CoChat
Sign Up
Cited by 40Medical Image Segmentation TechniquesAdvanced Neural Network ApplicationsArtificial Intelligence in Healthcare and Education

Fully Automated 3-D Ultrasound Segmentation of the Placenta, Amniotic Fluid, and Fetus for Early Pregnancy Assessment

Open access
Looney, Padraig, Yin, Yi, Collins, Sally L., Nicolaides, Kypros H., Plasencia, Walter, Molloholli, Malid, Natsis, Stavros, Stevenson, Gordon N.·2021·IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control
DOI 10.1109/tuffc.2021.3052143PMC PMC8154733
Open this paper in CoChat — free account

Abstract

Volumetric placental measurement using 3-D ultrasound has proven clinical utility in predicting adverse pregnancy outcomes. However, this metric cannot currently be employed as part of a screening test due to a lack of robust and real-time segmentation tools. We present a multiclass (MC) convolutional neural network (CNN) developed to segment the placenta, amniotic fluid, and fetus. The ground-truth data set consisted of 2093 labeled placental volumes augmented by 300 volumes with placenta, amniotic fluid, and fetus annotated. A two-pathway, hybrid (HB) model using transfer learning, a modified loss function, and exponential average weighting was developed and demonstrated the best performance for placental segmentation (PS), achieving a Dice similarity coefficient (DSC) of 0.84- and 0.38-mm average Hausdorff distances (HDAV). The use of a dual-pathway architecture improved the PS by 0.03 DSC and reduced HDAV by 0.27 mm compared with a naïve MC model. The incorporation of exponential weighting produced a further small improvement in DSC by 0.01 and a reduction of HDAV by 0.44 mm. Per volume inference using the FCNN took 7-8 s. This method should enable clinically relevant morphometric measurements (such as volume and total surface area) to be automatically generated for the placenta, amniotic fluid, and fetus. The ready availability of such metrics makes a population-based screening test for adverse pregnancy outcomes possible.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper presents a fully automated multi-class 3D ultrasound segmentation framework for the placenta, amniotic fluid, and fetus in early pregnancy, using a hybrid dual-pathway FCNN with transfer learning, a modified loss function, and exponential averaging. It reports state-of-the-art placental segmentation and clinically relevant inference times, with the goal of enabling population-based first-trimester screening.

Key findings

  • The hybrid model with exponential averaging (HBEA) achieved the best placental segmentation on the test set: mean DSC 0.82 (0.08) and HDAV 0.58 (0.70) mm; amniotic fluid DSC was 0.93 (0.04) and fetus DSC was 0.88 (0.04). The abstract reports placental DSC 0.84 and HDAV 0.38 mm.
  • A dual-pathway architecture improved placental segmentation by 0.03 DSC and reduced HDAV by 0.27 mm compared with a naïve multi-class model; exponential weighting gave a further DSC improvement of 0.01 and HDAV reduction of 0.44 mm.
  • Inference took 7–8 seconds per volume, and the HBEA model had a median segmentation time of 8.46 seconds versus 30 minutes for semi-automated segmentation.
  • In a blinded comparative Turing test, positive recognition rates for the Random Walker ground truth were MC 52.5%, MCTL 44.0%, HB 56.0%, and HBEA 45.8%; ICC repeatability was excellent for amniotic fluid and fetus and lower for placenta, with HBEA best among models for placenta.

Methods

  • Data came from 2,393 3D-US volumes of singleton pregnancies at 11+0 to 13+6 weeks at a UK tertiary referral hospital. The placental segmentation model used 2,093 labelled volumes (1,893 training, 150 validation, 50 test); 300 volumes were randomly selected for multi-class annotation (200 training, 40 validation, 60 test).
  • Ground truth was generated using the semi-automated Random Walker algorithm with expert seeding and quality control. Models included a 3D U-Net-like FCNN for placenta (PS/PSEA), multi-class models (MC, MCTL), and hybrid dual-pathway models (HB, HBEA) combining a fixed placental pathway with a trainable pathway for amniotic fluid and fetus.
  • Evaluation used DSC, Hausdorff distance, average Hausdorff distance, ICC, paired t-tests, and a blinded side-by-side Turing test. Post-processing included morphological filtering, largest-region retention, grayscale closing, and hole filling.

Limitations

  • The paper states that the 3D-US data were collected years ago using an ultrasound machine superseded by newer hardware; future image quality may improve segmentation, and ultrasound reconstruction methods such as spatial averaging may alter image texture and affect FCNN performance.
  • The paper states that model parameters were not investigated with full ablation studies, nor were other post-processing strategies fully evaluated.
  • The text makes evident that the multi-class dataset was relatively small (300 volumes, 60 test cases), and the paper notes that Random Walker ground truth has been shown equivalent to manual segmentation for the placenta but not for the fetus or amniotic fluid.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/tuffc.2021.3052143

Cites

Statistics

Nothing to recompute.

Claims the paper makes

0 of 38 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The PS model obtained the best mean DSC on the validation set of 0.85 (0.09) after 17,000 training steps.

    “The PS model obtained the best mean DSC (std. dev) on the validation set of 0.85 (0.09) after 17,000 training steps.”
    Quote in the paperSubstantial claim
  • Findingnone given

    Per-volume inference using the FCNN took 7-8 seconds.

    “Per volume inference using the FCNN took 7–8 seconds.”
    Quote in the paperSubstantial claim
  • Findingown data

    Repeatability measured by intra-class correlation coefficients showed good to excellent repeatability for the respective organs.

    “Repeatability measured by intra-class correlation coefficients (ICC), showed good to excellent repeatability (measured by ICC) for respective organs.”
    Quote in the paperSubstantial claim
  • Findingown data

    A two-pathway hybrid model using transfer learning, a modified loss function and exponential average weighting achieved the best performance for placental segmentation, with a DSC of 0.84 and an average Hausdorff distance of 0.38 mm.

    “A two pathway, hybrid model using transfer learning, a modified loss function and exponential average weighting was developed and demonstrated the best performance for placental segmentation, achieving a Dice similarity coefficient (DSC) of 0.84 and 0.38 mm average Hausdorff distance (HDAV).”
    Quote in the paperSubstantial claim
  • Findingown data

    The use of multiple pathway FCNNs trained on single and multi-class datasets, with a modified loss function, exponential averaging and transfer learning, improved average DSC to segment the placenta, amniotic fluid and fetus.

    “The use of multiple pathway FCNNs trained on single and multi-class datasets, which by addition of a modified loss function, exponential averaging (EA) and transfer learning improved average DSC to segment the placenta, amniotic fluid and fetus.”
    Quote in the paperSubstantial claim
  • May combine two claims. 29
  • Classified as probably not a substantial claim. 23
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingown data

    When trained with 50 training data over 40 epochs, the HBEA model achieved placental DSC 0.73 (0.1), amniotic fluid DSC 0.90 (0.06), and fetus DSC 0.83 (0.08); HD was 22.48 (8.78) mm for placenta, 14.20 (9.25) mm for amniotic fluid, and 21.74 (11.79) mm for fetus; HDAV was 1.23 (1.16) mm for placenta, 0.36 (0.84) mm for amniotic fluid, and 21.74 (11.79) mm for fetus.

    “Placental DSC was 0.73 (0.1), amniotic fluid DSC was 0.90 (0.06) and fetus DSC was 0.83 (0.08) for the HBEA model. For HD (mm), the placenta was 22.48 (8.78), amniotic fluid was 14.20 (9.25) and fetus was 21.74 (11.79). For HDAV (mm), the placenta was 1.23 (1.16), amniotic fluid was 0.36 (0.84) and fetus was 21.74 (11.79).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The 50 training cases used for comparison to Yang et al. are comparable to the dataset used by Yang et al.

  • May combine two claims.

    Findingreasoning

    Combining pathways from different models combined previous state-of-the-art placental segmentation with amniotic fluid and fetal segmentation comparable to the state of the art as measured by DSC and Hausdorff distance.

    “By combining pathways from different models and using these features, our previous state of the art segmentation performance on the placenta was combined with segmentation of the amniotic fluid and fetus that was comparable to the state of the art as measured by DSC and Hausdorff distance (HD)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Use of a dual-pathway architecture improved placental segmentation by 0.03 DSC and reduced HDAV by 0.27 mm compared with a naïve multi-class model.

    “Use of a dual-pathway architecture, improved placental segmentation by 0.03 DSC and reduced HDAV by 0.27mm when compared with a naïve multi-class model.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On the publicly available data published by [21], the PS model achieved a mean DSC of 0.67 (0.24), mean HD of 21.28 (14.18) mm, and mean HDAV of 1.59 (2.27) mm.

    “The mean DSC (std. dev) on this data set was 0.67 (0.24), mean HD was 21.28 (14.18) mm and mean HDAV was 1.59 (2.27) mm.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [21] Yushkevich PA, Pashchinskiy A, Oguz I, Mohan S, Schmitt JE, Stein JM, “ User-guided segmentation of multi-modality medical imaging datasets with ITK-SNAP,” Neuroinformatics, vol. 17, no. 1, pp. 83– 102, 2019. 29946897 PMC6310114

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Comparing the same HB model using the modified loss function to standard cross entropy loss, DSC for placenta increased by 0.01 and remained the same for fetus and amniotic fluid; HD increased for placenta by 0.44 mm, decreased for amniotic fluid by 0.01 mm, and increased for fetus by 1.4 mm; HDAV decreased for placenta by 0.04 mm, decreased for amniotic fluid by 0.01 mm, and was unchanged for the

    “Comparing the same HB model using the modified loss function defined in Eq. 1 to standard cross entropy loss, similarity metrics improved. DSC for placenta increased by 0.01 and remained the same for the fetus and amniotic fluid. For surface similarity metrics, small increases in HD (mm) (placenta +0.44; amniotic fluid −0.01; fetus 1.4) and decreases in HDAV (placenta −0.04mm; amniotic fluid −0.01”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Incorporation of exponential weighting produced a further improvement in DSC by 0.01 and a reduction in HDAV by 0.44 mm.

    “Incorporation of exponential weighting produced a further small improvement in DSC by 0.01 and a reduction of HDAV by 0.44mm.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On the 50 test cases, the PS model had a mean DSC of 0.85 (0.05) and the PSEA model had a mean DSC of 0.85 (0.05).

    “On the 50 test cases, the PS model had a mean DSC (std. dev) of 0.85 (0.05) and the PSEA model had a mean DSC of 0.85 (0.05)).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    A comparative Turing test indicated that automated segmentation was highly comparable to human performance, with less than 50% positive prediction rate, meaning the automated segmentation was selected blindly as better quality more often than the human.

    “An evaluation technique to test real-world performance using a comparative Turing test indicated that automated segmentation was highly comparable to human performance (<50% positive prediction rate; i.e. the automated segmentation was selected blindly as better quality more often than the human.).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    ICC values for the placenta were lower than those for the fetus and amniotic fluid, which showed excellent reproducibility.

    “ICC values for the placenta were lower than for those for the fetus and amniotic fluid which reported excellent reproducability.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The positive recognition rates for the Random Walker based ground-truth data compared to the four models were MC 52.5%, MCTL 44.0%, HB 56.0%, and HBEA 45.8%.

    “The positive recognition rates for the Random Walker based ground-truth data when compared to the four different models were: MC 52.5%, MCTL 44.0%, HB 56.0% and HBEA 45.8%.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    The MC and MCTL models have identical architectures, but the MC model's placenta segmentation performance is reduced compared to the MCTL model.

    “The MC and MCTL models have identical architectures but the performance of the MC model in segmenting the placenta is reduced in comparison to the MCTL model”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For the placenta, ICC values increased, with the HBEA model providing the best performance at a moderate level of reported ICC.

    “For the placenta, ICC values increased with the HBEA model providing best performance at a moderate level of reported ICC.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The HBEA model had a median segmentation time of 8.46 seconds, compared to 30 minutes for semi-automated segmentation.

    “The HBEA model had a median segmentation time of 8.46 seconds compared to the 30 minutes required for semi-automated segmenation.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    A semi-automated placental MRI segmentation technique combining multiple volumes and a single annotated slice with learned random forest and random field features and a 4D graph cut achieved a mean DSC of 0.89 ± 0.02 in 16 cases.

    “A semi-automated technique combining multiple volumes and a single annotated slice, with learned random forest and random field features combined with a 4D graph cut, demonstrated a mean Dice similarity coefficient (DSC) in 16 cases of 0.89 ± 0.02 (std. dev) [18].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [18] Wang G, Zuluaga MA, Pratt R, Aertsen M, Doel T, Klusmann M, “ Slic-seg: A minimally interactive segmentation of the placenta from sparse and motion-corrupted fetal mri in multiple views,” Medical image analysis, vol. 34, pp. 137– 147, 2016. 27179367 PMC5052128 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Volume estimation currently requires off-line manual annotation by a trained sonographer, which is time-consuming and cannot be performed within the 15 minutes of a standard scan.

    “However, volume estimation currently requires off-line manual annotation by a trained sonographer which is time-consuming and cannot be performed within the 15 minutes in which a standard scan takes place [6].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [6] Farina A, “ Systematic review on first trimester three-dimensional placental volumetry predicting small for gestational age infants,” Prenatal diagnosis, vol. 36, no. 2, pp. 135– 141, 2016. 26618611

      Not checked in the free run.

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    Placental insufficiency is the most common cause of stillbirth and is associated with other adverse pregnancy outcomes such as fetal growth restriction and pre-eclampsia.

    “Placental insufficiency is the most common cause of stillbirth [1] as well as other adverse pregnancy outcomes such as fetal growth restriction [2] and pre-eclampsia [3].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [1] Seaton SE, Field DJ, Draper ES, Manktelow BN, Smith GC, Springett A, “ Socioeconomic inequalities in the rate of stillbirths by cause: a population-based study,” BMJ open, vol. 2, no. 3, p. e001100, 2012. PMC3383980 22735165 source

      Not checked in the free run.

    • cites [2] Sankaran S and Kyle PM, “ Aetiology and pathogenesis of IUGR,” Best Practice & Research Clinical Obstetrics & Gynaecology, vol. 23, no. 6, pp. 765– 777, 2009. 19666240 source

      Not checked in the free run.

    • cites [3] Redman CW and Sargent IL, “ Latest advances in understanding preeclampsia,” Science, vol. 308, no. 5728, pp. 1592– 1594, 2005. 15947178 source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    A systematic literature review concluded that placental volume measured by 3D ultrasound could have value when integrated into a multivariable screening method for fetal growth restriction in the first trimester.

    “A systematic literature review concluded that placental volume, measured by 3D ultrasound (3D-US), could have value when integrated into a multivariable screening method for fetal growth restriction in the first trimester [6].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [6] Farina A, “ Systematic review on first trimester three-dimensional placental volumetry predicting small for gestational age infants,” Prenatal diagnosis, vol. 36, no. 2, pp. 135– 141, 2016. 26618611

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    An FCNN method applied to placental MRI data using a V-Net architecture in 12 patients provided a mean DSC of 0.75 ± 0.11.

    “An FCNN method applied to placental MRI data using a V-Net architecture in 12 patients provided a mean DSC of 0.75 ± 0.11 [19].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [19] Torrents-Barrena J, “ Lstm fully convolutional neural networks for umbilical cord segmentation in ttts foetal surgery planning,” in Proc. 32nd International Conference on Computer Assisted Radiology and Surgery, 2018.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Posterior placental implantation presents a significant segmentation challenge because the overlying fetus can attenuate and scatter the US signal, causing shadows and image artefact.

    “Placentas that implant on the posterior uterine wall present a significant segmentation challenge as the overlying fetus can attenuate and scatter the US signal causing shadows and image artefact.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [17] Torrents-Barrena J, “ Segmentation and classification in MRI and US fetal imaging: Recent trends and future prospects,” Medical Image Analysis, vol. 51, pp. 61– 88, 2019. 30390513 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    FCNNs are particularly suited to segmentation because their parameters are shared and independent of image size.

    “FCNNs are particularly suited to segmentation since the parameters are shared and independent of image size.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The placenta has a heterogeneous appearance and can implant on any of the uterine walls.

    “The placenta has a heterogeneous appearance and can implant on any of the uterine walls.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [17] Torrents-Barrena J, “ Segmentation and classification in MRI and US fetal imaging: Recent trends and future prospects,” Medical Image Analysis, vol. 51, pp. 61– 88, 2019. 30390513 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Manual labelling is highly operator dependent, with inter-observer ICCs of 0.59 (0.33-0.80) and 0.81 (0.68-0.91) reported in reproducibility studies.

    “Furthermore, manual labelling is highly operator dependent, inter-observer reproducability studies have demonstrated very different intra-class correlation coefficients (ICC; 95% CI)) of 0.59 (0.33–0.80) [7] and 0.81 (0.68–0.91) [8].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [7] Jones NW, Raine-Fenning NJ, Mousa HA, Bradley E, and Bugg GJ, “ Evaluating the intra- and interobserver reliability of three-dimensional ultrasound and power Doppler angiography (3D-PDA) for assessment of placental volume and vascularity in the second trimester of pregnancy,” UMB, vol. 37, no. 3, 2011. 21256663 source

      Not checked in the free run.

    • cites [8] Larsen M, Naver K, Kjaer M, Jorgensen F, and Nilas L, “ Reproducibility of 3-dimensional ultrasound measurements of placental volume at gestational ages 11–14 weeks,” Facts, Views & Vision in ObGyn, vol. 7, no. 4, p. 203, 2015. PMC5058409 27729965

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Use of multiple pathways to increase context of an FCNN has improved accuracy and has been applied to placental segmentation in MRI and 3D-US.

    “Use of multiple pathways, to increase the context of an FCNN, has improved accuracy [14] and has been successfully applied to placental segmentation in both MRI [15] and 3D-US [16].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [14] Kamnitsas K, Ledig C, Newcombe VF, Simpson JP, Kane AD, Menon DK, “ Efficient multi-scale 3D (CNN) with fully connected (CRF) for accurate brain lesion segmentation,” Medical Image Analysis, vol. 36, pp. 61 – 78, 2017. 27865153 source

      Not checked in the free run.

    • cites [15] Alansary A, Kamnitsas K, Davidson A, Khlebnikov R, Rajchl M, Malamateniou C, “ Fast fully automatic segmentation of the human placenta from motion corrupted MRI,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2016, pp. 589– 597. source

      Not checked in the free run.

    • cites [16] Looney P, Stevenson GN, Nicolaides KH, Plasencia W, Molloholli M, Natsis S, “ Automatic 3D ultrasound segmentation of the first trimester placenta using deep learning,” in Biomedical Imaging, IEEE 14th International Symposium on. IEEE, 2017, pp. 279– 282.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Real-time volume estimation has been achieved using fully convolutional neural networks (FCNNs), producing state-of-the-art performance in a range of medical imaging modalities.

    “Real-time volume estimation has been achieved using fully convolutional neural networks (FCNNs) [11] to produce state-of-the-art performance in a range of medical imaging modalities [12].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [11] Krizhevsky A, Sutskever I, and Hinton GE, “ Imagenet classification with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097– 1105. source

      Not checked in the free run.

    • cites [12] Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, “ A survey on deep learning in medical image analysis,” Medical image analysis, vol. 42, pp. 60– 88, 2017. 28778026 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Segmentation accuracy of FCNNs can be improved by technical enhancements such as loss function modification, which has improved prostate segmentation in MRI.

    “Segmentation accuracy of FCNNs can be further improved with technical enhancements such as loss function modification, which has demonstrated improved performance in segmenting the prostate in MRI images [13].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [13] Milletari F, Navab N, and Ahmadi S-A, “ V-net: Fully convolutional neural networks for volumetric medical image segmentation,” in 3D Vision (3DV), 2016 Fourth International Conference on. IEEE, 2016, pp. 565– 571.

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductcitation

    A blinded side-by-side comparison, akin to a comparative Turing test, was performed to assess an experienced operator's ability to discriminate between manual and automated outputs for the 60 multi-class test cases.

    “Finally, a blinded comparison was performed to assess an experienced operator’s (GS) ability to discriminate between the manual and automated outputs using a side-by-side comparison, akin to a comparative Turing test [39] for the 60 multi-class test cases.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [39] Gooding MJ, “ Comparative evaluation of autocontouring in clinical practice: A practical method using the turing test,” Medical Physics, vol. 45, no. 11, pp. 5105– 5115, 2018. 30229951 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The placenta segmentation model was selected by choosing the highest mean DSC on the validation set that was not improved upon by ten percent within the next five epochs.

    “The placenta segmentation (PS) model was chosen by selecting the highest mean DSC on the validation set that was not improved upon by ten percent within the next five epochs.”
    Quote in the paperNot a substantial claim

  • May combine two claims. The quote only partly states this claim.

    Study design and conductown data

    Placental ground truth segmentation was generated using the semi-automated Random Walker algorithm with expert seeding, independent quality control by a second clinical expert, and adjudication by a third expert for uncertain boundaries.

    “Segmentation of the placenta used as label map was performed using the semi-automated Random Walker algorithm, as described in our previous study [9]. Initialisation or ‘seeding’ of the placental segmentation was performed by a clinical expert (SN). These “seedings” then underwent quality control with each one being examined by a second, independent, clinical expert (MM)”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [9] Stevenson GN, Collins SL, Ding J, Impey L, and Noble JA, “ 3D ultrasound segmentation of the placenta using the Random Walker algorithm: reliability and agreement,” Ultrasound in Medicine & Biology, vol. 41, no. 12, pp. 3182– 3193, 2015. 26341043

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Automated methods of segmenting first-trimester placentas (n = 13) in an anterior position using joint label fusion and majority vote achieved a mean DSC of 0.83 ± 0.05.

    “Using 3D-US, automated methods of segmenting first trimester placentas (n = 13) in an anterior position using a joint label fusion and majority vote technique achieved a mean DSC of 0.83 ± 0.05 [20], [21].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [20] Oguz I, Pouch AM, Yushkevich N, Wang H, Gee JC, Schwartz N, “ Automated placenta segmentation from 3D ultrasound images,” in MICCAI workshop on perinatal, preterm and paediatric image analysis (PIPPI), 2016.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [21] Yushkevich PA, Pashchinskiy A, Oguz I, Mohan S, Schmitt JE, Stein JM, “ User-guided segmentation of multi-modality medical imaging datasets with ITK-SNAP,” Neuroinformatics, vol. 17, no. 1, pp. 83– 102, 2019. 29946897 PMC6310114 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Previously, using 2,393 cases, an FCNN segmented placental volumes to predict small for gestational age babies and obtained a mean DSC of 0.82 ± 0.10 for placental segmentation.

    “Previously using 2,393 cases, an FCNN segmented placental volumes to predict small for gestational age (SGA) babies [23] which obtained a mean DSC of 0.82 ± 0.10 (std. dev) for placental segmentation.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [23] Looney P, Stevenson GN, Nicolaides KH, Plasencia W, Molloholli M, Natsis S, “ Fully automated, real-time 3D ultrasound segmentation to estimate first trimester placental volume using deep learning,” JCI insight, vol. 3, no. 11, 2018. PMC6124436 29875312 source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductcitation

    Reproducibility of final placenta, fetus and amniotic fluid volumes was assessed using ICC(2,1) for inter-observer repeatability between the semi-automated Random Walker results and the newly generated FCNN outputs.

    “Reproducibility of the final placenta, fetus and amniotic fluid volume in millilitres was assessed using the ICC (2,1) [37], [38] for inter-observer repeatability between the semi-automated Random Walker results, regarded as the ground-truth used for performance evaluation, and the newly generated FCNN outputs.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [37] Shrout PE and Fleiss JL, “ Intraclass correlations: uses in assessing rater reliability.” Psychological bulletin, vol. 86, no. 2, p. 420, 1979. 18839484 source

      Not checked in the free run.

    • cites [38] Koo TK and Li MY, “ A guideline of selecting and reporting intraclass correlation coefficients for reliability research,” Journal of Chiropractic Medicine, vol. 15, no. 2, pp. 155– 163, 2016. 27330520 PMC4913118 source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Pairwise comparison between DSC measurements between models was assessed using Student's paired t-test with a significance threshold of P < .05.

    “The significance threshold was set at P < .05. Pairwise comparison between DSC measurements between models were assessed using Student’s paired t-test.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given

    The exponential moving average reduced volatility in the weights and improved model inference.

    “The exponential moving averaged reduced volatility in the weights which we propose improved model inference.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    Yang et al.'s strategy uses four times the number of features compared to the FCNN models used in this work.

    “This strategy uses four times the number of features compared to the FCNN models used in our work.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Differences in implementation and the greater number of training cases used may account for the increased performance in this study, particularly for placenta segmentation (DSC 0.82 vs 0.64).

    “These differences as well as the greater number of training cases used may account for the increased performance in this study, particularly in segmenting the placenta (DSC of 0.82 compared with 0.64).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Small differences between the PS and HB models can occur when background voxels are classified as placenta because their background probability was distributed among other classes such that placental probability was maximum.

    “Small differences between the PS and HB models can occur, when some background voxels can be classified as placenta if their background probability was distributed among the other classes such that the placental probability was maximum.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    The effect of the many parameters within the model has not been investigated with full ablation studies, nor has there been full evaluation of other post-processing strategies.

    “The effect of the many parameters within the model have not been investigated with full ablation studies nor full evaluation of other post-processing strategies.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Yang et al. used a 3D extension of U-Net and combined the output with a recurrent neural network.

    “They also used a 3D extension of U-Net [28] and combined the output with a recurrent neural network.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [28] Ronneberger O, Fischer P, and Brox T, “ U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234– 241. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    The 3D-US data were collected a number of years ago using an ultrasound machine that has been superseded by two newer generations of hardware.

    “The 3D-US data were collected a number of years ago using an US machine that has been superseded by two newer generations of hardware.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The Random-Walker technique has equivalent performance to manual segmentation in terms of both inter- and intra-operator variability.

    “Whilst it has been demonstrated that the much faster Random-Walker technique has equivalent performance to manual segmentation in terms of both inter and intra-operator variability [9]”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [9] Stevenson GN, Collins SL, Ding J, Impey L, and Noble JA, “ 3D ultrasound segmentation of the placenta using the Random Walker algorithm: reliability and agreement,” Ultrasound in Medicine & Biology, vol. 41, no. 12, pp. 3182– 3193, 2015. 26341043

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationreasoning

    These measures, combined with maternal characteristics and serum biomarkers, can now be used to develop a first trimester screening tool aimed at improving identification of pregnancies at-risk of later complications.

    “These measures, combined with maternal characteristics and serum biomarkers, can now be used to develop a first trimester screening tool aimed at improving identification of pregnancies at-risk of later complications.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The model realised in this work will allow rapid calculation of placental volume and other morphometrics such as shape and surface area of the utero-placental interface, since these can now be calculated using the MC segmentation.

    “Hence, the model realised in this work will allow rapid calculation of not only placental volume but other important morphometrics such as shape and surface area of the utero-placental interface since these can now also be calculated using the MC segmentation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The automated deep learning method achieves state-of-the-art performance, measured using DSC, HD, and HDAV, in segmenting the placenta, while obtaining similar values to state-of-the-art performance for the amniotic fluid and fetus.

    “We present an automated method based on deep learning that achieves state-of-the-art performance, measured using DSC, HD, and HDAV in segmenting the placenta while obtaining similar values to the state-of-the-art performance for the amniotic fluid and fetus.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The Random Walker generated ground truth is accurate for segmentation of the fetus and amniotic fluid.
43 of 43

  • The paper relies on this without support.

    The 300 multi-class volumes and 2,093 placental volumes are representative of the target population of singleton pregnancies at 11+0 to 13+6 weeks of gestation.

    “From the available 2,393 3D-US volumes with an existing placental segmentation, 300 volumes were randomly selected for multi-class segmentation.”
    “The remaining 2,093 cases were partitioned into 1,893 training cases, 150 validation cases and 50 test cases.”

  • The paper relies on this without support.

    The 50 training cases used for comparison to Yang et al. are comparable to the dataset used by Yang et al.

    “To compare performance to a reduced dataset size, as per Yang et al, using 50 training data the model was trained over 40 epochs.”

  • The paper relies on this without support.

    The Random Walker generated ground truth is accurate for segmentation of the fetus and amniotic fluid.

    “Whilst it has been demonstrated that the much faster Random-Walker technique has equivalent performance to manual segmentation in terms of both inter and intra-operator variability [9], the same has not been demonstrated for segmentation of the fetus or amniotic fluid.”

  • The paper relies on this without support.

    The PS model is more accurate than the multi-class models for placenta segmentation.

    “reflecting the fixed nature of the parameters of the top pathway which we used to exploit the more accurate PS model that has been trained on the largest set of cases.”

  • The paper relies on this without support.

    The border between the placenta and surrounding tissue often appears very diffuse, making it difficult for even highly experienced sonographers to distinguish the boundary between placenta and the uterine myometrium.

    “In this problem, the border between the placenta and the surrounding tissue often appears very diffuse making it difficult for even highly experienced sonographers to distinguish the boundary between placenta and the uterine myometrium.”
5 of 5

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Abstract and results disagree on the best placental segmentation metrics

    Major

    The abstract reports a placental DSC of 0.84 and HDAV of 0.38 mm for the best model, whereas the results report 0.82 and 0.58 mm for the HBEA model.

  • Discussion counts non-training cases as PS training data

    Minor

    The discussion says the PS model was trained on 2,093 placentas, but the methods assign only 1,893 of those cases to training.

  • Discussion overstates the single-pathway training-set size

    Minor

    The discussion says the single-pathway model was trained on 300 cases, but the methods allocate only 200 of the 300 multi-class cases to training.

3 of 3

The citation lists haven’t been fetched yet

References 38 · Cited by 40

Research works printed but not linked: 9

  1. 2013, R A Language and Environment for Statistical Computing
  2. johnson, 2013, The 692 ITK Software Guide
  3. looney, 0, IEEE Models
  4. looney, 2017, OxNNet
  5. stevenson, 2020, Imageturingtest
  6. torrents-barrena, 2018, LSTM fully convolutional neural networks for umbilical cord segmentation in ttts foetal surgery planning, Proc 32nd Int Conf Comput Assist Radiol Surg, 17s
  7. oguz, 2016, Automated placenta segmentation from 3D ultrasound images, Proc MICCAI Workshop Perinatal Preterm Paediatric Image Anal (PIPPI)
  8. looney, 2017, OxNNet
  9. looney, 2019, Plooney/KRETZ V1 1

Authors
Looney, P., Yin, Y., Collins, S. L., Nicolaides, K. H., Plasencia, W., Molloholli, M., Natsis, S., & Stevenson, G. N.
Affiliations
  • University of Oxford (GB)
  • John Radcliffe Hospital (GB)
  • Harris Birthright Research Centre for Fetal Medicine (GB)
  • Hospiten Group (ES)
  • Wexham Park Hospital (GB)
  • UNSW Sydney (AU)
ORCID
Venue
IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2021-06
Volume
68
Issue
6
Pages
2038-2047
ISSN
  • 0885-3010 (ISSN-L)
  • 1525-8955
DOI
10.1109/tuffc.2021.3052143
PMC
PMC8154733
OpenAlex
W3123759929
Semantic Scholar
780bd58fe679836e508041ed5ecd4d2c2d14ffec
Cited by
40
References
38
Citations by year
2026: 5 · 2025: 8 · 2024: 7 · 2023: 9 · 2022: 9 · 2021: 2
FWCI
2.43
Type
journal article
Language
en
Open access
Licence
Locations
  • IEEE Transactions on Ultrasonics Ferroelectrics and Frequency Control (publishedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
  • UNSWorks (University of New South Wales, Sydney, Australia) (submittedVersion, cc-by-nc-nd)
Funders
Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD) Human Placenta Project of the National Institutes of Health (UO1-HD087209)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
Keywords
3D ultrasound, hybrid neural network, transfer learning, exponential weighting, early pregnancy assessment, placental morphometry
MeSH terms
Amniotic Fluid, Female, Humans, Image Processing, Computer-Assisted, Placenta, Pregnancy, Ultrasonography, Neural Networks, Computer

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review