CoChat
Sign Up
Cited by 813Brain Tumor Detection and ClassificationMedical Image Segmentation TechniquesAdvanced Neural Network Applications

A deep learning model integrating FCNNs and CRFs for brain tumor segmentation

Open access
Zhao, Xiaomei, Wu, Yihong, Song, Guidong, Li, Zhenye, Zhang, Yazhuo, Fan, Yong·2018·Medical Image Analysis
DOI 10.1016/j.media.2017.10.002PMC PMC6029627
Open this paper in CoChat — free account

Abstract

HighlightsA deep learning model integrating FCNNs and CRFs for brain tumor segmentation.The integration of FCNNs and CRF‐RNN improves the segmentation robustness.A segmentation model with Flair, T1c, and T2 scans achieves competitive performance. Abstract Accurate and reliable brain tumor segmentation is a critical component in cancer diagnosis, treatment planning, and treatment outcome evaluation. Build upon successful deep learning techniques, a novel brain tumor segmentation method is developed by integrating fully convolutional neural networks (FCNNs) and Conditional Random Fields (CRFs) in a unified framework to obtain segmentation results with appearance and spatial consistency. We train a deep learning based segmentation model using 2D image patches and image slices in following steps: 1) training FCNNs using image patches; 2) training CRFs as Recurrent Neural Networks (CRF‐RNN) using image slices with parameters of FCNNs fixed; and 3) fine‐tuning the FCNNs and the CRF‐RNN using image slices. Particularly, we train 3 segmentation models using 2D image patches and slices obtained in axial, coronal and sagittal views respectively, and combine them to segment brain tumors using a voting based fusion strategy. Our method could segment brain images slice‐by‐slice, much faster than those based on image patches. We have evaluated our method based on imaging data provided by the Multimodal Brain Tumor Image Segmentation Challenge (BRATS) 2013, BRATS 2015 and BRATS 2016. The experimental results have demonstrated that our method could build a segmentation model with Flair, T1c, and T2 scans and achieve competitive performance as those built with Flair, T1, T1c, and T2 scans. Graphical abstract Figure. No Caption available.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a unified deep learning model integrating fully convolutional neural networks (FCNNs) and conditional random fields as recurrent neural networks (CRF-RNN) for brain tumor segmentation. It reports competitive performance on BRATS 2013, 2015, and 2016, including first place on the BRATS 2016 multi-temporal evaluation, while using only Flair, T1c, and T2 scans.

Key findings

  • On BRATS 2013, CRF-RNN and post-processing improved Dice and PPV for complete tumor, core, and enhancing regions, but generally reduced Sensitivity; CRF improved Sensitivity of the enhancing region, and FCNN + CRF + post-process obtained the best Dice and PPV.
  • Among the six post-processing steps, Step 3—removing small 3D-connected segmented tumor regions—played the most important role.
  • A model built on 3 imaging modalities (Flair, T1c, T2) achieved competitive performance to one built on 4 modalities (Flair, T1, T1c, T2).
  • The method ranked second on the BRATS 2013 Challenge dataset and first on the Leaderboard dataset at the time; it ranked first on the BRATS 2016 multi-temporal evaluation among 19 teams. Segmentation took 2–4 min per view per subject, and training took ~12 days on a shared GPU server.

Methods

  • The model integrates FCNNs and CRF-RNN in one network. Training has 3 steps: train FCNNs using image patches; train CRF-RNN using image slices with FCNNs fixed; fine-tune the whole network using image slices. Three models are trained for axial, coronal, and sagittal views and fused by majority voting, followed by post-processing.
  • Data: BRATS 2013 (65 glioma patients: 14 LGG, 51 HGG; training set 20 HGG and 10 LGG; Challenge 10 HGG; Leaderboard 21 HGG and 4 LGG), BRATS 2015 (training 220 HGG and 54 LGG; testing 110 cases with unknown grades), and BRATS 2016 (same training dataset as BRATS 2015; testing 191 cases with unknown grades).
  • Evaluation used BRATS metrics Dice, PPV, and Sensitivity for complete tumor, core region, and enhancing region, with results evaluated by the BRATS evaluation website. Preprocessing included N4ITK bias correction and robust intensity normalization; models were built on Caffe.

Limitations

  • The authors state that post-processing parameters were set based on BRATS 2013 and were not cross-validated because the BRATS 2013 training set was small, so they are not necessarily optimal; the same parameters were used for all BRATS experiments.
  • The authors state that 2D CNNs cannot take full advantage of 3D MRI information, and slice-based training has imbalanced class pixel numbers that may worsen performance; ongoing work is a fully 3D network.
  • The authors note that BRATS 2016 test data contained unprocessed or partially preprocessed images, requiring manual bounding boxes and post-processing Step 2; direct comparison with BRATS 2015 participants was not possible because only 53 testing cases were available during the challenge versus 110 now. They also state that accurate training time could not be estimated because a shared GPU server was used.

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

Integrity signals

Registry

Crossref · 10.1016/j.media.2017.10.002

Cites

48 of 48 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 14 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Post-processing reduced Sensitivity.

    “However, CRFs and post-process reduced Sensitivity.”
    Quote in the paperSubstantial claim
  • Findingown data

    CRFs improved Sensitivity of the enhancing region.

    “It is worth noting that CRFs improved Sensitivity of the enhancing region.”
    Quote in the paperNot a substantial claim
  • Findingown data

    The proposed post-processing procedure further improved segmentation performance compared to 3D CRF post-processing.

    “However, our proposed post-processing procedure could further improve the segmentation performance.”
    Quote in the paperSubstantial claim
  • Findingown data

    CRFs reduced Sensitivity on the complete and core regions.

    “In summary, CRFs improved both the Dice and PPV and decreased the Sensitivity on the complete and core regions, FCNN + CRF + post-process obtained the best performance with respect to Dice and PPV, but degraded the performance with respect to the Sensitivity, especially on the complete tumor region.”
    Quote in the paperSubstantial claim
  • Findingown data

    FCNN + CRF improved the spatial and appearance consistency of segmentation results.

    “These segmentation results demonstrated that FCNN + CRF could improve the spatial and appearance consistence of segmentation results, and FCNN + CRF + post-process could reduce false positives.”
    Quote in the paperSubstantial claim
  • May combine two claims. 9
  • Classified as probably not a substantial claim. 12
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingown data

    The proposed method can build a segmentation model using Flair, T1c, and T2 scans and achieve competitive performance compared with models built with Flair, T1, T1c, and T2 scans.

    “The experimental results have demonstrated that our method could build a segmentation model with Flair, T1c, and T2 scans and achieve competitive performance as those built with Flair, T1, T1c, and T2 scans.”
    • Also stated in: § 1. Introduction · hedged
      “Our method could achieve competitive segmentation performance based on 3 MR imaging modalities (Flair, T1c, T2), rather than 4 modalities (Flair, T1, T1c, T2)”
    Quote in the paperSubstantial claim

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

    Findingown data

    It took ~12 days to train our segmentation models on the shared GPU server.

    “On the shared GPU server, it took ~12 days to train our segmentation models.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Our method ranked second on the BRATS 2013 Challenge dataset and first on the Leaderboard dataset.

    “Particularly, the method proposed by Sergio Pereira et al (Pereira et al., 2016) ranked first on the Challenge dataset and second on the Leaderboard dataset right now, while our method ranked second on the Challenge dataset and first on the Leaderboard dataset right now.”
    Quote in the paperSubstantial claim

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

    Findingown data

    It took 2-4 min for one of the three views of our method to segment one subject's imaging data.

    “In general, it took 2–4 min for one of the three views of our method to segment one subject’s imaging data.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Generative model based brain tumor segmentation methods typically require prior information, which can be gained through probabilistic image atlases.

    “The generative model based brain tumor segmentation methods typically require prior information, which could be gained through probabilistic image atlases (Gooya et al., 2012 ; Cuadra et al., 2004 ; Menze et al., 2010).”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites Gooya et al., 2012

      Not checked in the free run.

    • cites Cuadra et al., 2004

      Not checked in the free run.

    • cites Menze et al., 2010

      Not checked in the free run.

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

    Premisecitation

    Existing automatic and semi-automatic brain tumor segmentation methods can be broadly categorized as generative model based or discriminative model based methods.

    “The existing automatic and semi-automatic brain tumor segmentation methods can be broadly categorized as either generative model based or discriminative model based methods (Menze et al., 2015).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Menze et al., 2015 source

      Not checked in the free run.

  • The quote only partly states this claim.

    Interpretationreasoning

    Using 3 MR imaging modalities instead of 4 could help reduce the cost of data acquisition and storage.

    “which could help reduce the cost of data acquisition and storage.”
    Quote in the paperSubstantial claim

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

    Premisenone given

    The performance of discriminative model based segmentation methods depends on the image features and classification algorithms.

    “The performance of discriminative model based segmentation methods are hinged on the image features and classification algorithms.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    The proposed method can segment brain images slice-by-slice and is much faster than image patch based segmentation methods.

    “Our method could segment brain images slice-by-slice, much faster than those based on image patches.”
    • Also stated in: § 1. Introduction · plain
      “The proposed method is able to segment brain images slice-by-slice, which is much faster than the image patch based segmentation methods.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Gliomas invade the surrounding brain tissues rather than displacing them, causing fuzzy boundaries.

    “(3) gliomas invade the surrounding brain tissues rather than displacing them, causing fuzzy boundaries (Goetz et al., 2016)”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Goetz et al., 2016

      Not checked in the free run.

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

    Interpretationreasoning

    The standard deviation calculated based on intensity mean is more sensitive to noise than the robust deviation.

    “Thus the standard deviation calculated based on intensity mean is more sensitive to noise than the robust deviation.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The intensity mean is more sensitive to noise than the gray value of the highest histogram bin.

  • Classified as probably not a substantial claim.

    Interpretationnone given

    The CRF-RNN takes the FCNNs' prediction results and image information as input to globally optimize appearance and spatial consistency of the segmentation results according to each pixel's intensity and position information.

    “The FCNNs predict the probability of assigning segmentation labels to each pixel, and the CRF-RNN takes the prediction results and image information as its input to globally optimize appearance and spatial consistency of the segmentation results according to each pixel’s intensity and position information.”
    Quote in the paperSubstantial claim

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

    Premisereasoning

    Image patch-based deep learning segmentation methods can avoid the training sample imbalance problem because the number and location of training image patches for each class can be controlled by changing the image patch sampling scheme.

    “Since the number and location of training image patches for each class can be easily controlled by changing the image patch sampling scheme, image patch-based deep learning segmentation methods can avoid the training sample imbalance problem.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    FCNNs can segment a testing image slice by slice with improved computational efficiency, even though the model is trained using image patches.

    “However, FCNNs can segment a testing image slice by slice with improved computational efficiency (Havaei et al., 2017), even though the model is trained using image patches.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Havaei et al., 2017 source

      Not checked in the free run.

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

    Premisenone given

    Among deep learning based tumor segmentation methods, methods built upon CNNs have achieved better performance.

    “Among the deep learning based tumor segmentation methods, the methods built upon CNNs have achieved better performance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationcitation

    The proposed deep learning model integrates FCNNs and CRFs, with CRFs formulated as Recurrent Neural Networks (CRF-RNN).

    “The proposed deep learning model for brain tumor segmentation integrates Fully Convolutional Neural Networks (FCNNs) and Conditional Random Fields (CRFs), as illustrated by Fig. 3. We formulated CRFs as Recurrent Neural Networks (RNNs), referred to as CRF-RNN (Zheng et al., 2015).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Zheng et al., 2015

      Not checked in the free run.

  • May combine two claims.

    Premisenone given

    3D-CNNs can potentially take full advantage of 3D information of MRI data, but increase network size and computational cost.

    “Although 3D-CNNs can potentially take full advantage of 3D information of the MRI data, the network size and computational cost are increased too.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisenone given

    Most CNN brain tumor segmentation methods assumed that each voxel's label is independent and did not take appearance and spatial consistency into consideration.

    “Most of the above CNN brain tumor segmentation methods assumed that each voxel’s label is independent, and they didn’t take the appearance and spatial consistency into consideration.”
    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

    The post-processing parameters were set based on the BRATS 2013 dataset and were not cross-validated because the number of BRATS 2013 training cases is small, so they are not necessarily optimal.

    “The parameters were set based on the BRATS 2013 dataset. Since the number of training cases of the BRATS 2013 is small, we did not cross-validate the parameters, therefore they are not necessarily optimal.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Equal numbers of training samples for different classes are extracted to avoid the data imbalance problem.

    “Equal numbers of training samples for different classes are extracted to avoid data imbalance problem.”
    Quote in the paperSubstantial claim
20 of 20

  • The paper relies on this without support.

    Manual annotations provide valid ground truth for evaluating brain tumor segmentation.

    “Ground truths were produced by manual annotations.”

  • The paper relies on this without support.

    Different MRI scans of the same modality have comparable image intensity information.

    “Preprocessing of MRI data plays an important role in the discriminative model based tumor segmentation methods that assume different MRI scans of the same modality have comparable image intensity information.”

  • The paper relies on this without support.

    The relationship among image patches is typically lost in image patch-based segmentation methods.

    “However, a limitation of image patch-based segmentation methods is that relationship among image patches is typically lost.”

  • The paper relies on this without support.

    The intensity mean is more sensitive to noise than the gray value of the highest histogram bin.

    “Besides, the intensity mean is more sensitive to noise than the gray value of the highest histogram bin.”

  • The paper relies on this without support.

    Isolated local areas with super high intensities are usually caused by imaging noise rather than tumors.

    “taking into consideration that isolated local areas with super high intensities are usually caused by imaging noise rather than tumors.”

  • The paper relies on this without support.

    Tumor tissues have high signal in at least one modality of Flair, T1c, and T2.

    “In general, tumor tissues have high signal in at least one modality of Flair, T1c, and T2.”

  • The paper relies on this without support.

    Voxels with low signal in Flair, T1c, and T2 at the same time are generally not tumor tissues.

    “Voxels with low signal in Flair, T1c, and T2 at the same time are generally not tumor tissues.”

  • The paper relies on this without support.

    Holes in Res are very likely to be necrosis.

    “Holes in Res are very likely to be necrosis.”

  • The paper relies on this without support.

    Our model may mistakenly label necrosis areas as enhancing core.

    “Our model may mistakenly label necrosis areas as enhancing core.”

  • The paper relies on this without support.

    Our tumor segmentation model is not sensitive to non-enhancing core.

    “Our tumor segmentation model is not sensitive to non-enhancing core.”

  • The paper relies on this without support.

    In our model, non-enhancing regions might be mistakenly labeled as edema, especially when the enhancing core region is very small.

    “In our model, non-enhancing regions might be mistakenly labeled as edema, especially when the enhancing core region is very small.”

  • The paper relies on this without support.

    In image slices, the numbers of pixels for different classes are different, which may worsen the segmentation performance of the trained network.

    “However, in image slices, the numbers of pixels for different classes are different, which may worsen the segmentation performance of the trained network.”
12 of 12

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?

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Zhao, X., Wu, Y., Song, G., Li, Z., Zhang, Y., & Fan, Y.
Affiliations
  • Chinese Academy of Sciences (CN)
  • Capital Medical University (CN)
Corresponding author
Wu, Yihong
ORCID
Venue
Medical Image Analysis
Publisher
Elsevier BV
Published
2018-01
Volume
43
Pages
98-111
ISSN
1361-8415 (ISSN-L)
DOI
10.1016/j.media.2017.10.002
PMC
PMC6029627
OpenAlex
W2587828787
Semantic Scholar
c01f6960618c9b357832d6d654c6bc2af4c90b84
Cited by
813
References
48
Citations by year
2026: 33 · 2025: 61 · 2024: 85 · 2023: 115 · 2022: 132 · 2021: 127 · 2020: 117 · 2019: 94 · 2018: 45 · 2017: 3
FWCI
35.28
Type
journal article
Language
en
Open access
Licence
www.elsevier.com/tdm/userlicense/1.0/ (tdm)
Locations
  • Medical Image Analysis (publishedVersion)
  • PubMed (publishedVersion)
  • Europe PMC (PubMed Central) (acceptedVersion)
  • PubMed Central (acceptedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • National High Technology Research and Development Program of China (2015AA020504)
  • National Natural Science Foundation of China (61572499, 61421004, 61473296)
  • NIH (EB022573, CA189523)
Topics
  • Life Sciences › Neuroscience › Neurology › Brain Tumor Detection and Classification
  • 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
Keywords
brain tumor segmentation, fully convolutional network, conditional random fields, multi-view image fusion, image patches, BraTS
MeSH terms
Brain Neoplasms, Glioma, Histological Techniques, Humans, Models, Neurological, Nerve Net

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 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

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