CoChat
Sign Up
Cited by 575Domain Adaptation and Few-Shot LearningMedical Image Segmentation TechniquesArtificial Intelligence in Healthcare and Education

Self-supervised learning for medical image analysis using image context restoration

Open access
Chen, Liang, Bentley, Paul, Mori, Kensaku, Misawa, Kazunari, Fujiwara, Michitaka, Rueckert, Daniel·2019·Medical Image Analysis
DOI 10.1016/j.media.2019.101539PMC PMC7613987
Open this paper in CoChat — free account

Abstract

Machine learning, particularly deep learning has boosted medical image analysis over the past years. Training a good model based on deep learning requires large amount of labelled data. However, it is often difficult to obtain a sufficient number of labelled images for training. In many scenarios the dataset in question consists of more unlabelled images than labelled ones. Therefore, boosting the performance of machine learning models by using unlabelled as well as labelled data is an important but challenging problem. Self-supervised learning presents one possible solution to this problem. However, existing self-supervised learning strategies applicable to medical images cannot result in significant performance improvement. Therefore, they often lead to only marginal improvements. In this paper, we propose a novel self-supervised learning strategy based on context restoration in order to better exploit unlabelled images. The context restoration strategy has three major features: 1) it learns semantic image features; 2) these image features are useful for different types of subsequent image analysis tasks; and 3) its implementation is simple. We validate the context restoration strategy in three common problems in medical imaging: classification, localization, and segmentation. For classification, we apply and test it to scan plane detection in fetal 2D ultrasound images; to localise abdominal organs in CT images; and to segment brain tumours in multi-modal MR images. In all three cases, self-supervised learning based on context restoration learns useful semantic features and lead to improved machine learning models for the above tasks.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes context restoration as a self-supervised pretraining task for medical images: randomly swap image patches and train a CNN to restore the original context. Across fetal ultrasound classification, abdominal CT organ localization, and brain MR tumor segmentation, context restoration pretraining outperformed autoencoder, relative-position, context-prediction, and no self-supervision.

Key findings

  • Context restoration pretraining improved CNNs in all three evaluated tasks: fetal standard scan plane classification, abdominal multi-organ localization, and brain tumor segmentation. It outperformed other self-supervised strategies and no self-supervision in the reported comparisons.
  • For fetal ultrasound classification, when using only a quarter of the training data, context restoration pretraining improved the SonoNet performance the most among the compared pretraining strategies.
  • For abdominal CT localization, context restoration pretraining produced more localization improvements than other pretraining methods; in some cases it was comparable to or better than no pretraining on more annotated training data.
  • For BraTS 2017 brain tumor segmentation, U-Nets initialized by context restoration pretraining achieved the best performance in total; using 50% of the training set gave similar performance to using the whole training set, with Dice in enhanced tumor core, sensitivity in non-enhanced and enhanced tumor cores, and Hausdorff distances even slightly better.

Methods

  • Self-supervised task: given an unlabelled image, randomly select two isolated patches and swap them; repeat T times to disorder context, then train a CNN to restore the original image using an L2 loss.
  • Evaluated on three datasets/tasks: fetal 2D ultrasound classification with 2694 examinations and SonoNet-64; abdominal CT multi-organ localization with 150 subjects and a SonoNet-like CNN; BraTS 2017 brain tumor segmentation with 285 subjects, 142 for training/validation and 143 for testing, using a 2D U-Net on 64×64 patches.
  • Compared pretraining strategies: autoencoder, relative position prediction, local context prediction, and context restoration; subsequent tasks used whole, half, or quarter of the training set. Training details include 58K iterations for ultrasound classification, 55K for CT localization, and 50K for segmentation.

Limitations

  • The paper states that the L2 loss used for context restoration results in blurry restored images.
  • The paper reports that data augmentation did not consistently improve performance and sometimes decreased segmentation accuracy; it also notes that self-supervised pretraining was not compared against supervised pretraining because supervised pretraining is difficult in medical imaging due to annotation requirements.
  • Evident from the text: evaluation is limited to three medical imaging tasks and datasets, and the paper notes that in brain tumor segmentation the performance differences among self-supervision strategies were not always significant.

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

Integrity signals

Registry

Crossref · 10.1016/j.media.2019.101539

Cites

72 of 72 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 23 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Balancing the numbers of instances in each class significantly improves performance in all three metrics (precision, recall, and F1-score) for fetal scan plane classification.

    “Balancing the numbers of instances in each class significantly improves the performance in all three metrics.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Initialising by pretrained features, particularly those from context restoration tasks, improves the CNN performance.

    “Initialising by pretrained features, particularly those from context restoration tasks, improves the CNN performance.”
    Quote in the paperSubstantial claim
  • Findingnone given

    The L2 loss results in image blur in context restoration.

    “A shortcoming is that the L2 loss results in image blur.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Self-supervised learning based on context restoration leads to improved machine learning models for classification, localization, and segmentation.

    “In all three cases, self-supervised learning based on context restoration learns useful semantic features and lead to improved machine learning models for the above tasks.”
    Quote in the paperSubstantial claim
  • Findingown data

    On a quarter of training data, data augmentation techniques can be advantageous over some self-supervised learning methods such as the CP method.

    “On quarter of training data, the data augmentation techniques can be advantageous over some self-supervised learning methods such as the CP method.”
    Quote in the paperNot a substantial claim
  • May combine two claims. 18
  • The quote only partly states this claim. 1
  • Classified as probably not a substantial claim. 21

  • May combine two claims.

    Findingown data

    When the SonoNet is trained on a quarter of the training data, the precision decreases significantly while the recall only slightly decreases.

    “when the SonoNet is trained on quarter of the training data, the precision decreases significantly while there is only slight decrease in terms of the recall.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In all three tasks (classification, localization, and segmentation), pretraining based on context restoration is superior to other self-supervised learning strategies and to no self-supervised training.

    “In all three tasks, the pretraining based on our context restoration strategy is superior to other self-supervised learning strategies, as well as no self-supervised training.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    When the SonoNet is trained on half of the training data, the precision and recall both decrease, leading to a decrease in the F1-score.

    “When the SonoNet is trained on half of the training data, the precision and recall both decrease, which lead to the decrease of the F1-score.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Data augmentation does not lead to improvement when the SonoNet is trained on half data, but leads to obvious improvement when trained on a quarter of data.

    “Data augmentation does not lead to improvement when the SonoNet is trained on half data; while it lead to obvious improvement when the SonoNet is trained on quarter of data.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingreasoning

    The SonoNet pretrained using context restoration can only offer marginal improvement on half data.

    “The SonoNet pretrained using context restoration can only offers marginal improvement.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    When learning using only a quarter of training images, the SonoNet with feature initialisation from the auto-encoder pretraining still cannot improve the baseline, while SonoNets using other pretraining strategies perform better than the baseline.

    “When learning using only a quarter of training images, the SonoNet with feature initialisation from the auto-encoder pretraining still cannot improve the baseline; while SonoNets using other pretraining strategies perform better than the baseline.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    For liver, the CNN performance decreases sharply with half training data, while it remains stable with quarter training data.

    “in terms of the liver, the CNN performance decreases sharply with half training data; while it remains stable with quarter training data.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Although the CNN pretrained using the RP method on quarter training data outperforms that using context restoration pretraining, it performs much worse in left kidney.

    “Although the CNN pretrained using the RP method on quarter training data outperforms that using context restoration pretraining, it performs much worse in left kidney.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    The Dice score in enhanced tumour core, the sensitivity in non-enhanced and enhanced tumour cores, and the Hausdorff distances in all aspects are even slightly better.

    “The Dice score in enhanced tumour core, the sensitivity in non-enhanced and enhanced tumour cores, and the Hausdorff distances in all aspects are even slightly better.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Pretraining based on relative positions and context prediction tasks improves the segmentations but they are not as good as pretraining based on context restoration task.

    “pretraining based on relative positions and context prediction tasks improves the segmentations but they are not as good as the pretraining based on the context restoration task.”
    Quote in the paperSubstantial claim

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

    Premisenone given

    For many common problems in medical imaging such as classification, localization, and segmentation, learning image context is key.

    “For many common problems in medical imaging such as classification, localization, and segmentation, learning image context is key.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    For subsequent segmentation tasks, sharing the whole CNN architecture between pretraining and segmentation allows almost all weights to be initialised from self-supervised pretraining, resulting in better segmentation results.

    “In case of subsequent segmentation tasks, the whole CNN architecture can be shared by the pretraining network and the subsequent segmentation network. As a result, almost all the weights of the subsequent segmentation CNN can be initialised using those learned in the self-supervised pretraining. This results in better segmentation results.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The removal of context in the CP method changes the image intensity distribution, so the resulting images belong to another domain and the learned features may not be useful for images in the original domain.

    “However, the removal of context changes the image intensity distribution. Thus the resulting images belong to another domain and the learned features may not be useful for images in the original domain.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Pathak et al., 2016

      Not checked in the free run.

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

    Premisenone given

    Pretrained models from the natural image domain are not useful in the medical imaging domain because the intensity distribution of natural images differs from that of medical images.

    “The pretrained models from the natural image domain are not useful in the medical imaging domain since the intensity distribution of natural images is different from that of medical images.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Exemplar learning is difficult to apply to large datasets.

    “Since each data instance is regarded as one class, the exemplar learning method is difficult to apply to large datasets.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Dosovitskiy et al., 2016 source

      Not checked in the free run.

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

    Premisereasoning

    CNN weights learned in the CP method can initialise subsequent CNNs for classification and segmentation, while CNN weights learned in the RP method can initialise subsequent classification CNNs and only the analysis part of subsequent segmentation CNNs.

    “More precisely, the CNN weights learned in the CP method can be used to initialise subsequent CNNs for classification and segmentation; while CNN weights learned in the RP method can initialise subsequent classification CNNs and only the analysis part of the subsequent segmentation CNNs.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Pathak et al., 2016

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation

    For subsequent classification tasks, the analysis part makes most contributions to context restoration, so the feature maps from the analysis part are more useful.

    “In case of subsequent classification tasks, simple CNN layers with a few deconvolution layers are sufficient (see Fig. 3) (Doersch and Zisserman, 2017). In this condition, the analysis part makes most contributions to the context restoration. Therefore, the feature maps resulting from the analysis part are more useful.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Doersch and Zisserman, 2017

      Not checked in the free run.

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

    Interpretationreasoning

    The lack of improvement from data augmentation may be because the augmented images do not cover the variance of organ shapes and positions and introduce additional variance to pixel intensities.

    “This may because the augmented images do not cover the variance of organ shapes and positions and introduce additional variance to pixel intensities.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Multi-task self-supervised learning requires significant computational resources, such as 64 GPUs for approximately 16.8K GPU hours.

    “The disadvantage of multi-task self-supervised learning is the training requires significant computational resources, i.e. 64 GPUs for approximately 16.8K GPU hours.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Doersch and Zisserman, 2017

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    In many medical image analysis problems, localizing anatomical structures is a prerequisite.

    “In many medical image analysis problems, localization anatomical structures is a prerequisite.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Heimann et al., 2009 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The L2 loss is sufficient for feature learning in context restoration, although the outputs may be blurry.

    “As suggested by Pathak et al. (2016), the L2 loss is sufficient for feature learning although the outputs from context restoration outputs may be blurry.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Pathak et al., 2016

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    In the half-data context-restoration scenario, the baseline is not far from the ceiling, so it is difficult to obtain improvements.

    “In this scenario, the baseline (i.e. random initialisation) is not far away from the ceiling (i.e. SonoNet on the whole training set). Therefore, it is difficult to obtain improvements.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    de Vos et al. (2017) proposed a novel approach which can localize anatomical structures in 3D medical images.

    “de Vos et al. (2017) proposed a novel approach which can localize anatomical structures in 3D medical images.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites de Vos et al., 2017 source

      Not checked in the free run.

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

    Premisecitation

    Multi-task self-supervised learning improves subsequent tasks more than single-task self-supervised learning.

    “The results showed that multitask self-supervised learning improves subsequent tasks more than single-task self-supervised learning.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Doersch and Zisserman, 2017

      Not checked in the free run.

  • May combine two claims.

    Premisenone given

    Existing self-supervised learning strategies applicable to medical images cannot result in significant performance improvement and often lead to only marginal improvements.

    “However, existing self-supervised learning strategies applicable to medical images cannot result in significant performance improvement. Therefore, they often lead to only marginal improvements.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Doersch et al., 2015

      Not checked in the free run.

    • cites Pathak et al., 2016

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    The CNNs for context restoration can be structured in flexible architectures depending on subsequent tasks.

    “the CNNs for context restoration can be structured in flexible architectures depending on subsequent tasks.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Gliomas are routinely assessed using MR imaging.

    “They are routinely assessed using MR imaging (Bakas et al., 2017b).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Bakas et al., 2017b

      Not checked in the free run.

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

    Premisecitation

    In computer vision, many CNNs are pretrained before the main task.

    “In computer vision, many CNNs are pretrained before the main task.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Ren et al., 2015

      Not checked in the free run.

    • cites Simonyan and Zis-serman, 2014

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    The implementation of the context restoration task is simple and straightforward, meaning that it can be widely used.

    “the implementation of the context restoration task is simple and straightforward, meaning that it can be widely used.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    The decrease in segmentation accuracy from data augmentation may be because the augmentation does not introduce any more useful variance in terms of tumour appearance.

    “2) The augmentation does not introduce any more useful variance in terms of tumour appearance.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    The decrease in segmentation accuracy from data augmentation may be because the augmentation introduces additional intensity variations which are not meaningful.

    “3) The augmentation introduces additional intensity variations which are not meaningful.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    The Faster R-CNN is based on the pretraining of the VGG-Net.

    “the Faster R-CNN (Ren et al., 2015) is based on the pretraining of the VGG-Net (Simonyan and Zis-serman, 2014).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites Ren et al., 2015

      Not checked in the free run.

    • cites Simonyan and Zis-serman, 2014

      Not checked in the free run.

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

    Interpretationreasoning

    The decrease in segmentation accuracy from data augmentation may be because the segmentation is based on patches so a large number of image patches can be sampled.

    “which may because of three reasons: 1) The segmentation is based on patches so a large number of image patches can be sampled.”
    Quote in the paperNot a substantial claim

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

    Interpretationcitation

    The ideal pretraining task should have a similar goal to the subsequent task.

    “We conclude from the existing self-supervised feature learning literature that the ideal pretraining task should have similar goal to the subsequent task.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    This type of pretraining leads to good detection results in the Faster R-CNN.

    “This type of pretraining leads to good detection results in the Faster R-CNN.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Ren et al., 2015

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    If less training data leads to significant decrease of results, self-supervised learning is likely to improve the results significantly.

    “It is noteworthy that if less training data leads to significant decrease of results, self-supervised learning is likely to improve the results significantly.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    If the reduction of training data causes significant performance decrease, the context restoration pretraining can offer significant performance improvement over the baselines.

    “if the reduction of training data causes significant performance decrease, the context restoration pretraining can offer significant performance improvement over the baselines.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Compared with existing strategies such as relative positions and context prediction, solving the context restoration task requires pattern recognition and prediction, which ensures the context restoration task offers more efficient image semantics.

    “Compared with the existing strategies such as relative positions and context prediction, solving the context restoration task requires pattern recognition and prediction, which ensures the context restoration task offers more efficient image semantics.”
    Quote in the paperNot a substantial claim
38 of 38

  • The paper relies on this without support.

    Training CNNs only using a small number of labelled images cannot always achieve satisfactory results and does not exploit the potentially large number of unlabelled images that may be available.

    “Training CNNs only using the small number of labelled images cannot always achieve satisfactory results and does not exploit the potentially large number of unlabelled images that may be available.”

  • The paper relies on this without support.

    It is usually easier to access a large number of unlabelled medical images rather than a large number of annotated images.

    “This means it is usually easier to access a large number of unlabelled medical images rather than a large number of annotated images.”

  • The paper relies on this without support.

    In medical imaging, most images are in greyscale so that no colour information is available.

    “In medical imaging, most images are in greyscale so that no colour information is available.”

  • The paper relies on this without support.

    The key challenge for self-supervised learning is identifying a suitable self-supervision task, i.e. generating input and output instance pairs from data.

    “The key challenge for self-supervised learning is identifying a suitable self supervision task, i.e. generating input and out-put instance pairs from data.”

  • The paper relies on this without support.

    Manual cropping requires expert knowledge and is costly.

    “However, manual cropping requires expert knowledge and costly.”

  • The paper relies on this without support.

    Not all patients have two kidneys.

    “This is because not all patients have two kidneys.”

  • The paper relies on this without support.

    Gliomas are the major brain tumours occurring in adults.

    “Gliomas are the major brain tumours occurring in adults.”

  • The paper relies on this without support.

    Accurate segmentation of gliomas on MR image is a key step for quantification.

    “Accurate segmentation of gliomas on MR image is a key step for quantification.”

  • The paper relies on this without support.

    In medical image analysis, the image context is the common feature for classification, localization/detection, and segmentation tasks.

    “Particularly, in medical image analysis, the image context is the common feature for classification, localization/detection, and segmentation tasks.”
9 of 9

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.

  • Corrupted dataset is defined as a function of itself

    Minor

    The first equation says the corrupted dataset is obtained by applying f to that same corrupted dataset. The later inverse mapping instead implies that f maps each original image to its corrupted version.

1 of 1

Cites

Flagged first, then by citations

1 of the 73 references couldn’t be matched to a known paper.

Cited by

Sorted by citations

Related

Authors
Chen, L., Bentley, P., Mori, K., Misawa, K., Fujiwara, M., & Rueckert, D.
Affiliations
  • Imperial College London (GB)
  • Nagoya University (JP)
  • Aichi Cancer Center (JP)
  • Nagoya University Hospital (JP)
Corresponding author
Chen, Liang
ORCID
Venue
Medical Image Analysis
Publisher
Elsevier BV
Published
2019-12
Volume
58
Pages
101539
ISSN
1361-8415 (ISSN-L)
DOI
10.1016/j.media.2019.101539
PMC
PMC7613987
OpenAlex
W2964744899
Semantic Scholar
d6e78069dbbb35592f013187ef3432286def8f18
Cited by
575
References
73
Citations by year
2026: 44 · 2025: 85 · 2024: 99 · 2023: 133 · 2022: 105 · 2021: 74 · 2020: 34 · 2019: 1
FWCI
24.53
Type
journal article
Language
en
Open access
Licence
Locations
  • Medical Image Analysis (publishedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
  • Spiral (Imperial College London) (submittedVersion, cc-by-nc-nd)
Funders
Wellcome Trust (102431)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
Keywords
self-supervised learning, unlabeled data, CT images, brain tumor segmentation, multimodal MRI, image classification, image localization, image segmentation
MeSH terms
Abdomen, Supervised Machine Learning, Brain Neoplasms, Humans, Magnetic Resonance Imaging, Tomography, X-Ray Computed, Ultrasonography, Prenatal

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