3D MRI Image Segmentation using 3D UNet Architectures: Technical

Review

Vijaya Kamble

and Rohin Daruwala

Research Scholar Department of Electronics Engineering, Veermata Jijabai Technological Institute (VJTI),

Mumbai, Maharashtra, India

Department of Electronics Engineering, Veermata Jijabai Technological Institute (VJTI), Mumbai, Maharashtra, India

Keywords: 3D UNet, MRI

Images, Segmentation, Brain Tumors

Abstract: From last few decades machine learning & deep convolutional neural networks (CNNs) used extensively and

have shown remarkable performance in almost all fields including medical diagnostics. It is used in medical

domain for automatic tissue, lesion detection, segmentation, anatomical or structure segmentation classifica-

tion & survival predictions. In this paper we presented an extensive technical literature review on 3D CNN

U-Net architectures applied for 3D brain magnetic resonance imaging (MRI) analysis. We mainly focused on

the architectures, its modifications, pre-processing techniques, types datasets, data preparation, methodology,

GPU, tumor disease types and per architectures evaluation measures in this works. Our primary goal for this

extensive technical review is to report how different 3D U-Net architectures or CNN architectures have been

used to differentiate between state-of-the-art strategies, compare their results obtained using public/clinical

datasets and examine their effectiveness. This paper is intended to present detailed reference for further re-

search activity or plan of strategy to use 3D U-Nets for brain MRI automated tumor diseases detection, seg-

mentation & survival prediction analysis. Finally, we are presenting a novel perspective to assist research

directions on the future of CNNs & 3D U-Net architectures to explore in subsequent years to help doctors &

radiologist.

1 INTRODUCTION

Over last few decades the use of machine learning and

deep learning techniques revolutionized medical

imaging field for tumor or disease segmentation,

detection & survival predication. It is helping

physicians to diagnose brain cancers quickly to boost

prognosis. A patient’s MRI is the three-dimensional

brain anatomy (Oday Ali Hassen, et al., 2021). MRI

images of different modalities such as T1-weighted,

T2-weighted, T1c, and Flair as T1c has precise data

such as tumor form, location, and scale. Different

MRI modalities are used in brain tumor extraction

and segmentation. Among all types of brain tumors

Gliomas are most common tumors in brain cancer

with high mortality rate. These brain tumors

originating from the glial cells in the central nervous

system. Gliomas are 70% of all brain tumors. The

survival duration of patients with high grade gliomas

https://orcid.org/0000-0002-6469-8837

https://orcid.org/0000-0002-3267-6270

(HGG) lead less than 2 years if prognosis is poor.

Compared with HGG, prognosis of low grade

gliomas (LGG) are more effective (Chandan

Yogananda, et al., 2020).

Different architectures of CNNs used in medical

imaging and other applications from year 1990s.

Medical Image data is sensitive patients data and not

available easily. Earlier limitations on on

performance of CNN networks years as less labeled

medical data available. But now large annotated

medical public & clinical data sets available online &

on demand and more powerful graphics processing

units (GPUs) available for data processing so this is

enabling researchers to continue working in the area

to help doctors (Chandan Yogananda, et al., 2020).

Automated or semi automated segmentation

methods saving physicians time and provide an

accurate reproducible solutions for 3D brain tumor

analysis and patient monitoring. Convolutional neural

Kamble, V. and Daruwala, R.

3D MRI Image Segmentation using 3D UNet Architectures: Technical Review.

DOI: 10.5220/0010851300003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 2: BIOIMAGING, pages 141-146

ISBN: 978-989-758-552-4; ISSN: 2184-4305

141

networks (CNN) able to learn from examples so they

demonstrate state-of-the-art segmentation accuracy

both in 2D natural images (Andriy Myronenko, 2019)

and in 3D medical image modalities. Its difficult to

differentiate brain tumors from normal tissues

because tumor boundaries are ambiguous and there is

a high degree of variability in the shape, location,

intensity in homogeneity, or different intensity ranges

between the same sequences and acquisition scanners

and extent of the patient (Li Sun, et al., 2019). This

can influence the segmentation accuracy and correct

detection of tumor. Different hospitals shows

different gray-scale values for the same tumorous

cells may when they are scanned differently.

Although advance automatic algorithms used for

brain tumor segmentation, the problem is still remains

a challenging task.

To address issues in this research area we have

done extensive comparative review of most cited

research papers based on 3D U-Net architectures &

3D medical imaging modalities, with different

processing techniques use of powerful GPUs

different software's with various high grade tumors

classification segmentation & survival prediction of

patients.

Summary of this extensive most cited research is

mentioned in table no 1 with reference to paper, few

prominent U-Net model parameter & methodology

discussed in short with figure. Different imaging

modalities, preprocessing techniques datasets,

evaluation parameters advantages & limitations also

mentioned. Most of the reviewed content got dice

scores above 0.75 to 0.89 range for whole tumor core

tumors & enhancing tumors. Some of the papers got

excellent accuracy, sensitivity & specificity.

2 CNN ARCHITECTURES

CNN architectures used in medical imaging for

segmentation detection & predictions of disease

diagnosis prognosis. CNN architectures can be

grouped around five sub types:

I) Based on interconnected operating modules,

II) Selection of types of input MRI modalities,

III) Selection of input patch dimension,

IV) Number of Predictions at a time

V) Based on implicit and explicit contextual

information.

In this summary of the literature review methods dis-

tinguish with the different CNN architectures mostly

on types of U-Nets, pre processing, post-processing

and target of the segmentation & tumor types.

2.1 UNet Architectures Literature

Survey

In medical imaging for brain tumor disease diagnosis

prognosis for image semantic segmentation and

classifications mostly U-Net ResNet architectures are

used.

The U-Net is one of the most popular convolu-

tional neural network end-to-end architectures in the

field of semantic segmentation.a that is designed for

fast and precise segmentation of images. In several

challenges U-Net has performed extremely well.

Figure 1: UNet Architecture (Xue Feng et al.).

U-Net Architecure split the network into two parts:

Encoder: The encoder path is the backbone. The

encoder captures features at different scales of the

images by using a traditional stack of convolutional

and max pooling layers. A block in the encoder

consists of the repeated use of two convolutional

layers (k=3, s=1), each followed by a non-linearity

layer, and a max-pooling layer (k=2, s=2). For every

convolution block and its associated max pooling

operation, the number of feature maps is doubled to

ensure that the network can learn the complex

structures effectively.

Decoder: The decoder path is a symmetric expanding

counterpart that uses transposed convolutions. This

type of convolutional layer is an up-sampling method

with trainable parameters and performs the reverse of

(down)pooling layers such as the max pool. Similar to

the encoder, each convolution block is followed by

such an up-convolutional layer. The number of feature

maps is halved in every block. Because recreating a

segmentation mask from a small feature map is a

rather difficult task for the network, the output after

every up-convolutional layer is appended by the

feature maps of the corresponding encoder block. The

feature maps of the encoder layer are cropped if the

BIOIMAGING 2022 - 9th International Conference on Bioimaging

142

dimensions exceed the one of the corresponding

decoder layers.

In the end, the output passes another convolution

layer (k=1, s=1) with the number of feature maps

being equal to the number of defined labels. The

result is a u-shaped convolutional network that offers

an elegant solution for good localization and use of

context. Let’s take a look at the code. max pooling

operations (in each dimension) are the most

appropriate.

In these section the from most cited literature

review best architecture discussed. Researchers

proposed common U-nets, cascaded U-Nets,

modified type of Unet architectures for brain tumor

detection & survival predictions.

Xue Feng et al. explained generic 3D U-Net

structure with different hyper-parameters, deployment

of each model is for full volume prediction and final

ensemble modeling. Model fitting done for the survival

task feature extraction (Xue Feng, et al., 2019).

Figure 2: Cascaded Unet Architecture (Yan Hu, et al.).

Yan Hu et al proposed algorithm for intra-tumor

structure segmentation using three cascaded U-Net

models as shown in figure 2. They are concatenated

and further processed by two convolutional layers to

detect tumor region. The feature maps generated by

three cascaded U-Net models using T1, T1c, T2 and

FLAIR modalities.Patches are cropped within tumor

region detected for classification model (Yan Hu , et

al., 2018).

2.2 Pre-processing

In computer vision or image processing domain pre-

processing of image is preliminary but important task.

Brain MRI volumes acquired from scanners, these

volumes are with nonbrain tissues, parts of the head

or skull, eyes, fat, spinal cord. From acquired MRI

volumes extracting the brain tissue from non-brain

image is the pre processing primary task. This is

known as skull strippings. This is an essential step for

subsequent segmentation task. To achieve a good

performance in training supervised models such as

CNNs, or Unets the input training data hugely

influences the performance of the model, so having

preprocessed and well-annotated data is a crucial step

in MRI image processing. This step is very important

as it has direct impact on the performance of auto-

mated segmentation methods. Inclusion of skull or

eyes as brain tissue in MRI analysis may lead to

unexpected results missed classification or tumor

detection. In this review context every researcher

used some preprocessing methods & post processing

method for correct segmentation, predictions results.

In MRI image preprocessing there are few

common but fixed steps or algorithms stated as below:

i) Intensity Normalization-as there are different

image modalities,

ii) Bias field Corrections,

iii) Skull stripping,

iv) Image registration.

After Image preprocessing step there is data

preparation phase in CNN algorithms in that data

augmentation, 2D, 3D patch extraction before

segmentation & classification task.

2.3 Input Modalities

There are various types of MRI images based on their

scanning techniques, acquisition modes, intensities.

Basically in MRI four modalities are popular among

research community T1,T2 T2c Flair. In the literature

strategies of selection of modality for processing can

also be grouped according to the number of

modalities that are processed at the same time. The

major categories are two: single- and multi-modality.

3 EVALUATION

MEASUREMENTS

Final Deep CNN models performance majorly

depend on the types of dataset, modalities,

types of

tumors regions, sub-regions & model parameters. In

this extensive research survey of 3D UNet & MRI

imaging following evaluation metrics were used for

segmentation & classification of tumors:

i) Global accuracy,

ii) Dice coefficient,

iii) Recall,

iv) Precision and

v) Housedraf distance measure.

3D MRI Image Segmentation using 3D UNet Architectures: Technical Review

143

Following are the 6 Equations for evaluation

parameters with the well known terms False Positives

(FP), False Negatives (FN), True Negative(TN):

𝐺𝑙𝑜𝑏𝑎𝑙𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 

𝑇𝑃  𝑇𝑁

𝑇𝑃



𝑇𝑁



𝐹𝑃



𝐹𝑁

(1)

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 

𝑇𝑃



𝐹𝑃

(2)

𝑅𝑒𝑐𝑎𝑙𝑙 

𝑇𝑃



𝐹𝑁

(3)

𝐹1  2

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∗ 𝑅𝑒𝑐𝑎𝑙𝑙

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛  𝑅𝑒𝑐𝑎𝑙𝑙

(4)

𝐷𝑖𝑐𝑒 

2𝑇𝑃



𝐹𝑃



𝐹𝑁

(5)

𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 

𝑇𝑁



𝐹𝑃

(6)

The main evaluation measures for the challenges

mentioned pre-viously are DSC, specificity,

sensitivity, positive predictive value (pre-cision),

average surface distance (ASD), average volumetric

difference (AVD) and modified Hausdorff distance

(MHD).

Table 1 summarizes the types of architectures,

databases, numbers of samples, MRI modalities

considered, tumor diseases types, GPU types,

Software's used, evaluation measurements applied

and corresponding results reported of the extensive

technical surveyed work (Chandan Yogananda, et al.,

2020 - http://www.tomography.org/).

4 MEDICAL CNN-BASED

SOFTWARES

Now a days most of researchers release their winning

competitions or challenges source codes to the public

it helps for research in the medical & other fields.

There are few free deep learning libraries for MRI

segmentation as listed below:

i) Tensorflow,

ii) Theano,

iii) Caffe,

iv) Keras

v) PyTorch.

There are few CNN open-source frameworks

namely NiftyNet 17 and DLTK.

Researchers work on clinical & publically

available datasets depending on their application.

5 CONCLUSION AND FUTURE

SCOPE

Automatic brain tumor segmentation for cancer

diagnosis & prediction is challenging task. Most recent

advancements in medical diagnostic research using

Deep Convolution Neural network 3D Unet

architectures discussed in this technical most cited

literature review paper. This 3D Unet architectures &

modified frameworks indicate significant potential to

segment classify &predict the brain tumors lesions

from the 3D MRI images. Even though MRI images

are of different modalities intensities and categories

still complex features from these MRI images can be

automatically extracted from 3D Unet architectures it

also segment tumor with subregions. There is always

chance of improvements and modifications in CNN

architectures, Unet architectures to improve the

efficiency of segmentation, detection & predictions of

cancerous brain tumors.

With this deep technical review we observed and

analyzed that most of the proposed methods are based

on specific 3D MRI modalities for high grade tumor

segmentation so they have computational

complexities as well as memory constraints & in need

of specific GPU speed for software's. In most of

papers deep learning software libraries are used to

implement layers of deep CNNs. They are arranged

either parallel or distributed or cascaded frameworks,

which help researchers to train their models in multi-

core architectures or GPUs. Mostly Nvidia GPUs &

Intel GPUs used for training and implementation of 3

D Unet CNN models. It is observed from evaluation

measures that the training and validation for brain

image analysis is significantly affected by the data

imbalance problem .Lesions are smaller than the

entire volume so it affects generalization & robust

model. We observed that the full capacity of 3D Unet

CNN architectures has not yet been fully leveraged in

brain MRI analysis. More sophisticated dedicated

softwares are available for Medical imaging or Brain

MRI analysis. But there is always a challenge for

domain adaptation techniques, more research in this

sense is needed for permanent solutions for high

grade and low grade tumors for correct diagnosis

without experts interventions.

BIOIMAGING 2022 - 9th International Conference on Bioimaging

144

Table 1: Comparison with different U-Net models.

SrNo Article Dataset Number of scans Model GPU Softwares Segmentation

tasks

Evaluation Measure

1 Yue Zhao

et al.

BraTS 2018 75 low grade and 210 high

grade gliomas

3D recurrent

multi-fiber network

NVIDIA Tesla

V100 32 GB GPU.

Pytorch MS lesion

Whole brain,

tissue and sub-

cortical

structure

Dice scores of WT 89.62%,

TC 83.65% and ET 78.72%

2 Chandan

Ganesh

Bangalore

Yogananda

et al

BraTS2017,

BraTS2018,

Oslo data set 52

LGG and HGG

(age > 18 years)

scanned from

2003 to 2012

(200 cases 150 HGG and 50

LGG), validation (65 48 HGG

and 17 LGG) and 10% (20 12

HGG and 8 LGG)

3D-Dense-Unets

Combination of WT-net,

TC net, ET net

Tesla P100, P40 or

K80 NVIDIA-

GPUs

Tensorflow,

Keras,

python

ackage, and

Pycharm

IDEs with

(Adam)

MS lesion

Whole brain,

tissue and sub-

cortical

structure,

Stroke

Dice-scores for WT 0.90, TC

0.84, and ET 0.80

3 Oday Ali

Hassen et al

BRATS 2019

and BRATS

2017. At

BRATS 2019

3D-MRI of 336 heterogeneous

gliomas patients, 259 HGG and

76 Low-Grade Gliomas LGG

Population-based Artificial

Bee Colony Clustering (P-

ABCC) methodology, K-

means

Intel (R), Core

(TM) i3 CPU, 8.00

GB RAM

MATLAB

R2018a

Brain Tumour Entire Tumor (WT), Tumor

Center (TC), Improved (ET) by

0.03%, 0.03%, and 0.01%

respectively. At BRATS 2017,

an increase in precision for WT

was reached by 5.27%.

Jing Huang

and Minhua

Zheng et al

BRATS 2017

285 patients, 210 HGG images,

75 LGG images

NVIDIA RTX

TITAN 24GB GPU

Brain Tumour

Dice scores Similarity

WT 0.9089, TC 0.7165, and

ET 0.8398

5 Parvez

Ahmad et

BRATS

2018

228 training images 57 testing

images out of data set is 285.

Residual 3D U-net, Dense

inception-like architecture

with multiple dilated

convolutional layers

Keras Brain Tumour Dice Similarity

WT 87.16, ET 84.81, 80.20,

Whole 86.42, Sensitivity

Core, Enhancing 82.15,80.01

6 Hassan A.

Khalil et al

BRATS 2017 Clustering technique

integrates k-means and the

dragonfly algorithm

Intel, Core i3 CPU

with 8.00 GB of

RAM

MATLAB

software

R2018a

Brain Tumour Accuracy 98.20, Recall 95.13,

Precision 93.21

7 Xue Feng

et al

CBICA’s Image

Processing

Portal

163 training subjects, 285

training subjects, 66 subjects

were provided as validation

An ensemble of 3D U-Nets

with different hyper-

parameters for brain tumor

segmentation

Nvidia Titan Xp

GPU with 12 Gb

Tensorflow

framework

was used

with Adam

optimizer

Brain Tumour Accuracy was 0.321, MSE

was 99115.86, median SE

was 77757.86, std SE was

104291.596 and Spearman

Coefficient was 0.264

Andriy

Myronenko

et al

BraTS 2018

285 Training cases validation

(66 cases) and the testing sets

(191 cases)

Encoder-decoder based

CNN architecture

asymmetrically larger

encoder to smaller decoder

NVIDIA Tesla,

V100 32 GB GPU

Tensorflow Brain Tumour

Dice Similarity ET 0.7664,

WT 0.8839 and TC 0.8154

Wei Chen,

Boqiang

Liu et al.

BraTS 2018

285 subjects, of which 210 are

GBM/HGG and 75 are LGG

Separable 3D U-Net

GeForce GTX

1080Ti GPU

PyTorch

toolbox,

Adam

Brain Tumor

Dice scores of ET 0. 68946,

WT 0. 83893 and TC 0.

78347

10 Xiaojun Hu

et al.

BRATS 2015,

ISLES 2017

database

3D Brain SegNet 4 Titan Xp

GPUs,8G memory

for each GPU

Pytorch Brain Tumor Dice Score 0.30± 0.22,

0.35±0.27, 0.43±0.27

11 Li Sun et

al.

BraTS 2018 210 HGG and 75 LGG Three different 3D CNN

architectures (CA-CNN,

DFKZ Net, 3D U-Net,

Wnet, Tnet, Enet)

Brain Tumor,

survival

prediction

61.0% accuracy

12 Dmitry

Lachinov et

BraTS 2018 285 MRIs for training (210

high grade and 75 low grade

glioma images), 67 validation

and 192 testing MRIs.

Multiple Encoders Unet,

Cascaded UNet

GTX 1080TI MXNet

framework

Brain Tumor,

survival

prediction

Dice score of ET 0.720, WT

0.878, TC 0.785

Ping Liu et

BraTS 2017

285 samples with manually

annotated and confirmed

ground truth labels

Deep supervised 3D

Squeeze-and-Excitation V-

Net (DSSE-V-Net)

4 NVIDIA Titan

1080 TI 11GB

GPUs

Pytorch

Brain Tumor,

survival

prediction

Dices of WT and TC of DS-U-

Net increased to 0.8953 and

0.7828 from 0.8799 and 07693

of 3D U-Net, respectively

Pawel

Mlynarski

et al

BRATS 2017

285 scans (210 high grade

gliomas and 75 low grade

gliomas

CNN-based model,short-

range 3D context and the

long-range 2D context

Keras,

Tensor

Flow

Brain Tumor,

survival

prediction

Dice scores of WT 0.918, TC

0.883 ET 0.854

15 Suting

Peng et al

BraTS 2015 220 HGG and 54 LGG Multi-Scale 3D U-Nets

architecture

NVIDIA GeForce

GTX 1080Ti GPU

Brain Tumor,

survival

prediction

Dice similarity

WT 0.85, ET 0.72, WC 0.61

16 Mina

Ghaffari et

BraTS 2018 230 cases for training, and the

remaining 55 cases were

reserved for testing.

Modified version of the

well-known U-Net

architecture

4 x

NVIDIA Tesla

Pascal P100

Brain Tumor Dice similarity WT,0.87, ET

0.79 WC0.66

17 Parvez

Ahmad et

BraTS 2018 80% of subjects for training and

20% for validation

3D Dense Dilated

Hierarchical Architecture

Brain Tumor, Dice similarity

WT 0.8480, TC 0.8574 CT

0.8219

Shangfeng

Lu et al

Brats2019

259 high-grade gliomas (HGG)

and 76 low-grade gliomas

(LGG)

Multipath feature extraction

3D CNN

NVIDIA 1080ti

GPU with 11G

RAM

pytorch Brain Tumour

Dice similarity

WT 0.881, TC 0.837, ET

0.815

Saqib

Qamar et al

BraTS 2018

210 patients, to train and test

our model

3D Hyper-dense Connected

Convolutional Neural

Network

Brain Tumour

Dice similarity

WT 0.87, ET 0.81, CT 0.84

0 Yan Hu et

BraTS 2017 285 training subjects, 46

validation subjects and 146 test

bjects.

3D Deep neural network Intel Xeon 2.10

GHz CPU, NVIDIA

GTX 1080 Ti GPU,

32 GB RAM

Brain Tumour Dice similarity

WT 0.81, CT 0.69 and ET 0.55

3D MRI Image Segmentation using 3D UNet Architectures: Technical Review

145

REFERENCES

Chandan Ganesh Bangalore Yogananda, Bhavya R. Shah,

Maryam Vejdani-Jahromi, Sahil S. Nalawade,

Gowtham K. Murugesan, Frank F. Yu, Marco C. Pinho,

Benjamin C. Wagner, Kyrre E. Emblem, Atle Bjørne-

rud, Baowei Fei, Ananth J. Madhuranthakam, and Jo-

seph A. Maldjian “A Fully Automated Deep Learning

Network for Brain Tumor Segmentation” tomogra-

phy.org, volume 6, number 2, June 2020, ISSN 2379-

1381 https://doi.org/10.18383/j.tom.20 19.00026

Oday Ali Hassen, Sarmad Omar Abter, Ansam A. Ab-

dulhussein, Saad M. Darwish, Yasmine M. Ibrahim 4

and Walaa Sheta, “Nature-Inspired Level Set Segmen-

tation Model for 3D-MRI Brain Tumor Detection”,

2021.

Yue Zhao, Xiaoqiang Ren, Kun Hou, Wentao Li, “Recur-

rent Multi-Fiber Network for 3D MRI Brain Tumor

Segmentation”, Symmetry 2021, 13, 320.

https://doi.org/10.3390/sym13020320,https://www.md

pi.com/journal/symmetry

Parvez Ahmad, Hai Jin, Saqib Qamar, Ran Zheng, and

Wenbin Jiang, “Combined 3D CNN for Brain Tumor

Segmentation”, 2020 IEEE Conference on Multimedia

Information Processing and Retrieval (MIPR), 978-1-

7281-4272-2/20/, DOI 10.1109/MIPR49039.2020.0

0029,

Hassan A. Khalil, Saad Darwish, Yasmine M. Ibrahim and

Osama F. Hassan, “3D-MRI Brain Tumor Detection

Model Using Modified Version of Level Set Segmen-

tation Based on Dragonfly Algorithm”, Symmetry

2020, 12, 1256; doi:10.3390/sym12081256,

www.mdpi.com/journal/symmetry,

Xue Feng, Nicholas Tustison, Craig Meyer, “Brain Tumor

Segmentation Using an Ensemble of 3D U-Nets and

Overall Survival Prediction Using Radiomic Features”,

Springer Nature Switzerland AG 2019. A. Crimi et al.

(Eds.): BrainLes 2018, LNCS 11384, pp. 279–288,

2019. https://doi.org/10.1007/978-3-030-11726-9_25,

Andriy Myronenko, “3D MRI Brain Tumor Segmentation

Using Autoencoder Regularization, “Springer Nature

Switzerland AG 2019 A. Crimi et al. (Eds.): BrainLes

2018, LNCS 11384, pp. 311–320, 2019.

https://doi.org/10.1007/978-3-030-11726-9 _ 28.

Wei Chen, Boqiang Liu, Suting Peng, Jiawei Sun, and Xu

Qiao, “S3D-UNet: Separable 3D U-Net for Brain

Tumor Segmentation”, Springer Nature Switzerland

AG 2019, A. Crimi et al. (Eds.): BrainLes 2018, LNCS

11384, pp. 358–368, 2019. https://doi.org/10.1007/

978-3-030-11726-9_32.

Xiaojun Hu, Weijian Luo, Jiliang Hu, Sheng Guo, Weilin

Huang, Matthew R. Scott, Roland Wiest, Michael

Dahlweid and Mauricio Reyes, “Brain SegNet: 3D

local refinement network for brain lesion

segmentation”, Hu et al. BMC Medical Imaging, (2020)

20:17, https://doi.org/s12880-020-0409-2.

Li Sun, Songtao Zhang, Hang Chen, Lin Luo, “Brain Tu-

mor Segmentation and Survival Prediction Using Mul-

timodal MRI Scans With Deep Learning”, Frontiers in

Neuroscience, www.frontiersin.org, August 2019, Vol-

ume 13, Article 810.

Dmitry Lachinov, Evgeny Vasiliev, and Vadim Turlapov,

“Glioma Segmentation with Cascaded Unet”, Springer

Nature Switzerland AG 2019, A. Crimi et al. (Eds.):

BrainLes 2018, LNCS 11384, pp. 189–198, 2019,

https://doi.org/10.1007/978-3-030-11726-9, 17.

Ping Liu, Qi Dou, Qiong Wang and Pheng-Ann Heng, “An

Encoder-Decoder Neural Network With 3D Squeeze-

and-Excitation and Deep Supervisionfor Brain Tumor

Segmentation”, IEEE Access 2020, Digital Object

Identifier 10.1109/ACCESS.2020.2973707.

Pawel Mlynarski, Hervé Delingette, Antonio Criminisi, Nich-

olas Ayache, “3D convolutional neural networks for tu-

mor segmentation using long-range 2D context,

https://doi.org/10.1016/j.compmedimag.2019.02.001

0895-6111/, 2019 Elsevier Ltd.

Suting Peng, Wei Chen, Jiawei Sun, Boqiang Liu, “Multi-

Scale 3D U-Nets: An approach to automatic segmenta-

tion of brain tumor”, 2019, Wiley Periodicals, Inc.,

DOI: 10.1002/ima.22368

Mina Ghaffari, Arcot Sowmya, Ruth Oliver, Len Hamey,

“Multimodal brain tumour segmentation using densely

connected 3D convolutional neural network”, 978-1-

7281-3857-2/19/, 2019, IEEE.

Parvez Ahmad, Hai Jin, Saqib Qamar, Ran Zheng, Wenbin

Jiang, Belal Ahmad, Mohd Usama, “3D Dense Dilated

Hierarchical Architecture for Brain Tumor Segmenta-

tion”, 2019, Association for Computing Machinery,

ACM, ISBN 978-1-4503-6278-8/19/05, http://doi.org/

10.1145/3335484.3335516

Shangfeng Lu, Xutao Guo, Ting Ma, Chushu Yang, Tong

Wang, Member, Pengzheng Zhou, “Effective Multipath

Feature Extracion 3D CNN for Multimodal Brain Tu-

mor Segmentation”, 2019 International Conference on

Medical Imaging Physics and Engineering (ICMIPE).

Saqib Qamar, Hai Jin, Ran Zheng, Parvez Ahmad, “3D Hy-

per-dense Connected Convolutional Neural Network

for Brain Tumor Segmentation”, 978-1-7281-0441-

6/18/, DOI 10.1109/SKG.2018.00024.

Jing Huang and Minhua Zheng, Peter X. Liu, “Automatic

Brain Tumor Segmentation Using 3D Architecture

Based on ROI Extraction”, Proceeding of the IEEE In-

ternational Conference on Robotics and Biomimetics,

978-1-7281-6321-5/19/.

Yan Hu, Yong Xia, “3D Deep Neural Network-Based Brain

Tumor Segmentation Using Multimodality Magnetic

Resonance Sequences”, Springer International Publish-

ing AG, part of Springer Nature, 2018. A. Crimi et al.

(Eds.): BrainLes 2017, LNCS 10670, pp. 423–434,

2018. https://doi.org/10.1007/978-3-319-75238-9 _ 36.

Despotovic I., Goossens B., Philips W., MRI segmentation

of the human brain: challenges, methods, and applica-

tions. Computational and mathematical methods in

medicine, 2015.

www.tomography.org

BIOIMAGING 2022 - 9th International Conference on Bioimaging

146