A modified fcn based method for left ventricle endocardium and epicardium segmentation with new block modules

A Modified FCN Based Method for Left Ventricle Endocardium and Epicardium Segmentation with New Block Modules A modified FCN based method for Left Ventricle endocardium and epicardium segmentation wit[.]

A modified FCN-based method for Left Ventricle endocardium and epicardium segmentation with

new block modules

Do-Hai-Ninh Nham School of Applied Mathematics and Informatics

Hanoi University of Science and Technology

ninh.ndh182714@sis.hust.edu.vn

Minh-Nhat Trinh School of Electrical and Electronic Engineering Hanoi University of Science and Technology minhnhattrinh312@gmail.com

Tien-Thanh Tran

Smart Health Center

Vingroup Big Data Institute

trantienthanh081298@gmail.com

Van-Truong Pham School of Electrical and Electronic Engineering Hanoi University of Science and Technology truong.phamvan@hust.edu.vn

Thi-Thao Tran School of Electrical and Electronic Engineering

Hanoi University of Science and Technology

thao.tranthi@hust.edu.vn

Abstract—Cardiac segmentation of medical magnetic

reso-nance images has been crucial nowadays owing to its necessity

for cardiac problems diagnosis In the increasing demand of

advanced procedures for cardiac disease diagnosis and inspired

by the structure of receptive field block; in this paper, we

propose a new block module then further assembling into a

deep fully convolutional neural network to deal with automated

left ventricle segmentation With only one learning stage, our

proposed model is trained end-to-end, pixels-to-pixels and

val-idated on two popular cardiac MRI benchmarks, ACDC and

SunnyBrook datasets Several experiments have proved that our

new model architecture has a better performance than previous

segmentation methods with enhanced feature discriminability

and robustness, despite having much less training parameters

Index Terms—kernel size; dilation rate; fully convolutional

neural network; cardiac MRI segmentation

I INTRODUCTION

To interpret the cardiovascular function, cardiac magnetic

resonance imaging (MRI) has been extensively utilized

Seg-menting the left ventricle (LV) from these MRI images is

an imperative step for cardiac problem diagnosis; as many

information like systolic and diastolic issues could be sooner

observed [1] Thus there is a desperate necessity for upgrade in

LV segmentation methods nowadays As clinical approaches

could be tedious and sometimes might lead to human error

risks, developed automated approaches such as deep

learning-based approaches (DL) [2], [3], active contour models

ap-proaches (ACMs) [4], [5] have proved their efficiencies

As regards ACM methods, the level set-based active contour

model is shown to be effective on enabling the curve to update

Identify applicable funding agency here If none, delete this.

its topology throughout the segmentation process Despite revealing benefits, there are certain deficiencies that unless contour initialization has to be constructed, the final result might have less acceptable objects boundary information There is also remarkable success indicated in LV segmentation with reference to deep learning-based methods, as Tran [6] has employed a deep fully convolutional neural network (FCN) into cardiac MRI segmentation In his paper, traditional fully connected layers are replaced with convolutional layers to seize a coarse inference map, before deconvoluting it with fractional stride convolution to yield temporal coherence and discriminative features then obtain prominent results Notably methods using U-Net [7]; due to skip connections operation, which deep, coarse, semantic feature maps from the encoder

to the corresponding layers in the decoder could be allowed

to propagate, the output segmentation map is proved to be more proper Trinh et al [3] have combined the vanilla U-Net with modules of Swish activation [8] and Squeeze-and-Excite block [9] for better tunning parameters and highlighting the best salient boundary features Tran et al [10] have formulated the issue as a two-stage problem and build a hybrid algorithm, where the first phase hypothesizes raw segmentation features of all masks by a vanilla U-Net In the second phase these features are handled as initial level set functions for the multiphase active contour model (MP-ACM) before providing persistently outstanding cardiac MRI contours prediction performance However, all these previous deep network methods have resulted in high computational costs, low inference speed and borderline data insufficiency

as a consequence of strides and dilated convolution defection Motivated by these limitations, in this work, we develop

and validate a new approach for automatic contours detection

of cardiac MR images Our main contributions could be

summarized as follows:

• We propose new block modules to optimize the receptive

field on feature maps, thus aiming to enhance deep

features, improving feature detection and correct

occa-sional misalignment in MRI segmentation performance,

in condition of sensitive contour initialization in cardiac

MRI segmentation performance

• Using the backbone of FCN, we replace FCN’s

convo-lutional block modules by our proposed modules before

training the model end-to-end

• We show that our new approach achieves state-of-the-art

results on the ACDC and SunnyBrook datasets at a real

time processing speed with only one training stage, and

demonstrate the effectiveness of our notably light-weight

model to other models

II RELATEDWORK

Fully Convolutional Neural Network (FCN) Previously,

FCN has been exploited in many proposals and applications

With fully convolutional inference, a convolutional network

has been exemplified for raw multi-class segmentation of

particular tissues [11] Also within a convolutional network,

multi-scale and sliding window technique could be constructed

for classification, localization and even detection [12]

Evolv-ing from [12], FCN has been adopted as a segmentation

method [13], which includes FCN-8s, FCN-16s and FCN-32s

FCN-8s and FCN-16s are validated to have more

outstand-ing segmentation performances, as the FCN-32s outcome is

rougher due to the loss of spatial location and local

informa-tion Furthermore, FCN in [6] highlighted the first time ever to

be employed in pixel-wise labelling or per-pixel classification,

benchmarked on cardiac MRI datasets

Receptive Field In neuroscience, receptive field is a region

in the sensory periphery within which stimuli can influence the

electrical activity of sensory cells [15] Similarly, in a deep

learning context, the Receptive Field (RF) is basically defined

as a measure of association of an output feature of any layer to

the input region [16] Inception architecture [17] has overcome

the hurdles of huge variation locaction, being overfitted and

expensive computatation cost by putting multiple RFs sizes

(1 × 1, 3 × 3, 5 × 5) into use Thus the network structure

would essentially get a bit wider rather than deeper Inception

variants [18], [19] have factorized convolutions of filter size

n × n into a combination of 1 × n and n × 1

convolu-tions; therefore attaining promising performances on object

detection and classification In the Deformable Convolutional

Network (DCN) [20], several RF sizes of deformable filters

are correlated with object scales and shapes, indicating that

the deformation is efficiently learned from input image or

feature maps Compared with larger but fixed dilation value

of atrous convolution [21], deformable convolution utilizes

different dilation values applied to each point in the grid

during convolution In addition, DCN is also reported to be

an extremely light-weight Spatial Transformer Network [22];

hence DCN has won as the runner-up in the COCO Detection

as well as Segmentation Challenge

III METHODOLOGY

A The Proposed Block Module

In general, assuming that the input shape is nh× nw and the convolutional kernel shape is kh× kw, the output shape will be (nh− kh+ 1) × (nw− kw+ 1) Accordingly, the output shape of the convolutional layer is affected by the shape of the input and the shape of the convolution kernel Convolutional kernels with odd height and width values (such as 1, 3, 5, or 7) are frequently utilized; which the spatial dimensionality could

be preserved while padding with the same number of rows

on top and bottom, and the same number of columns on left and right This usage of odd kernels and padding to precisely preserve dimensionality reveals certain efficiency; that for any two-dimensional tensor X, when the kernel size is odd and the padding rows and columns number of all sides are all similar, obtaining an output with equal height and width as the input, we know that the output y[i, j] is determined by cross-correlation of the input and convolutional kernel with the window centered on x[i, j] [23]

The question is, how to decide the most optimized kernels for convolutional operation? Salient parts of the input image could have substantially wide variation in shape and size Ow-ing to this, adoptOw-ing the precise kernel size for convolutional operation has become a tough issue If a smaller kernel is adopted for locally distributed information, a larger one is preferred for more global features Very deep network makes

it hard for gradient to be passed during updating through the whole network; also, naively stacking large convolutional operations is computationally expensive Consequently, con-volutional filters with several shapes on the same level need

to be taken into account

As above, asymmetric n×n convolution could be performed

as the integration of 1×n and n×1 convolutions For instance,

a 3 × 3 convolution is proved to be equivalent to a 1 × 3 then

3 × 1 convolution when their computational cost of is proved

to be 33% cheaper Thus, instead of employing 2 consecutive

3 × 3, 5 × 5, 7 × 7 and 9 × 9 operations, we consider a merger

of 1 × 3 and 3 × 1, 1 × 5 and 5 × 1, 1 × 7 and 7 × 1, 1 × 9 and

9 × 1 respectively; before applying an asymmetric convolution operation

One primary obstacle with the above modules is that even

a modest number of 5 × 5 and 7 × 7 convolutions could

be prohibitively expensive on top of a convolutional layer with a large number of filters A 1 × 1 convolutional layer is judiciously exploited to overcome this dispute, as it could offer

a channel-wise pooling, also called feature map pooling or a projection layer Such simple technique with low dimensional embeddings could be utilized in dimensionality reduction whilst retaining salient features; also, generating a one-to-one projection of stack of feature maps to pool features across channels after conventional pooling layers

To obtain a promising performance on MRI segmentation, convolutional procedures have to analyze both globally and

locally distributed features If familiar discrete convolutions

are applied throughout the network architecture, it will be

nec-essary to assign large kernels in order to achieve global view,

which is responsible for parameters surging Therefore, we

adopt dilated convolutions [24] at the asymmetric convolution

layers as it supports RF exponential expansion without loss of

resolution, while the parameters number increases linearly, to

control the features’ eccentricities

Between two consecutive layers, an activation function

and a Mean-Variance normalization layer (MVN) [6] are

used for humbling the pixel distribution shifting right after

a convolutional operation Compare with Batch Normalization

[14], which reduces internal covariate shift and accelerates the

gradient flow through the network like MVN; MVN is still

effective though being much simpler as it primarily centers

and standardizes a single batch at a time About the activation,

instead of using ReLU activation; Swish [8] is selected as the

only activation function throughout the training

Overall, our proposed block modules could be briefly

de-scribed as in Fig 1 We create 4 block modules In block

Fig 1 Our proposed block module (n = 2i + 1; i ∈ [1, 4])

module i, (i ∈ [1, 4]), there are 4 convolutional blocks, whose

kernel sizes are (1, 1), (1, n), (n, 1), (n, n) respectively that

n = 2i + 1 All the dilation rates are set up to be equal to

(1, 1) except for the last one, which is (n − 2, n − 2)

B Model Architecture

Assembling the new module to the FCN backbone [6], we

construct an advanced one-stage segmentation model Even

though having such a lightweight module, our modified FCN

delivers relatively comparable results to some of the

start-of-the-art approaches and moreover retaining a very fast speed

Our new 16-layer model architecture is displayed as Fig 2

Fig 2 Our modified FCN based on the new receptive field block module in Fig 1

C Loss function Sigmoid activation is employed to form the loss and the binary image from the output layer, which contains c = 2 labels standing for the total number of classes for each pixel

to be classified into Denoting there are totally N pixels for prediction as well as groundtruth, P and L be the predicted set and the groundtruth set respectively which |P | = |L| = N

Pic and Lic are the element of P and L in order, with condition that i ∈ {1, 2, , N } and c ∈ {0, 1}; Pic ∈ [0, 1]; Lic∈ {0, 1} displaying ground truth label and predicted label probability correspondingly

Dice Loss is our loss function usage in the training process,

it is calculated as:

Dice Loss =

PN i=1(Pic+ Lic) − 2PicLic

PN i=1(Pic+ Lic) + ϵ (1)

D Evaluation Metrics

In medical image analysis, while the Dice Similarity Co-efficient (DSC) is a statistical tool for measuring similarity between segmentation maps, the Intersection over Union index (IoU) is also a statistical tool but to gauge the similarity and diversity of sample pixel sets They are determined by:

DSC(P, L) =

PN i=12PicLic+ ϵ

PN i=1(Pic+ Lic) + ϵ (2) IoU (P, L) =

PN i=1PicLic+ ϵ

PN i=1(Pic+ Lic− PicLic) + ϵ (3)

The smooth coefficient ϵ is provided for preventing zero

division, in experiment we suppose ϵ to be 1e − 15

In a diagnostic test, while sensitivity (SEN) is determined

as the rate of people who are diagnosed as positive out

of those who actually have the condition, specificity (SPE)

is determined as the rate of people who are diagnosed as

negative Mathematically, sensitivity=number of true

posi-tives (TP)/(number of true posiposi-tives (TP) + number of false

negatives (FN)) and specificity=number of true negatives

(TN)/(number of true negatives (TN) + number of false

positives (FP)):

SP E = T N

Accuracy (ACC) is also used as a statistical measure of how

well a binary classification test correctly identifies or excludes

a condition:

T P + T N + F P + F N (6)

IV EXPERIMENTALRESULTS

A Benchmarks

There are 45 cine-MRI images in the Sunnybrook dataset

[25] (2009 Cardiac MR Left Ventricle Segmentation Challenge

data), comprising a mixed of patients and pathologies: healthy,

hypertrophy, heart failure with infarction and heart failure

without infarction While a part of this dataset was first used

in the automated myocardium segmentation challenge from

short-axis MRI, held by a MICCAI workshop in 2009; the

entire dataset is now available in the CAP database with public

domain license There are three different subsets of 15 cases

each: training, validation, and online with groundtruth

respec-tively The training set is utilized for training model for LV

endocardium and epicardium segmentation, when validation

and online sets would be used for evaluation

Having been released by the ”Automatic Cardiac Diagnosis

Challenge (ACDC)” workshop held in conjunction with the

20th International Conference on Medical Image Computing

and Computer Assisted Intervention (MICCAI), the public

ACDC dataset [26] comprises 100 patient 4D cine-CMR scans

each consisting of segmentation masks for the left ventricle

(LV), the myocardium (Myo) and the right ventricle (RV) at

the end-systolic (ES) and end-diastolic (ED) phases of each

patient Then the training database containing the manual

segmentation masks is separated into the training set and

test set with ratio 8:2 to assess the proposed model, before

comparing the performance between several automatic

meth-ods on the segmentation of the left ventricular endocardium

and epicardium for both end diastolic and end systolic phase

instances

B Training

We have performed our customized FCN to segment cardiac

left ventricles on MRIs It is necessary to note that in each

segmentation task of each dataset, we break down the task into

2 sub-tasks respectively for endo- and epi-cardium contour prediction Our models are trained end-to-end, with cost minimization on various epochs (base on different cases) is operated by applying NADAM optimizer [27] with an original learning rate of 0.001 Learning rate is multiplied by 0.5 every 10 epochs, before reaching 0.00001 and being constantly kept through the remainder training period All images are all rigidly pre-processed by cropping center and normalization, before being augmented by randomly flipping horizontally and vertically, and rotated within π radians, before being trained with NVIDIA Tesla P100 16GB GPU

C Evaluation

TABLE I

ENDOCARDIUM OF S UNNY B ROOK D ATASET

Trinh [3] 32.4M 0.7850 0.6996 0.8991 0.9920 0.9746 FCN[1][6] 10.9M 0.9156 0.8537 0.9336 0.9934 0.9866

-U-Net [7] 31.0M 0.6123 0.5438 0.7294 0.9914 0.9731

-Ours 9.1M 0.9196 0.8580 0.9314 0.9942 0.9870

[1] w/o finetune and Xavier unit.

TABLE II

OF S UNNY B ROOK D ATASET

Trinh [3] 32.4M 0.8620 0.7940 0.9277 0.9843 0.9726 FCN [1] [6] 10.9M 0.9466 0.9009 0.9542 0.9886 0.9814

-U-Net [7] 31.0M 0.6896 0.6347 0.7150 0.9921 0.9568

-Ours 9.1M 0.9505 0.9100 0.9507 0.9910 0.9826

[1] w/o finetune and Xavier unit.

We compare our algorithm with several algorithms to eval-uate the efficiency of our new approach and write down the mean value of each metric index into 4 tables Ablation test results recorded in the Table I and Table II show that our proposed method is confirmed to be the best on endocardium segmentation of the SunnyBrook dataset according to testing measurement on DSC (0.9196), IoU (0.8580), SPE (0.9942) and ACC (0.9870), also on epicardium segmentation of the SunnyBrook dataset with DSC (0.9505), IoU (0.9100), SPE (0.9910) and ACC (0.9826) This indicates that our proposed method is more convinced of the predicted regions and pro-vides more accurate boundary maps, even with exceptionally light-weighted backbone (9.1 million trainable parameters), comparing with Tran et al [10] (which is a two-stage ar-chitecture) and other popular architectures

TABLE III

FCN [1] [6] 10.9M 0.8576 0.8047 0.9272 0.9810 0.9798

-Trinh [3] 32.4M 0.9253 0.8964 0.9661 0.9966 0.9940

SegNet [28] 29.5M 0.8192 0.7470 0.9300 0.9829 0.9783

U-Net [7] 31.0M 0.8810 0.8225 0.9504 0.9909 0.9920

Ours 9.1M 0.8798 0.8360 0.9512 0.9852 0.9823

[1] w/o finetune and Xavier unit. [2]with finetune and Xavier unit.

TABLE IV

OF ACDC D ATASET

FCN [1] [6] 10.9M 0.9167 0.8803 0.9631 0.9947 0.9918

-SegNet [28] 29.5M 0.8896 0.8321 0.9526 0.9942 0.9907

U-Net [7] 31.0M 0.9189 0.8712 0.9618 0.9945 0.9918

Ours 9.1M 0.9273 0.8978 0.9656 0.9969 0.9943

[1] w/o finetune and Xavier unit. [2]with finetune and Xavier unit.

To present a fairer verification of our proposed approach,

corresponding outcomes of ours and five different approaches

have been displayed in Table III and Table IV on the

bench-mark of the ACDC dataset From Table IV, our proposed

method is confirmed to outperform other compared methods

on epicardium segmentation of the ACDC dataset according

to testing measurement on DSC (0.9273), IoU (0.8978), SPE

(0.9969) and ACC (0.9943), which is very comparable to some

of the state-of-the-art approaches Nevertheless, with regards

to the endocardium segmentation in Table III, Trinh et al [3]

has gained the best results with Dice score of 0.9253, while

our result stays modest with Dice score of 0.8798

D Representative Results

In this section, we provide quantitative results produced by

our proposed model for difficult input cases As indicated in

Fig 3-5, our simple model can achieve better results than

some of existing methods; which verifies the strength and

light-weighting of our proposed method have obtained quite a

comparable performance on inputs with high complexity

From the results on figures and tables, we demonstrate other

fundamental observations that:

• U-Net-based networks seem to have poorer performances

than FCN-based networks do on the SunnyBrook dataset,

as vanilla U-Net, in some particular cases it could not

provide an acceptable prediction as Fig 3 and Fig 4

have illustrated

• In contrast, FCN-based methods seem to have more

mediocre results than U-Net-based methods However,

our FCN-based performance on the epicardium of ACDC

Fig 3 Endocardium contour prediction on the SunnyBrook dataset between different models in some complex cases

Fig 4 Epicardium contour prediction on the SunnyBrook dataset between different models in some complex cases

Fig 5 Epicardium contour prediction on the ACDC dataset between different models in some complex cases

dataset has verified our approach’s convincement on pro-viding comparable results for 2 different datasets (which very little number of algorithms could do and no light-weighted models could do), as shown in Fig 5 as our light-weighted network still outperforms vanilla U-Net on

a side of the ACDC dataset segmentation

V CONCLUSION

In this paper, we have introduced a new light-weighted deep-learning framework for cardiac MRI contours prediction Our experimental outcomes through the training process

ev-idently express that our new model has achieved prominent

performances on the two popular datasets SunnyBrook and

ACDC Even though, those are just small datasets in the deep

learning era Therefore, in the future, we will dedicate more

effort on complex MRI segmentation datasets to minimize

the segmentation errors of the deep-learning approaches and

maximize the acceleration, efficiency and reliability of the

deep-learning network

ACKNOWLEDGEMENT

This research is funded by the Hanoi University of Science

and Technology (HUST) under project number T2021-PC-005

Minh-Nhat Trinh was funded by Vingroup JSC and supported

by the Master, PhD Scholarship Programme of Vingroup

Innovation Foundation (VINIF), Institute of Big Data, code

VINIF.2021.ThS.33

REFERENCES [1] Miller, Christopher and Pearce, Keith and Jordan, Peter and Argyle,

Rachel and Clark, David and Stout, Martin and Ray, Simon and Schmitt,

Matthias, “Comparison of real-time three-dimensional echocardiography

with cardiovascular magnetic resonance for left ventricular volumetric

assessment in unselected patients” European heart journal cardiovascular

Imaging, vol 13, pp 187–195, doi 10.1093/ejechocard/jer248,

Novem-ber 2011.

[2] M R Avendi and A Kheradvar and H Jafarkhani, A Combined

Deep-Learning and Deformable-Model Approach to Fully Automatic

Segmentation of the Left Ventricle in Cardiac MRI, eprint 1512.07951,

archivePrefix, primaryClass cs.CV, 2015.

[3] M.N Trinh and N.T Nguyen and T.T Tran and V.T Pham, A Deep

Learning-based Approach with Image-driven Active Contour Loss for

Medical Image Segmentation, The 2nd International Conference on Data

Science and Applications (ICDSA 2021).

[4] Lynch, Michael and Ghita, Ovidiu and Whelan, Paul, Left-ventricle

myocardium segmentation using a coupled level-set with a priori

knowl-edge, Computerized Medical Imaging and Graphics, 30 (4) pp 255-262.

ISSN 0895-6111, doi 10.1016/j.compmedimag.2006.03.009, July 2006.

[5] Pham, Van-Truong and Tran, Thi-Thao, Active Contour Model and

Nonlinear Shape Priors with Application to Left Ventricle Segmentation

in Cardiac MR Images, Optik - International Journal for Light and

Electron Optics, vol 127, doi 10.1016/j.ijleo.2015.10.162, November

2015.

[6] Phi Vu Tran, A Fully Convolutional Neural Network for Cardiac

Seg-mentation in Short-Axis MRI, eprint 1604.00494, archivePrefix arXiv,

primaryClass cs.CV, 2017.

[7] Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas, U-Net:

Convolutional Networks for Biomedical Image Segmentation, LNCS,

vol 9351, pp 234–241, isbn 978-3-319-24573-7, doi

10.1007/978-3-319-24574-4 28, October 2015.

[8] Ramachandran, Prajit and Zoph, Barret and Le, Quoc, Swish: a

Self-Gated Activation Function, October 2017.

[9] Jie Hu and Li Shen and Samuel Albanie and Gang Sun and Enhua

Wu, Squeeze-and-Excitation Networks, eprint 1709.01507, archivePrefix

arXiv, primaryClass cs.CV, 2019.

[10] Tran, Tien-Thanh and Tran, Thi-Thao and Ninh, Quoc-Cuong and

Bui, Minh-Duc and Pham, Van-Truong, Segmentation of Left Ventricle

in Short-Axis MR Images Based on Fully Convolutional Network

and Active Contour Model, pp 49–59, isbn 978-3-030-62323-4, doi

10.1007/978-3-030-62324-1 5, October 2020.

[11] Ning, Feng and Delhomme, Damien and Lecun, Yann and Piano, Fabio

and Bottou, L´eon and Barbano, Paolo, Toward automatic phenotyping of

developing embryos from videos, Image Processing, IEEE Transactions

on, vol 14, pp 1360–1371, doi 10.1109/TIP.2005.852470, October 2005.

[12] Sermanet, Pierre and Eigen, David and Zhang, Xiang and Mathieu,

Michael and Fergus, Rob and Lecun, Yann, OverFeat: Integrated

Recognition, Localization and Detection using Convolutional Networks,

International Conference on Learning Representations (ICLR) (Banff),

December 2013.

[13] Long, Jonathan and Shelhamer, Evan and Darrell, Trevor, Fully con-volutional networks for semantic segmentation, pp 3431–3440, doi 10.1109/CVPR.2015.7298965, June 2015.

[14] Sergey Ioffe and Christian Szegedy, Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift, eprint 1502.03167, archivePrefix arXiv, primaryClass cs.LG, 2015.

[15] Levitt and Jonathan B., Receptive field, accessed: 01.10.2021, https://www.britannica.com/science/receptive-field.

[16] Nikolas Adaloglou, Understanding the receptive field

of deep convolutional networks, accessed: 02.07.2020, https://theaisummer.com/receptive-field/.

[17] Christian Szegedy and Wei Liu and Yangqing Jia and Pierre Sermanet and Scott Reed and Dragomir Anguelov and Dumitru Erhan and Vincent Vanhoucke and Andrew Rabinovich, Going Deeper with Convolutions, eprint 1409.4842, archivePrefix arXiv, primaryClass cs.CV, 2014 [18] Christian Szegedy and Vincent Vanhoucke and Sergey Ioffe and Jonathon Shlens and Zbigniew Wojna, Rethinking the Inception Archi-tecture for Computer Vision, eprint 1512.00567, archivePrefix arXiv, primaryClass cs.CV, 2015.

[19] Christian Szegedy and Sergey Ioffe and Vincent Vanhoucke and Alex Alemi, Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning, eprint 1602.07261, archivePrefix arXiv, pri-maryClass cs.CV, 2016.

[20] Jifeng Dai and Haozhi Qi and Yuwen Xiong and Yi Li and Guodong Zhang and Han Hu and Yichen Wei, Deformable Convolutional Net-works, eprint 1703.06211, archivePrefix arXiv, primaryClass cs.CV, 2017.

[21] Liang-Chieh Chen and George Papandreou and Iasonas Kokkinos and Kevin Murphy and Alan L Yuille, DeepLab: Semantic Image Segmen-tation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs, eprint 1606.00915, archivePrefix arXiv, primaryClass cs.CV, 2017.

[22] Sik-Ho Tsang, Review: DCN / DCNv1 — Deformable Convolutional Networks, 2nd Runner Up in 2017 COCO Detection (Object Detection), accessed: 02.02.2019, https://towardsdatascience.com/review-dcn- deformable-convolutional-networks-2nd-runner-up-in-2017-coco-detection-object-14e488efce44.

[23] Aston Zhang and Zack C Lipton and Mu Li and Alex J Smola, Dive into Deep Learning, https://d2l.ai/chapter convolutional-neural-networks/padding-and-strides.html.

[24] Liang-Chieh Chen and George Papandreou and Florian Schroff and Hartwig Adam, Rethinking Atrous Convolution for Semantic Image Segmentation, eprint 1706.05587, archivePrefix arXiv, primaryClass cs.CV, 2017.

[25] Krasnobaev, Arseny and Sozykin, Andrey, An Overview of Techniques for Cardiac Left Ventricle Segmentation on Short-Axis MRI, ITM Web

of Conferences, vol 8, pp 01003, doi 10.1051/itmconf/20160801003, January 2016.

[26] Bernard, Olivier and Lalande, Alain and Zotti, Clement and Cerve-nansky, Frederick and Yang, Xin and Heng, Pheng-Ann and Cetin, Irem and Lekadir, Karim and Camara, Oscar and Gonz´alez Ballester, Miguel ´ Angel and Sanroma, Gerard and Napel, Sandy and Petersen, Steffen and Tziritas, Georgios and Ilias, Grinias and Khened, Mahendra and Kollerathu, Varghese and Krishnamurthi, Ganapathy and Rohe, Marc-Michel and Jodoin, Pierre-Marc, Deep Learning Techniques for Automatic MRI Cardiac Multi-Structures Segmentation and Diagnosis:

Is the Problem Solved?, IEEE Transactions on Medical Imaging, vol.

PP, doi 10.1109/TMI.2018.2837502, May 2018.

[27] Timothy Dozat, Incorporating Nesterov Momentum into Adam, 2016 [28] Vijay Badrinarayanan and Alex Kendall and Roberto Cipolla, SegNet:

A Deep Convolutional Encoder-Decoder Architecture for Image Seg-mentation, eprint 1511.00561, archivePrefix arXiv, primaryClass cs.CV, 2016.

[29] Queir´os, Sandro and Barbosa, Daniel and Heyde, Brecht and Morais, Pedro and Vilac¸a, Jo˜ao and Friboulet, Denis and Bernard, Olivier and D’hooge, Jan, Fast Automatic Myocardial Segmentation in 4D cine CMR datasets, Medical Image Analysis, vol 18, doi 10.1016/j.media.2014.06.001, October 2014.

[30] Hu, Huaifei and Liu, Haihua and Gao, Zhiyong and Huang, Lu, Hybrid segmentation of left ventricle in cardiac MRI using Gaussian-mixture model and region restricted dynamic programming, Magnetic resonance imaging, vol 31, doi 10.1016/j.mri.2012.10.004, December 2012.