). 10 Bhuiyan et al

2001