CoChat
Sign Up
Cited by 843Medical Image Segmentation TechniquesAdvanced Neural Network ApplicationsBrain Tumor Detection and Classification

Interactive medical image segmentation using deep learning with image-specific fine tuning

Open access
Wang, Guotai, Li, Wenqi, Zuluaga, Maria A., Pratt, Rosalind, Patel, Premal A., Aertsen, Michael, Doel, Tom, David, Anna L. +3 more·2018·IEEE Transactions on Medical Imaging
DOI 10.1109/tmi.2018.2791721PMC PMC6051485
Open this paper in CoChat — free account

Abstract

Convolutional neural networks (CNNs) have achieved state-of-the-art performance for automatic medical image segmentation. However, they have not demonstrated sufficiently accurate and robust results for clinical use. In addition, they are limited by the lack of image-specific adaptation and the lack of generalizability to previously unseen object classes (a.k.a. zero-shot learning). To address these problems, we propose a novel deep learning-based interactive segmentation framework by incorporating CNNs into a bounding box and scribble-based segmentation pipeline. We propose image-specific fine tuning to make a CNN model adaptive to a specific test image, which can be either unsupervised (without additional user interactions) or supervised (with additional scribbles). We also propose a weighted loss function considering network and interaction-based uncertainty for the fine tuning. We applied this framework to two applications: 2-D segmentation of multiple organs from fetal magnetic resonance (MR) slices, where only two types of these organs were annotated for training and 3-D segmentation of brain tumor core (excluding edema) and whole brain tumor (including edema) from different MR sequences, where only the tumor core in one MR sequence was annotated for training. Experimental results show that: 1) our model is more robust to segment previously unseen objects than state-of-the-art CNNs; 2) image-specific fine tuning with the proposed weighted loss function significantly improves segmentation accuracy; and 3) our method leads to accurate results with fewer user interactions and less user time than traditional interactive segmentation methods.

Integrity signals

Registry

Crossref · 10.1109/tmi.2018.2791721

Cites

42 of 42 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 15 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Using image-specific fine-tuning together with the proposed weighted loss markedly increases segmentation accuracy.

    “2) image-specific fine-tuning with the proposed weighted loss function significantly improves segmentation accuracy;”
    Quote in the paperSubstantial claim
  • Findingown data

    The proposed model segments previously unseen objects more robustly than state-of-the-art CNNs.

    “Experimental results show that 1) our model is more robust to segment previously unseen objects than state-of-the-art CNNs;”
    Quote in the paperSubstantial claim
  • Findingown data

    BIFSeg improved accuracy from the initial segmentation more than CRF or BIFSeg(-w) did in the 2D fetal MRI task.

    “It shows that BIFSeg achieves a larger improvement of accuracy from the initial segmentation when compared with the use of CRF or BIFSeg(-w).”
    Quote in the paperSubstantial claim
  • Findingown data

    After one refinement round, BIFSeg was significantly more accurate (p < 0.05) than P-Net + CRF and BIFSeg(-w) for the placenta, unseen fetal lungs, and maternal kidneys.

    “BIFSeg achieves significantly better accuracy (p-value < 0.05) for the placenta, and previously unseen fetal lungs and maternal kidneys compared with P-Net + CRF and BIFSeg(-w).”
    Quote in the paperSubstantial claim
  • Findingown data

    For the placenta and fetal brain, FCN, U-Net, and P-Net produced visually similar results close to the reference.

    “For the placenta and fetal brain, FCN, U-Net and P-Net achieve visually similar results that are close to the ground truth.”
    Quote in the paperNot a substantial claim
  • May combine two claims. 39
  • Classified as probably not a substantial claim. 13

  • May combine two claims.

    Findingnone given

    The proposed method achieves accurate results with fewer user interactions and less user time than conventional interactive segmentation approaches.

    “3) our method leads to accurate results with fewer user interactions and less user time than traditional interactive segmentation methods.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    With the same scribbles, BIFSeg produced better results than P-Net + CRF and BIFSeg(-w).

    “In contrast, BIFSeg achieves better results with the same set of scribbles.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    DeepMedic performed poorly on T1c but well on FLAIR, while HighRes3DNet showed the opposite pattern.

    “It shows that the performance of DeepMedic is low for T1c but high for FLAIR, and that of HighRes3DNet is the opposite.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Average machine time was 3.87 s for 3D GrabCut, 65.31 s for DeepMedic, and 3.83 s for PC-Net on a laptop, and 1.10 s for HighRes3DNet on a cluster.

    “The average machine time for 3D GrabCut, DeepMedic, and PC-Net is 3.87s, 65.31s and 3.83s, respectively (on the laptop), and that for HighRes3DNet is 1.10s (on the cluster).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    CRF improved segmentation somewhat, but large under-segmentation areas remained.

    “CRF improves the segmentation to some degree, but large areas of under-segmentation still exist.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For fetal MRI segmentation, BIFSeg used noticeably less user time while achieving similar or higher accuracy than Slic-Seg, GrabCut, and Random Walks.

    “It shows that BIFSeg takes noticeably less user time with similar or higher accuracy compared with the other three interactive segmentation methods.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    GrabCut segmented fetal MRI organs poorly, except for the fetal brain, where target-background contrast was good.

    “It can be observed that GrabCut achieves a poor segmentation except for the fetal brain where there is a good contrast between the target and the background.”
    Quote in the paperNot a substantial claim

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

    Findingown data

    In the 2D case, BIFSeg required 0.72 s on average for unsupervised image-specific fine-tuning.

    “In this 2D case, BIFSeg takes 0.72s in average for unsupervised image-specific fine-tuning.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    HighRes3DNet and PC-Net gave similar results on T1c tumor core, but PC-Net was less complex and used less memory.

    “HighRes3DNet and PC-Net obtain similar results, but PC-Net is less complex and has lower memory consumption.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    P-Net gave the best accuracy for unseen fetal lungs and maternal kidneys, with an average machine time of 0.16 s.

    “It shows that P-Net achieves the best accuracy for unseen fetal lungs and maternal kidneys with average machine time 0.16s.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For FLAIR whole tumor, CNNs outperformed 3D GrabCut, and DeepMedic and PC-Net performed better than HighRes3DNet.

    “and the CNNs outperform 3D GrabCut, with DeepMedic and PC-Net performing better than HighRes3DNet.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    BIFSeg performed better than CRF and BIFSeg(-w) for unsupervised brain tumor refinement.

    “In contrast, BIFSeg performs better than CRF and BIFSeg(-w).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    BIFSeg raised average Dice from 82.66% to 86.13% for tumor core and from 83.52% to 86.29% for whole tumor.

    “BIFSeg improves the average Dice score from 82.66% to 86.13% for the tumor core, and from 83.52% to 86.29% for the whole tumor.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Image-specific fine-tuning acts as post-processing after initial segmentation and outperforms CRF.

    “This acts as a post-processing step after the initial segmentation and out-performs CRF.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    Because scribbles are placed only in mis-segmented areas, their position and length vary less than freely drawn scribbles in methods such as Random Walks and Slic-Seg, and BIFSeg output varies only slightly with different scribbles.

    “Since the scribbles are provided only in mis-segmented areas, the variations of position and length of scribbles are limited and much smaller than that of freely drawn scribbles used in traditional methods such as Random Walks [18] and Slic-Seg [21], and the output of BIFSeg also varies slightly with varying scribbles, as shown in Fig. 8.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [18] Grady L Schiwietz T Aharon S Westermann R Random walks for interactive organ segmentation in two and three dimensions: implementation and validation MICCAI 2005 773 780 16686030 source

      Not checked in the free run.

    • cites [21] Wang G Zuluaga MA Pratt R Aertsen M Doel T Klusmann M David AL Deprest J Vercauteren T Ourselin S Slic-Seg: A minimally interactive segmentation of the placenta from sparse and motion-corrupted fetal MRI in multiple views Medical Image Analysis 2016 34 137 147 27179367 PMC5052128 source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    BIFSeg achieved better refined results for both T1c tumor core and FLAIR whole tumor.

    “In contrast, BIFSeg achieves better refined results for both the tumor core in T1c and the whole tumor in FLAIR.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    DeepMedic takes over 60 s for inference, and HighRes3DNet uses too much memory to run on a laptop.

    “DeepMedic takes over 60 seconds for inference, while HighRes3DNet has too large a memory consumption to work on a laptop.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Experiments on 2D fetal MRI organs and 3D brain tumors show that the method performs well on previously unseen objects and that image-specific fine-tuning outperforms CRF.

    “Experiments on segmenting multiple organs from 2D fetal MRI and brain tumors from 3D MRI show that our method performs well on previously unseen objects and the image-specific fine-tuning outperforms CRF.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    BIFSeg achieves similar or higher accuracy with fewer user interactions and less user time than traditional interactive segmentation methods.

    “BIFSeg achieves similar or higher accuracy with fewer user interactions and less user time than traditional interactive segmentation methods.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given

    The proposed framework is general and flexible enough to perform both 2D and 3D segmentation with minimal constraints on network architecture.

    “Our framework is general, flexible and can handle both 2D and 3D segmentations with few assumptions of network structures.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    The energy minimization problem stays well-posed when the weighted loss is used.

    “Note that the energy optimization problem of Eq. (1) remains well-posed with Eq. (9), (10), and (11).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    Scribbles provided by users are much more reliable than other pixels and should influence the loss more strongly.

    “First, user-provided scribbles have much higher confidence than the other pixels, and they should have a higher impact on the loss function, leading to a weighted version of Eq. (3):”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    HighRes3DNet handles whole 3D images with relatively fast inference but requires substantial GPU memory.

    “HighRes3DNet [8] works on 3D whole images with relatively fast inference but needs a large amount of GPU memory, leading to high hardware requirements.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8] Li W Wang G Fidon L Ourselin S Cardoso MJ Vercauteren T On the compactness, efficiency, and representation of 3D convolutional networks: brain parcellation as a pretext task IPMI 2017 348 360

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

  • May combine two claims.

    Premisecitation

    DeepMedic processes 3D local patches to lower memory use, but its inference is slow.

    “For example, DeepMedic [7] works on 3D local patches to reduce memory requirements but results in a slow inference.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [7] Kamnitsas K Ledig C Newcombe VFJ Simpson JP Kane AD Menon DK Rueckert D Glocker B Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation Medical Image Analysis 2017 36 61 78 27865153

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The CNNs are built and trained to capture shared features like saliency, contrast, and hyper-intensity across objects, aiding generalization to unseen objects.

    “This is based on the fact that our CNNs are designed and trained to learn some common features, such as saliency, contrast and hyper-intensity, across different objects, which helps to generalize to unseen objects.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Current CNNs do not generalize well to previously unseen object classes that are absent from the training set.

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

    Premisecitation

    DeepIGeoS integrates CNNs with user interactions and has shown good interactivity.

    “DeepIGeoS [5] combines CNNs with user interactions and has demonstrated good interactivity.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [5] Wang G Zuluaga MA Li W Pratt R Patel PA Aertsen M Doel T Klusmann M David AL Deprest J Ourselin S DeepIGeoS: A deep interactive geodesic framework for medical image segmentation arXiv preprint arXiv:1707.00652 2017 PMC6594450 29993532

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    Adapting a pre-trained Gaussian mixture model to a specific image can increase segmentation accuracy.

    “It has been shown that image-specific adaptation of a pre-trained Gaussian Mixture Model (GMM) helps to improve segmentation accuracy [6].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [6] Ribeiro HL Gonzaga A Hand image segmentation in video sequence by GMM: a comprarative analysis SIBGRAPI 2006 357 364

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

  • May combine two claims.

    Study design and conductown data

    In the 3D brain tumor experiments, the tumor core in T1c images had been seen during training, while the whole tumor in FLAIR images had not.

    “Here the tumor core in T1c images was previously seen while the whole tumor in FLAIR images was previously unseen for the CNNs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    DeepMedic's small receptive field and reliance on local features explain its low T1c and high FLAIR performance; local features are hard for T1c due to complex appearance but easier for FLAIR because its appearance is less complex.

    “This is because DeepMedic has a small receptive field and tends to rely on local features. It is difficult to use local features to deal with T1c due to its complex appearance but easier to deal with FLAIR since the appearance is less complex.”
    Quote in the paperSubstantial claim

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

    Study design and conductown data

    For the 3D brain tumor task, parameters were chosen by grid search on training data only; whole tumor images were excluded from parameter learning.

    “Parameter setting was λ = 10.0, σ = 0.1, t0 = 0.2, t1 = 0.6, ϵ = 0.2, ω = 5.0 based on a grid search with the training data (i.e., whole tumor images were not used for parameter learning).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductown data

    In the 2D fetal MRI experiments, the placenta and fetal brain had been seen during training, whereas the fetal lungs and maternal kidneys had not.

    “Here the fetal brain and the placenta were previously seen objects, and the fetal lungs and maternal kidneys were previously unseen objects.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The authors split data by patient, using images from 10, 2, and 6 patients for training, validation, and testing.

    “We performed data splitting at patient level and used images from 10, 2, 6 patients for training, validation and testing, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given

    HighRes3DNet has a more complex model and tends to overfit the tumor core.

    “HighRes3DNet has a more complex model and tends to over-fit the tumor core.”
    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 conductown data

    For the 2D fetal MRI task, parameters were chosen by grid search on training data only; fetal lungs and maternal kidneys were excluded from parameter learning.

    “Parameter setting was λ = 3.0, σ = 0.1, t0 = 0.2, t1 = 0.7, ϵ = 0.2, ω = 5.0 based on a grid search with the training data (i.e., fetal lungs and maternal kidneys were not used for parameter learning).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    DeepMedic achieved higher accuracy for whole tumor because the 3D task favors recognizing hyper-intensity regions where local features matter more than global features, and DeepMedic's smaller receptive field suits this better than PC-Net or HighRes3DNet.

    “Table IV shows that DeepMedic achieves higher accuracy for the whole tumor. The reason is that our 3D experiment tends to learn to recognize hyper-intensity regions where local features have a higher influence than global features, and the relatively smaller receptive field of DeepMedic is more suitable for this task compared with the PC-Net and HighRes3DNet.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given

    The four organs were imaged with the same protocol and share similar signal-to-noise ratio and common features such as saliency, contrast, and hyper-intensity.

    “First, these four organs were imaged with the same protocol. They have similar signal to noise ratio and share some common features, such as saliency, contrast and hyper-intensity.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Applying pre-trained models to previously unseen objects is a zero-shot learning problem.

    “Applying pre-trained models to previously unseen objects is a zero-shot learning problem [33].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [33] Zhang Z Saligrama V Zero-shot learning via semantic similarity embedding ICCV 2015 4166 4174

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

  • May combine two claims.

    Interpretationreasoning

    This reduces effort for collecting and annotating training data and can be applied directly to some unseen organs.

    “It therefore reduces the efforts needed for gathering and annotating training data and can be applied to some unseen organs directly.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    The tumor core and whole tumor share common features, such as lower or higher intensity than the remaining brain regions.

    “Similarly, the tumor core and whole tumor have some common features, e.g., lower or higher intensity than the remaining brain regions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Unsupervised fine-tuning corrects small mis-segmentations when initial automated performance is satisfactory, but it may underperform on complex cases due to distribution mismatch between training and testing data.

    “Though unsupervised fine-tuning helps to correct small mis-segmentations when the initial fully-automated performance is satisfactory, it may lead to under-performance when dealing with some complex cases, considering the distribution mismatch between the training and testing data.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation

    Handling unseen objects is a major advantage over traditional CNNs and even transfer learning or weakly supervised learning, because for some objects it requires no annotated training instances.

    “Dealing with unseen objects is a major advantage compared with traditional CNNs and even transfer learning [25] or weakly supervised learning [3], since for some objects it does not require annotated instances for training at all.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [25] Tajbakhsh N Shin JY Gurudu SR Hurst RT Kendall CB Gotway MB Liang J Convolutional neural networks for medical image analysis: full training or fine tuning? TMI 2016 35 5 1299 1312 26978662 source

      Not checked in the free run.

    • cites [3] Rajchl M Lee M Oktay O Kamnitsas K Passerat-Palmbach J Bai W Rutherford M Hajnal J Kainz B Rueckert D DeepCut: Object segmentation from bounding box annotations using convolutional neural networks TMI 2017 36 2 674 683 PMC7115996 27845654 source

      Not checked in the free run.

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

    Study design and conductreasoning

    To check PC-Net's suitability for brain tumor segmentation despite its efficiency, the authors compared initial fully automated outputs of PC-Net, DeepMedic, and HighRes3D, then used only PC-Net in the interactive BIFSeg pipeline.

    “To ensure that PC-Net was suitable for the brain tumor segmentation task despite the gain in efficiency, we compared the initial fully automated output of PC-Net, DeepMedic and HighRes3D. Then, we only used PC-Net for the interactive segmentation pipeline of BIFSeg.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    PC-Net is more compact than HighRes3DNet and less prone to overfitting, which improves its ability to handle the unseen whole tumor.

    “PC-Net is more compact than HighRes3DNet and less likely to achieve over-fitting, leading to better ability to deal with the unseen whole tumor.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Compared with FCN and U-Net, P-Net has far fewer parameters without reducing receptive field, which allows it to generalize better to unseen objects.

    “Second, compared with FCN and U-Net, P-Net has far fewer parameters without reduction of the receptive field. Therefore, it can generalize better to previously unseen objects.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    The BIFSeg framework is theoretically applicable to different CNN models.

    “Our BIFSeg framework is theoretically applicable to different CNN models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The framework uses a bounding-box-based CNN for binary segmentation and can segment previously unseen objects.

    “It uses a bounding box-based CNN for binary segmentation and can segment previously unseen objects.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    In the experiments, BIFSeg hyper-parameters such as λ were fixed globally during testing.

    “In our experiments, hyper-parameters of BIFSeg (e.g., λ) were fixed globally in the testing stage.”
    Quote in the paperNot a substantial claim
47 of 47

  • The paper relies on this without support.

    Interactive segmentation remains the state of the art for existing commercial surgical planning and navigation products.

    “As such, interactive segmentation remains the state of the art for existing commercial surgical planning and navigation products.”

  • The paper relies on this without support.

    Current CNNs do not generalize well to previously unseen object classes that are absent from the training set.

    “The first challenge of using CNNs for interactive segmentation is that current CNNs do not generalize well to previously unseen object classes that are not present in the training set.”

  • The paper relies on this without support.

    Current CNNs are not adaptive to different test images because their parameters are learned from training images and then fixed during testing.

    “but current CNNs are not adaptive to different test images, as parameters of the model are learned from training images and then fixed in the testing stage, without image-specific adaptation.”

  • The paper relies on this without support.

    Interactive segmentation often requires image-specific learning to handle large context variations among different images.

    “Second, interactive segmentation often requires image-specific learning to deal with large context variations among different images”

  • The paper relies on this without support.

    Annotations for medical images are often expensive to acquire.

    “For medical images, annotations are often expensive to acquire as both expertise and time are needed to produce accurate annotations.”

  • The paper relies on this without support.

    Transitioning from GMMs to CNNs for image-specific adaptation has not yet been demonstrated.

    “However, transitioning from simple GMMs to powerful but complex CNNs in this context has not yet been demonstrated.”
6 of 6

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?

  • In 2018, McClure et al. [667] used CL with DWC: Distributed weight consolidation for segmentation of brain segmentation on 5 datasets.

    “McClure al. [667] | et | 2018 | CL | Segmentation | Brain segmentation | 5 | DWC: Distributed weight consolidation”

    From A review of uncertainty quantification in deep learning: Techniques, applications and challenges (2021) · cites this paper as [667]

    The Abstract and Experiments and Results describe image-specific fine-tuning for fetal-organ and brain-tumor segmentation, not distributed weight consolidation. The brain-tumor Data section identifies the 2015 BRATS training set, while the results identify the method as BIFSeg. Thus the paper does cover brain segmentation, but it contradicts the claimed method and five-dataset description. (LLM, conf=0.90)

    Checked by a reasoning model — may be wrong.

    Experimental — accuracy not yet measured

  • Existing deep learning-based interactive segmentation methods focus mainly on limited tasks and image modalities.

    “while existing deep learning-based interactive segmentation methods focus mainly on limited tasks and image modalities.”

    From Segment anything in medical images (2024) · cites this paper as [9]

    The Introduction says there are very few studies using CNNs for interactive segmentation and that current CNNs struggle to generalize to object classes absent from their training sets, supporting the limited-tasks point. It does not establish that existing interactive methods focus mainly on limited image modalities; its discussion of MRI applications concerns the proposed method. (LLM, conf=0.83)

    Experimental — accuracy not yet measured

  • For FCN, U-Net and P-Net, the segmentation accuracy of TTD remains close to that of the single-prediction baseline.

    “We found that for all of these three networks, the segmentation accuracy of TTD remains close to that of the single-prediction baseline.”

    From Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks (2019) · cites this paper as [4]

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.

  • Scribble penalties reverse the stated label constraint

    Minor

    Equation (6) assigns an infinite penalty when a pixel matches its scribble label and zero penalty when it disagrees. This is the opposite of the constraint in Equation (4), so Equation (5) is not the claimed unconstrained equivalent.

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 2

  1. sermanet, 2014, Overfeat Integrated Recognition Localization and Detection Using Convolutional Networks
  2. vezhnevets, 2005, GrowCut: Interactive multi-label ND image segmentation by cellular automata, Proc GraphiCon, 150

Cited by

Sorted by citations

Related

Authors
Wang, G., Li, W., Zuluaga, M. A., Pratt, R., Patel, P. A., Aertsen, M., Doel, T., David, A. L., Deprest, J., Ourselin, S., & Vercauteren, T.
Affiliations
  • Wellcome / EPSRC Centre for Interventional and Surgical Sciences (GB)
  • KU Leuven (BE)
Corresponding author
Wang, Guotai
ORCID
Venue
IEEE Transactions on Medical Imaging
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2018-07
Volume
37
Issue
7
Pages
1562-1573
ISSN
  • 0278-0062 (ISSN-L)
  • 1558-254X
DOI
10.1109/tmi.2018.2791721
PMC
PMC6051485
OpenAlex
W2763160469
Cited by
843
References
42
Citations by year
2026: 37 · 2025: 82 · 2024: 102 · 2023: 112 · 2022: 132 · 2021: 146 · 2020: 130 · 2019: 76 · 2018: 26
FWCI
47.27
Type
journal article
Language
en
Open access
Licence
Locations
  • IEEE Transactions on Medical Imaging (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • Lirias (submittedVersion, other-oa)
  • UCL Discovery (University College London) (submittedVersion)
  • Research Portal (King's College London) (submittedVersion, cc-by)
  • arXiv (Cornell University)
Funders
  • Wellcome Trust (WT101957, WT97914, HICF-T4-275)
  • Engineering and Physical Sciences Research Council (NS/A000027/1, EP/H046410/1, EP/J020990/1, EP/K005278, NS/A000050/1, 203145Z/16/Z)
  • Royal Society (RG160569)
  • University College London
  • Great Ormond Street Hospital Charity
  • Nvidia
  • Emerald, a GPU-accelerated High Performance Computer, made available by the Science and Engineering South Consortium operated in partnership with the STFC Rutherford-Appleton Laboratory
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
  • Life Sciences › Neuroscience › Neurology › Brain Tumor Detection and Classification
Keywords
zero-shot learning, scribble-based segmentation, weighted loss function, uncertainty estimation, brain tumor segmentation, magnetic resonance imaging, convolutional neural networks
MeSH terms
Deep Learning, Brain, Brain Neoplasms, Female, Fetus, Humans, Image Interpretation, Computer-Assisted, Magnetic Resonance Imaging, Pregnancy, Prenatal Diagnosis

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 2026-10-05.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-05.

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