Recurrent Residual Convolutional Neural

Network based on U-Net (R2U-Net) for

Medical Image Segmentation
Md Zahangir Alom1*, Student Member, IEEE, Mahmudul Hasan2, Chris Yakopcic1, Member, IEEE,
Tarek M. Taha1, Member, IEEE, and Vijayan K. Asari1, Senior Member, IEEE

 are available for training CNN models [1]. However, in most

Abstract—Deep learning (DL) based semantic segmentation cases, models are explored and evaluated using classification
methods have been providing state-of-the-art performance in the tasks on very large-scale datasets like ImageNet [1], where the
last few years. More specifically, these techniques have been outputs of the classification tasks are single label or probability
successfully applied to medical image classification, segmentation,
and detection tasks. One deep learning technique, U-Net, has
values. Alternatively, small architecturally variant models are
become one of the most popular for these applications. In this used for semantic image segmentation tasks. For example, a
paper, we propose a Recurrent Convolutional Neural Network fully-connected convolutional neural network (FCN) also
(RCNN) based on U-Net as well as a Recurrent Residual provides state-of-the-art results for image segmentation tasks in
Convolutional Neural Network (RRCNN) based on U-Net models, computer vision [2]. Another variant of FCN was also proposed
which are named RU-Net and R2U-Net respectively. The proposed which is called SegNet [10].
models utilize the power of U-Net, Residual Network, as well as
RCNN. There are several advantages of these proposed
architectures for segmentation tasks. First, a residual unit helps
when training deep architecture. Second, feature accumulation
with recurrent residual convolutional layers ensures better feature
representation for segmentation tasks. Third, it allows us to design
better U-Net architecture with same number of network
parameters with better performance for medical image
segmentation. The proposed models are tested on three
benchmark datasets such as blood vessel segmentation in retina
images, skin cancer segmentation, and lung lesion segmentation.
The experimental results show superior performance on
segmentation tasks compared to equivalent models including U-
Net and residual U-Net (ResU-Net).
Fig. 1. Medical image segmentation: retina blood vessel segmentation in the
Index Terms—Medical imaging, Semantic segmentation, left, skin cancer lesion segmentation, and lung segmentation in the right.
Convolutional Neural Networks, U-Net, Residual U-Net, RU-Net,
and R2U-Net. Due to the great success of DCNNs in the field of computer
vision, different variants of this approach are applied in
different modalities of medical imaging including
I. INTRODUCTION segmentation, classification, detection, registration, and

N OWADAYS DL provides state-of-the-art performance for

image classification [1], segmentation [2], detection and
tracking [3], and captioning [4]. Since 2012, several Deep
medical information processing. The medical imaging comes
from different imaging techniques such as Computer
Tomography (CT), ultrasound, X-ray, and Magnetic Resonance
Convolutional Neural Network (DCNN) models have been Imaging (MRI). The goal of Computer-Aided Diagnosis (CAD)
proposed such as AlexNet [1], VGG [5], GoogleNet [6], is to obtain a faster and better diagnosis to ensure better
Residual Net [7], DenseNet [8], and CapsuleNet [9][65]. A DL treatment of a large number of people at the same time.
based approach (CNN in particular) provides state-of-the-art Additionally, efficient automatic processing without human
performance for classification and segmentation tasks for involvement to reduce human error and also reduces overall
several reasons: first, activation functions resolve training time and cost. Due to the slow process and tedious nature of
problems in DL approaches. Second, dropout helps regularize
the networks. Third, several efficient optimization techniques

Md Zahangir Alom1*, Chris Yakopcic1, Tarek M. Taha1, and Vijayan K. Mahmudul Hasan2, is with Comcast Labs, Washington, DC, USA. (e-mail:
Asari1 are with the University of Dayton, 300 College Park, Dayton, OH, [email protected]).
45469, USA. (e-mail: {alomm1, cyakopcic1, ttaha1, vasari1}@udayton.edu).
 Fig. 2. U-Net architecture consisted with convolutional encoding and decoding units that take image as input and produce the segmentation feature maps with
respective pixel classes.

manual segmentation approaches, there is a significant demand as universal learning approaches, where a single model can be
for computer algorithms that can do segmentation quickly and utilized efficiently in different modalities of medical imaging
accurately without human interaction. However, there are some such as MRI, CT, and X-ray.
limitations of medical image segmentation including data According to a recent survey, DL approaches are applied to
scarcity and class imbalance. Most of the time the large number almost all modalities of medical imagining [20, 21].
of labels (often in the thousands) for training is not available for Furthermore, the highest number of papers have been published
several reasons [11]. Labeling the dataset requires an expert in on segmentation tasks in different modalities of medical
this field which is expensive, and it requires a lot of effort and imaging [20, 21]. A DCNN based brain tumor segmentation and
time. Sometimes, different data transformation or augmentation detection method was proposed in [22].
techniques (data whitening, rotation, translation, and scaling) From an architectural point of view, the CNN model for
are applied for increasing the number of labeled samples classification tasks requires an encoding unit and provides class
available [12, 13, and 14]. In addition, patch based approaches probability as an output. In classification tasks, we have
are used for solving class imbalance problems. In this work, we performed convolution operations with activation functions
have evaluated the proposed approaches on both patch-based followed by sub-sampling layers which reduces the
and entire image-based approaches. However, to switch from dimensionality of the feature maps. As the input samples
the patch-based approach to the pixel-based approach that traverse through the layers of the network, the number of
works with the entire image, we must be aware of the class feature maps increases but the dimensionality of the feature
imbalance problem. In the case of semantic segmentation, the maps decreases. This is shown in the first part of the model (in
image backgrounds are assigned a label and the foreground green) in Fig. 2. Since, the number of feature maps increase in
regions are assigned a target class. Therefore, the class the deeper layers, the number of network parameters increases
imbalance problem is resolved without any trouble. Two respectively. Eventually, the Softmax operations are applied at
advanced techniques including cross-entropy loss and dice the end of the network to compute the probability of the target
similarity are introduced for efficient training of classification classes.
and segmentation tasks in [13, 14]. As opposed to classification tasks, the architecture of
Furthermore, in medical image processing, global segmentation tasks requires both convolutional encoding and
localization and context modulation is very often applied for decoding units. The encoding unit is used to encode input
localization tasks. Each pixel is assigned a class label with a images into a larger number of maps with lower dimensionality.
desired boundary that is related to the contour of the target The decoding unit is used to perform up-convolution (de-
lesion in identification tasks. To define these target lesion convolution) operations to produce segmentation maps with the
boundaries, we must emphasize the related pixels. Landmark same dimensionality as the original input image. Therefore, the
detection in medical imaging [15, 16] is one example of this. architecture for segmentation tasks generally requires almost
There were several traditional machine learning and image double the number of network parameters when compared to
processing techniques available for medical image the architecture of the classification tasks. Thus, it is important
segmentation tasks before the DL revolution, including to design efficient DCNN architectures for segmentation tasks
amplitude segmentation based on histogram features [17], the which can ensure better performance with less number of
region based segmentation method [18], and the graph-cut network parameters.
approach [19]. However, semantic segmentation approaches This research demonstrates two modified and improved
that utilize DL have become very popular in recent years in the segmentation models, one using recurrent convolution
field of medical image segmentation, lesion detection, and networks, and another using recurrent residual convolutional
localization [20]. In addition, DL based approaches are known networks. To accomplish our goals, the proposed models are
 Fig. 3. RU-Net architecture with convolutional encoding and decoding units using recurrent convolutional layers (RCL) based U-Net architecture. The residual
units are used with RCL for R2U-Net architecture.

evaluated on different modalities of medical imagining as segmentation [2]. One of the image patch-based architectures is
shown in Fig. 1. The contributions of this work can be called Random architecture, which is very computationally
summarized as follows: intensive and contains around 134.5M network parameters.
The main drawback of this approach is that a large number of
1) Two new models RU-Net and R2U-Net are introduced for
pixel overlap and the same convolutions are performed many
medical image segmentation.
2) The experiments are conducted on three different times. The performance of FCN has improved with recurrent
modalities of medical imaging including retina blood vessel neural networks (RNN), which are fine-tuned on very large
segmentation, skin cancer segmentation, and lung datasets [27]. Semantic image segmentation with DeepLab is
segmentation. one of the state-of-the-art performing methods [28]. SegNet
3) Performance evaluation of the proposed models is consists of two parts, one is the encoding network which is a
conducted for the patch-based method for retina blood vessel 13-layer VGG16 network [5], and the corresponding decoding
segmentation tasks and the end-to-end image-based approach network uses pixel-wise classification layers. The main
for skin lesion and lung segmentation tasks. contribution of this paper is the way in which the decoder up-
4) Comparison against recently proposed state-of-the-art samples its lower resolution input feature maps [10]. Later, an
methods that shows superior performance against equivalent improved version of SegNet, which is called Bayesian SegNet
models with same number of network parameters. was proposed in 2015 [29]. Most of these architectures are
The paper is organized as follows: Section II discusses related explored using computer vision applications. However, there
work. The architectures of the proposed RU-Net and R2U-Net are some deep learning models that have been proposed
models are presented in Section III. Section IV, explains the specifically for the medical image segmentation, as they
datasets, experiments, and results. The conclusion and future consider data insufficiency and class imbalance problems.
direction are discussed in Section V. One of the very first and most popular approaches for
semantic medical image segmentation is called “U-Net” [12].
II. RELATED WORK A diagram of the basic U-Net model is shown in Fig. 2.
According to the structure, the network consists of two main
Semantic segmentation is an active research area where
DCNNs are used to classify each pixel in the image parts: the convolutional encoding and decoding units. The basic
convolution operations are performed followed by ReLU
individually, which is fueled by different challenging datasets
activation in both parts of the network. For down sampling in
in the fields of computer vision and medical imaging [23, 24,
and 25]. Before the deep learning revolution, the traditional the encoding unit, 2×2 max-pooling operations are performed.
In the decoding phase, the convolution transpose (representing
machine learning approach mostly relied on hand engineered
up-convolution, or de-convolution) operations are performed to
features that were used for classifying pixels independently. In
up-sample the feature maps. The very first version of U-Net was
the last few years, a lot of models have been proposed that have
used to crop and copy feature maps from the encoding unit to
proved that deeper networks are better for recognition and
segmentation tasks [5]. However, training very deep models is the decoding unit. The U-Net model provides several
advantages for segmentation tasks: first, this model allows for
difficult due to the vanishing gradient problem, which is
the use of global location and context at the same time. Second,
resolved by implementing modern activation functions such as
Rectified Linear Units (ReLU) or Exponential Linear Units it works with very few training samples and provides better
performance for segmentation tasks [12]. Third, an end-to-end
(ELU) [5,6]. Another solution to this problem is proposed by
He et al., a deep residual model that overcomes the problem pipeline process the entire image in the forward pass and
directly produces segmentation maps. This ensures that U-Net
utilizing an identity mapping to facilitate the training process
preserves the full context of the input images, which is a major
[26].
In addition, CNNs based segmentation methods based on advantage when compared to patch-based segmentation
approaches [12, 14].
FCN provide superior performance for natural image
 However, U-Net is not only limited to the applications in the are named RU-Net and R2U-Net. These two approaches utilize
domain of medical imaging, nowadays this model is massively the strengths of all three recently developed deep learning
applied for computer vision tasks as well [30, 31]. Meanwhile, models. RCNN and its variants have already shown superior
different variants of U-Net models have been proposed, performance on object recognition tasks using different
including a very simple variant of U-Net for CNN-based benchmarks [42, 43]. The recurrent residual convolutional
segmentation of Medical Imaging data [32]. In this model, two operations can be demonstrated mathematically according to
modifications are made to the original design of U-Net: first, a the improved-residual networks in [43]. The operations of the
combination of multiple segmentation maps and forward Recurrent Convolutional Layers (RCL) are performed with
feature maps are summed (element-wise) from one part of the respect to the discrete time steps that are expressed according
network to the other. The feature maps are taken from different to the RCNN [41]. Let’s consider the 𝑥𝑙 input sample in the 𝑙 𝑡ℎ
layers of encoding and decoding units and finally summation layer of the residual RCNN (RRCNN) block and a pixel located
(element-wise) is performed outside of the encoding and at (𝑖, 𝑗) in an input sample on the kth feature map in the RCL.
𝑙
decoding units. The authors report promising performance Additionally, let’s assume the output of the network 𝑂𝑖𝑗𝑘 (𝑡) is
improvement during training with better convergence at the time step t. The output can be expressed as follows as:
compared to U-Net, but no benefit was observed when using a
𝑓 𝑇 𝑓(𝑖,𝑗)
summation of features during the testing phase [32]. However, 𝑙 (𝑡)
𝑂𝑖𝑗𝑘 = (𝑤𝑘 ) ∗ 𝑥𝑙 (𝑡) + (𝑤𝑘𝑟 )𝑇 ∗ 𝑥𝑙𝑟(𝑖,𝑗) (𝑡 − 1) + 𝑏𝑘 (1)
this concept proved that feature summation impacts the 𝑓(𝑖,𝑗)
Here 𝑥𝑙 (𝑡) and 𝑥𝑙𝑟(𝑖,𝑗) (𝑡 − 1) are the inputs to the
performance of a network. The importance of skipped
connections for biomedical image segmentation tasks have standard convolution layers and for the 𝑙 𝑡ℎ RCL respectively.
𝑓
been empirically evaluated with U-Net and residual networks The 𝑤𝑘 and 𝑤𝑘𝑟 values are the weights of the standard
[33]. A deep contour-aware network called Deep Contour- convolutional layer and the RCL of the kth feature map
Aware Networks (DCAN) was proposed in 2016, which can respectively, and 𝑏𝑘 is the bias. The outputs of RCL are fed to
extract multi-level contextual features using a hierarchical the standard ReLU activation function 𝑓 and are expressed:
architecture for accurate gland segmentation of histology 𝑙 𝑙
ℱ(𝑥𝑙 , 𝑤𝑙 ) = 𝑓(𝑂𝑖𝑗𝑘 (𝑡)) = max(0, 𝑂𝑖𝑗𝑘 (𝑡)) (2)
images and shows very good performance for segmentation
[34]. Furthermore, Nabla-Net: a deep dig-like convolutional ℱ(𝑥𝑙 , 𝑤𝑙 ) represents the outputs from of lth layer of the
architecture was proposed for segmentation in 2017 [35]. RCNN unit. The output of ℱ(𝑥𝑙 , 𝑤𝑙 ) is used for down-sampling
Other deep learning approaches have been proposed based and up-sampling layers in the convolutional encoding and
on U-Net for 3D medical image segmentation tasks as well. The decoding units of the RU-Net model respectively. In the case of
3D-Unet architecture for volumetric segmentation learns from R2U-Net, the final outputs of the RCNN unit are passed through
sparsely annotated volumetric images [13]. A powerful end-to- the residual unit that is shown Fig. 4(d). Let’s consider that the
end 3D medical image segmentation system based on output of the RRCNN-block is 𝑥𝑙+1 and can be calculated as
volumetric images called V-net has been proposed, which follows:
consists of a FCN with residual connections [14]. This paper
also introduces a dice loss layer [14]. Furthermore, a 3D deeply 𝑥𝑙+1 = 𝑥𝑙 + ℱ(𝑥𝑙 , 𝑤𝑙 ) (3)
supervised approach for automated segmentation of volumetric Here, 𝑥𝑙 represents the input samples of the RRCNN-block.
medical images was presented in [36]. High-Res3DNet was The 𝑥𝑙+1 sample is used the input for the immediate succeeding
proposed using residual networks for 3D segmentation tasks in sub-sampling or up-sampling layers in the encoding and
2016 [37]. In 2017, a CNN based brain tumor segmentation decoding convolutional units of R2U-Net. However, the
approach was proposed using a 3D-CNN model with a fully number of feature maps and the dimensions of the feature maps
connected CRF [38]. Pancreas segmentation was proposed in for the residual units are the same as in the RRCNN-block
[39], and Voxresnet was proposed in 2016 where a deep voxel shown in Fig. 4 (d).
wise residual network is used for brain segmentation. This
architecture utilizes residual networks and summation of
feature maps from different layers [40].
Alternatively, we have proposed two models for semantic
segmentation based on the architecture of U-Net in this paper.
The proposed Recurrent Convolutional Neural Networks
(RCNN) model based on U-Net is named RU-Net, which is
shown in Fig. 3. Additionally, we have proposed a residual
RCNN based U-Net model which is called R2U-Net. The
following section provides the architectural details of both Fig. 4. Different variant of convolutional and recurrent convolutional units (a)
models. Forward convolutional units, (b) Recurrent convolutional block (c) Residual
convolutional unit, and (d) Recurrent Residual convolutional units (RRCU).
III. RU-NET AND R2U-NET ARCHITECTURES
The proposed deep learning models are the building blocks
Inspired by the deep residual model [7], RCNN [41], and U-
of the stacked convolutional units shown in Fig. 4(b) and (d).
Net [12], we propose two models for segmentation tasks which
There are four different architectures evaluated in this work. copying unit from the basic U-Net model and use only
First, U-Net with forward convolution layers and feature concatenation operations, resulting a much-sophisticated
concatenation is applied as an alternative to the crop and copy architecture that results in better performance.
method found in the primary version of U-Net [12]. The basic
convolutional unit of this model is shown in Fig. 4(a). Second,
U-Net with forward convolutional layers with residual
connectivity is used, which is often called residual U-net
(ResU-Net) and is shown in Fig. 4(c) [14]. The third
architecture is U-Net with forward recurrent convolutional
layers as shown in Fig. 4(b), which is named RU-Net. Finally,
the last architecture is U-Net with recurrent convolution layers
with residual connectivity as shown in Fig. 4(d), which is
named R2U-Net. The pictorial representation of the unfolded
RCL layers with respect to time-step is shown in Fig 5. Here
t=2 (0 ~ 2), refers to the recurrent convolutional operation that
includes one single convolution layer followed by two sub-
sequential recurrent convolutional layers. In this
implementation, we have applied concatenation to the feature
maps from the encoding unit to the decoding unit for both RU-
Net and R2U-Net models.

Fig. 6. Example images from training dataset: left column from DRIVE dataset,
middle column from STARE dataset and right column from CHASE-DB1
dataset. The first row shows the original images, second row shows fields of
view (FOV), and third row shows the target outputs.

There are several advantages of using the proposed

architectures when compared with U-Net. The first is the
efficiency in terms of the number of network parameters. The
proposed RU-Net, and R2U-Net architectures are designed to
have the same number of network parameters when compared
to U-Net and ResU-Net, and RU-Net and R2U-Net show better
Fig. 5. Unfolded recurrent convolutional units for t = 2 (left) and t = 3 (right). performance on segmentation tasks. The recurrent and residual
operations do not increase the number of network parameters.
The differences between the proposed models with respect to However, they do have a significant impact on training and
the U-Net model are three-fold. This architecture consists of testing performance. This is shown through empirical evidence
convolutional encoding and decoding units same as U-Net. with a set of experiments in the following sections [43]. This
However, the RCLs and RCLs with residual units are used approach is also generalizable, as it easily be applied deep
instead of regular forward convolutional layers in both the learning models based on SegNet [10], 3D-UNet [13], and V-
encoding and decoding units. The residual unit with RCLs helps Net [14] with improved performance for segmentation tasks.
to develop a more efficient deeper model. Second, the efficient
feature accumulation method is included in the RCL units of IV. EXPERIMENTAL SETUP AND RESULTS
both proposed models. The effectiveness of feature To demonstrate the performance of the RU-Net and R2U-Net
accumulation from one part of the network to the other is shown models, we have tested them on three different medical imaging
in the CNN-based segmentation approach for medical imaging. datasets. These include blood vessel segmentations from retina
In this model, the element-wise feature summation is performed images (DRIVE, STARE, and CHASE_DB1 shown in Fig. 6),
outside of the U-Net model [32]. This model only shows the skin cancer lesion segmentation, and lung segmentation from
benefit during the training process in the form of better 2D images. For this implementation, the Keras, and
convergence. However, our proposed models show benefits for TensorFlow frameworks are used on a single GPU machine
both training and testing phases due to the feature accumulation with 56G of RAM and an NIVIDIA GEFORCE GTX-980 Ti.
inside the model. The feature accumulation with respect to
A. Database Summary
different time-steps ensures better and stronger feature
representation. Thus, it helps extract very low-level features 1) Blood Vessel Segmentation
which are essential for segmentation tasks for different We have experimented on three different popular datasets for
modalities of medical imaging (such as blood vessel retina blood vessel segmentation including DRIVE, STARE,
segmentation). Third, we have removed the cropping and and CHASH_DB1. The DRIVE dataset is consisted of 40 color
retinal images in total, in which 20 samples are used for training randomly. A 20-sample set is used for training and the
and remaining 20 samples are used for testing. The size of each remaining 8 samples are used for testing.
original image is 565×584 pixels [44]. To develop a square As the dimensionality of the input data larger than the entire
dataset, the images are cropped to only contain the data from DRIVE dataset, we have considered 250,000 patches in total
columns 9 through 574, which then makes each image 565×565 from 20 images for both STARE and CHASE_DB1. In this case
pixels. In this implementation, we considered 190,000 225,000 patches are used for training and the remaining 25,000
randomly selected patches from 20 of the images in the DRIVE patches are used for validation. Since the binary FOV (which
dataset, where 171,000 patches are used for training, and the is shown in second row in Fig. 6) is not available for the STARE
remaining 19,000 patches used for validation. The size of each and CHASE_DB1 datasets, we generated FOV masks using a
patch is 48×48 for all three datasets shown in Fig. 7. The second similar technique to the one described in [47]. One advantage
dataset, STARE, contains 20 color images, and each image has of the patch-based approach is that the patches give the network
a size of 700×605 pixels [45, 46]. Due to the smaller number of access to local information about the pixels, which has impact
samples, two approaches are applied very often for training and on overall prediction. Furthermore, it ensures that the classes of
testing on this dataset. First, training sometimes performed with the input data are balanced. The input patches are randomly
randomly selected samples from all 20 images [53]. sampled over an entire image, which also includes the outside
region of the FOV.
2) Skin Cancer Segmentation
This dataset is taken from the Kaggle competition on skin
lesion segmentation that occurred in 2017 [49]. This dataset
contains 2000 samples in total. It consists of 1250 training
samples, 150 validation samples, and 600 testing samples. The
original size of each sample was 700×900, which was rescaled
to 256×256 for this implementation. The training samples
include the original images, as well as corresponding target
binary images containing cancer or non-cancer lesions. The
target pixels are represented with a value of either 255 or 0 for
Fig. 7. Example patches in the left and corresponding outputs of patches are
shown in the right.
the pixels outside of the target lesion.
3) Lung Segmentation
The Lung Nodule Analysis (LUNA) competition at the
Kaggle Data Science Bowl in 2017 was held to find lung lesions
in 2D and 3D CT images. The provided dataset consisted of 534
2D samples with respective label images for lung segmentation
[50]. For this study, 70% of the images are used for training and
the remaining 30% are used for testing. The original image size
was 512×512, however, we resized the images to 256×256
pixels in this implementation.
B. Quantitative Analysis Approaches
For quantitative analysis of the experimental results, several
performance metrics are considered, including accuracy (AC),
sensitivity (SE), specificity (SP), F1-score, Dice coefficient
(DC), and Jaccard similarity (JS). To do this we also use the
variables True Positive (TP), True Negative (TN), False
Fig. 8. Experimental outputs for DRIVE dataset using R2UNet: first row shows Positive (FP), and False Negative (FN). The overall accuracy is
input image in gray scale, second row show ground truth, and third row shows calculated using Eq. (4), and sensitivity is calculated using Eq.
the experimental outputs. (5).
𝑇𝑃+𝑇𝑁
Another approach is the “leave-one-out” method, in which 𝐴𝐶 = (4)
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁
each image is tested, and training is conducted on the remaining 𝑇𝑃
19 samples [47]. Therefore, there is no overlap between training 𝑆𝐸 = (5)
𝑇𝑃+𝐹𝑁
and testing samples. In this implementation, we used the “leave- Furthermore, specificity is calculated using the following Eq.
one-out” approach for STARE dataset. The CHASH_DB1 (6).
dataset contains 28 color retina images and the size of each
𝑇𝑁
image is 999×960 pixels [48]. The images in this dataset were 𝑆𝑃 = (6)
𝑇𝑁+𝐹𝑃
collected from both left and right eyes of 14 school children.
The dataset is divided into two sets where samples are selected The DC is expressed as in Eq. (7) according to [51]. Here GT
refers to the ground truth and SR refers the segmentation result.
 |𝐺𝑇∩𝑆𝑅| validation accuracy for the STARE dataset is shown in Figs. 12
𝐷𝐶 = 2 |𝐺𝑇|+|𝑆𝑅|
(7)
and 13 respectively.
The JS is represented using Eq. (8) as in [52]. R2U-Net shows a better performance than all other models
|𝐺𝑇∩𝑆𝑅|
during training. In addition, the validation accuracy in Fig. 13
𝐽𝑆 = |𝐺𝑇∪𝑆𝑅|
(8) demonstrates that the RU-Net and R2U-Net models provide
better validation accuracy when compared to the equivalent U-
However, the area under curve (AUC) and the receiver Net and ResU-Net models. Thus, the performance demonstrates
operating characteristics (ROC) curve are common evaluation the effectiveness of the proposed approaches for segmentation
measures for medical image segmentation tasks. In this tasks.
experiment, we utilized both analytical methods to evaluate the
performance of the proposed approaches considering the
mentioned criterions against existing state-of-the-art
techniques.

Fig. 11. Experimental outputs of STARE dataset using R2UNet: first row shows
input image after performing normalization, second row show ground truth, and
third row shows the experimental outputs.

Fig. 9. Training accuracy of the proposed models of RU-Net, and R2U-Net

against ResU-Net and U-Net.

C. Results
1) Retina Blood Vessel Segmentation Using the DRIVE
Dataset
The precise segmentation results achieved with the proposed
R2U-Net model are shown in Fig. 8. Figs. 9 and 10 show the
training and validation accuracy when using the DRIVE
dataset. These figures show that the proposed R2U-Net and
RU-Net models provide better performance during both the
training and validation phase when compared to U-Net and
Fig. 12. Training accuracy in STARE dataset for R2U-Net, RU-Net, ResU-Net,
ResU-Net. and U-Net.

Fig. 10. Validation accuracy of the proposed models against ResU-Net and U-
Net.
Fig. 13. Validation accuracy in STARE dataset for R2U-Net, RU-Net, ResU-
2) Retina blood vessel segmentation on the STARE dataset Net, and U-Net.
The experimental outputs of R2U-Net when using the
STARE dataset are shown in Fig. 11. The training and 3) CHASE_DB1
For qualitative analysis, the example outputs of R2U-Net are
shown in Fig. 14. For quantitative analysis, the results are given
in Table I. From the table, it can be concluded that in all cases, validation during training with a batch size of 32 and 150
the proposed RU-Net and R2U-Net models show better epochs.
performance in terms of AUC and accuracy. The ROC for the The training accuracy of the proposed models R2U-Net and
highest AUCs for the R2U-Net model on each of the three retina RU-Net was compared with that of ResU-Net and U-Net for an
blood vessel segmentation datasets is shown in Fig. 15. end-to-end image based segmentation approach. The result is

TABLE I. EXPERIMENTAL RESULTS OF PROPOSED APPROACHES FOR RETINA BLOOD VESSEL SEGMENTATION AND COMPARISON AGAINST OTHER
TRADITIONAL AND DEEP LEARNING-BASED APPROACHES.
Dataset Methods Year F1-score SE SP AC AUC
DRIVE Chen [53] 2014 - o.7252 0.9798 0.9474 0.9648
Azzopardi [54] 2015 - 0.7655 0.9704 0.9442 0.9614
Roychowdhury[55] 2016 - 0.7250 0.9830 0.9520 0.9620
Liskowsk [56] 2016 - 0.7763 0.9768 0.9495 0.9720
Qiaoliang Li [57] 2016 - 0.7569 0.9816 0.9527 0.9738
U-Net 2018 0.8142 0.7537 0.9820 0.9531 0.9755
Residual U-Net 2018 0.8149 0.7726 0.9820 0.9553 0.9779
Recurrent U-Net 2018 0.8155 0.7751 0.9816 0.9556 0.9782
R2U-Net 2018 0.8171 0.7792 0.9813 0.9556 0.9784
STARE Marin et al. [58] 2011 - 0.6940 0.9770 0.9520 0.9820
Fraz [59] 2012 - 0.7548 0.9763 0.9534 0.9768
Roychowdhury[55] 2016 - 0.7720 0.9730 0.9510 0.9690
Liskowsk [56] 2016 - 0.7867 0.9754 0.9566 0.9785
Qiaoliang Li [57] 2016 - 0.7726 0.9844 0.9628 0.9879
U-Net 2018 0.8373 0.8270 0.9842 0.9690 0.9898
Residual U-Net 2018 0.8388 0.8203 0.9856 0.9700 0.9904
Recurrent U-Net 2018 0.8396 0.8108 0.9871 0.9706 0.9909
R2U-Net 2018 0.8475 0.8298 0.9862 0.9712 0.9914
CHASE_DB1 Fraz [59] 2012 - 0.7224 0.9711 0.9469 0.9712
Fraz [60] 2014 - - - 0.9524 0.9760
Azzopardi [54] 2015 - 0.7655 0.9704 0.9442 0.9614
Roychowdhury[55] 2016 - 0.7201 0.9824 0.9530 0.9532
Qiaoliang Li [57] 2016 - 0.7507 0.9793 0.9581 0.9793
U-Net 2018 0.7783 0.8288 0.9701 0.9578 0.9772
Residual U-Net 2018 0.7800 0.7726 0.9820 0.9553 0.9779
Recurrent U-Net 2018 0.7810 0.7459 0.9836 0.9622 0.9803
R2U-Net 2018 0.7928 0.7756 0.9820 0.9634 0.9815

shown in Fig. 16. The validation accuracy is shown in Fig. 17.

In both cases, the proposed models show better performance
when compared with the equivalent U-Net and ResU-Net
models. This clearly demonstrates the robustness of the
proposed models in end-to-end image-based segmentation
tasks.

Fig. 14. Qualitative analysis for CHASE_DB1 dataset. The segmentation

outputs of 8 testing samples using R2U-Net. First row shows the input images,
second row is ground truth, and third row shows the segmentation outputs using
R2U-Net.
Fig. 15. AUC for retina blood vessel segmentation for the best performance
achieved with R2U-Net.
4) Skin Cancer Lesion Segmentation
In this implementation, this dataset is preprocessed with
mean subtraction and normalized according to the standard
deviation. We used the ADAM optimization technique with a
learning rate of 2×10-4 and binary cross entropy loss. In
addition, we also calculated MSE error during the training and
validation phase. In this case 10% of the samples are used for
 clearly shows the robustness of the proposed segmentation
method.
We have compared the performance of the proposed
approaches against recently published results with respect to
sensitivity, specificity, accuracy, AUC, and DC. The proposed
R2U-Net model provides a testing accuracy 0.9424 with a
higher AUC, which is 0.9419. The average AUC for skin lesion
segmentation is shown in Fig. 19. In addition, we calculated the
average DC in the testing phase and achieved 0.8616, which is
around 1.26% better than recently proposed alternatives [62].
Furthermore, the JSC and F1 scores are calculated and the R2U-
Net model obtains 0.9421 for JSC and 0.8920 for F1 score for
Fig. 16. Training accuracy for skin lesion segmentation. skin lesion segmentation with t=3. These results are achieved

TABLE II. EXPERIMENTAL RESULTS OF PROPOSED APPROACHES FOR SKIN CANCER LESION SEGMENTATION AND COMPARISON AGAINST OTHER
EXISTING APPROACHES. JACCARD SIMILARITY SCORE (JSC).
Methods Year SE SP JSC F1-score AC AUC DC
Conv. classifier VGG-16 [61] 2017 0.533 - - - 0.6130 0.6420 -
Conv. classifier Inception-v3[61] 2017 0.760 - - - 0.6930 0.7390 -
Melanoma detection [62] 2017 - - - - o.9340 - 0.8490
Skin Lesion Analysis [63] 2017 0.8250 0.9750 - - 0.9340 - -
U-Net (t=2) 2018 0.9479 0.9263 0.9314 0.8682 0.9314 0.9371 0.8476
ResU-Net (t=2) 2018 0.9454 0.9338 0.9367 0.8799 0.9367 0.9396 0.8567
RecU-Net (t=2) 2018 0.9334 0.9395 0.9380 0.8841 0.9380 0.9364 0.8592
R2U-Net (t=2) 2018 0.9496 0.9313 0.9372 0.8823 0.9372 0.9405 0.8608
R2U-Net (t=3) 2018 0.9414 0.9425 0.9421 0.8920 0.9424 0.9419 0.8616

with a R2U-Net model that only contains about 1.037 million

The quantitative results of this experiment were compared (M) network parameters. Contrarily, the work presented in [61]
against existing methods as shown in Table II. Some of the evaluated VGG-16 and Incpetion-V3 models for skin lesion
example outputs from the testing phase are shown in Fig. 18. segmentation, but those networks contained around 138M and
The first column shows the input images, the second column 23M network parameters respectively.
shows the ground truth, the network outputs are shown in the
third column, and the fourth column demonstrates the final
outputs after performing post processing with a threshold of 0.5.
Figure 18 shows promising segmentation results.

Fig. 17. Validation accuracy for skin lesion segmentation.

In most cases, the target lesions are segmented accurately

with almost the same shape of ground truth. However, if we
observe the second and third rows in Fig. 18, it can be clearly
seen that the input images contain two spots, one is a target
lesion and the other bright spot which is not a target. This result
is obtained even though the non-target lesion is brighter than
the target lesion shown in the third row in Fig. 18. The R2U- Fig. 18. This results demonstrates qualitative assessment of the proposed R2U-
Net for skin cancer segmentation task with t=3. First column is the input
Net model still segments the desired part accurately, which sample, second column is ground truth, third column shows the outputs from
network, and fourth column show the final resulting after performing respectively. However, we also experimented with U-Net,
thresholding with 0.5.
ResU-Net, RU-Net, and R2U-Net models with following
5) Lung Segmentation structure: 116326412864  32161. In this
Lung segmentation is very important for analyzing lung case we used a time-step of t=3, which refers to one forward
related diseases, and can be applied to lung cancer segmentation convolution layer followed by three subsequent recurrent
and lung pattern classification for identifying other problems. convolutional layers. This network was tested on skin and lung
In this experiment, the ADAM optimizer is used with a learning lesion segmentation. Though the number of network parameters
rate of 2×10-4. We used binary cross entropy loss, and also increase little bit with respect to the time-step in the recurrent
calculated MSE during training and validation. In this case 10% convolution layer, further improved performance can be clearly
of the samples were used for validation with a batch size of 16 seen in the last rows of Table II and III. Furthermore, we have
and 150 epochs 150. Table III shows the summary of how well evaluated both of the proposed models for patch-based
the proposed models performed against equivalent U-Net and modeling on retina blood vessel segmentation and end-to-end
ResU-Net models. The experimental results show that the image-based methods for skin and lung lesion segmentation.
proposed models outperform the U-Net and ResU-Net models In both cases, the proposed models outperform existing state-

TABLE III. EXPERIMENTAL OUTPUTS OF PROPOSED MODELS OF RU-NET AND R2U-NET FOR LUNG SEGMENTATION AND COMPARISON
AGAINST RESU-NET AND U-NET MODELS.
Methods Year SE SP JSC F1-Score AC AUC
U-Net (t=2) 2018 0.9696 0.9872 0.9858 0.9658 0.9828 0.9784
ResU-Net(t=2) 2018 0.9555 0.9945 0.9850 0.9690 0.9849 0.9750
RU-Net (t=2) 2018 0.9734 0.9866 0.9836 0.9638 0.9836 0.9800
R2U-Net (t=2) 2018 0.9826 0.9918 0.9897 0.9780 0.9897 0.9872
R2U-Net (t=3) 2018 0.9832 0.9944 0.9918 0.9823 0.9918 0.9889

with same number of network parameters. of-the-art methods including ResU-Net and U-Net in terms of
AUC and accuracy on all three datasets. The network
architectures with different numbers of network parameters
with respect to the different time-step are shown in Table IV.
The processing times during the testing phase for the STARE,
CHASE_DB, and DRIVE datasets were 6.42, 8.66, and 2.84
seconds per sample respectively. In addition, skin cancer
segmentation and lung segmentation take 0.22 and 1.145
seconds per sample respectively.

TABLE IV. ARCHITECTURE AND NUMBER OF NETWORK PARAMETERS.

t Network architectures Number of parameters
(million)
2 1-> 16->32->64>128->64 –> 32- 0.845
>16->1
Fig. 19. ROC-AUC for skin segmentation four models with t=2 and t=3. 3 1-> 16->32->64>128->64 –> 32- 1.037
>16->1
Furthermore, many models struggle to define the class
boundary properly during segmentation tasks [64]. However, if
we observe the experimental outputs shown in Fig. 20, the
outputs in the third column show different hit maps on the
border, which can be used to define the boundary of the lung
region, while the ground truth tends to have a smooth boundary.
In addition, if we observe the input, ground truth, and output
of this proposed approaches in the second row, it can be
observed that the output of the proposed approaches shows
better segmentation with appropriate contour. The ROC with
AUCs are shown Fig. 21. The highest AUC is achieved with the
proposed approach of R2U-Net with t=3.
D. Evaluation
Most of the cases, the networks are evaluated for different
segmentation tasks with following architectures:
164128256512256  128641 that require
4.2M network parameters and 164128256512256
 128641, which require about 8.5M network parameters
Fig. 20. Qualitative assessment of R2U-Net performance on Lung segmentation
dataset: first column input images, second column ground truth, and third
column outputs with R2U-Net.
Fig. 21. ROC curve for lung segmentation four models with t=2 and t=3.

E. Computational time REFERENCES

[1] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "ImageNet
The computational time for testing per sample is shown in classification with deep convolutional neural networks." Advances in
Table V for blood vessel segmentation for retina images, skin neural information processing systems. 2012.
cancer, and lung segmentation respectively. [2] Long, J., Shelhamer, E., & Darrell, T. (2015). Fully convolutional
networks for semantic segmentation. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition (pp. 3431-
TABLE V. COMPUTATIONAL TIME FOR TESTING PHASE. 3440).
Dataset Time (Sec.)/ sample [3] Wang, Naiyan, et al. "Transferring rich feature hierarchies for robust
Blood vessel DRIVE 6.42 visual tracking." arXiv preprint arXiv: 1501.04587 (2015).
segmentation STARE 8.66 [4] Mao, Junhua, et al. "Deep captioning with multimodal recurrent neural
CHASE_DB1 2.84 networks (m-rnn)." arXiv preprint arXiv: 1412.6632 (2014)
Skin cancer segmentation 0.22 [5] Simonyan, Karen, and Andrew Zisserman. "Very deep convolutional
Lung segmentation 1.15 networks for large-scale image recognition." arXiv preprint arXiv:
1409.1556 (2014).
[6] Szegedy, Christian, et al. "Going deeper with convolutions." Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition.
V. CONCLUSION AND FUTURE WORKS
2015.
In this paper, we proposed an extension of the U-Net [7] He, Kaiming, et al. "Deep residual learning for image recognition."
Proceedings of the IEEE Conference on Computer Vision and Pattern
architecture using Recurrent Convolutional Neural Networks
Recognition. 2016.
and Recurrent Residual Convolutional Neural Networks. The [8] Huang, Gao, et al. "Densely connected convolutional networks." arXiv
proposed models are called “RU-Net” and “R2U-Net” preprint arXiv:1608.06993 (2016).
respectively. These models were evaluated using three different [9] Sabour, Sara, Nicholas Frosst, and Geoffrey E. Hinton. "Dynamic routing
between capsules." Advances in Neural Information Processing Systems.
applications in the field of medical imaging including retina 2017.
blood vessel segmentation, skin cancer lesion segmentation, [10] Badrinarayanan, Vijay, Alex Kendall, and Roberto Cipolla. "Segnet: A
and lung segmentation. The experimental results demonstrate deep convolutional encoder-decoder architecture for image
segmentation." arXiv preprint arXiv:1511.00561(2015).
that the proposed RU-Net, and R2U-Net models show better [11] Ciresan, Dan, et al. "Deep neural networks segment neuronal membranes
performance in segmentation tasks with the same number of in electron microscopy images." Advances in neural information
network parameters when compared to existing methods processing systems. 2012.
[12] Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net:
including the U-Net and residual U-Net (or ResU-Net) models Convolutional networks for biomedical image
on all three datasets. In addition, results show that these segmentation." International Conference on Medical image computing
proposed models not only ensure better performance during the and computer-assisted intervention. Springer, Cham, 2015.
[13] Çiçek, Özgün, et al. "3D U-Net: learning dense volumetric segmentation
training but also in testing phase. In future, we would like to from sparse annotation." International Conference on Medical Image
explore the same architecture with a novel feature fusion Computing and Computer-Assisted Intervention. Springer International
strategy from encoding to the decoding units. Publishing, 2016.
[14] Milletari, Fausto, Nassir Navab, and Seyed-Ahmad Ahmadi. "V-net:
Fully convolutional neural networks for volumetric medical image
segmentation." 3D Vision (3DV), 2016 Fourth International Conference
on. IEEE, 2016.
[15] Yang, Dong, et al. "Automated anatomical landmark detection ondistal
femur surface using convolutional neural network." Biomedical Imaging
(ISBI), 2015 IEEE 12th International Symposium on. IEEE, 2015.
[16] Cai, Yunliang, et al. "Multi-modal vertebrae recognition using
transformed deep convolution network." Computerized Medical Imaging
and Graphics 51 (2016): 11-19.
[17] Ramesh, N., J-H. Yoo, and I. K. Sethi. "Thresholding based on histogram
approximation." IEE Proceedings-Vision, Image and Signal
Processing 142.5 (1995): 271-279.
[18] Sharma, Neeraj, and Amit Kumar Ray. "Computer aided segmentation of
medical images based on hybridized approach of edge and region based
techniques." Proceedings of International Conference on Mathematical
 Biology', Mathematical Biology Recent Trends by Anamaya Publishers. [41] Liang, Ming, and Xiaolin Hu. "Recurrent convolutional neural network
2006. for object recognition." Proceedings of the IEEE Conference on
[19] Boykov, Yuri Y., and M-P. Jolly. "Interactive graph cuts for optimal Computer Vision and Pattern Recognition. 2015.
boundary & region segmentation of objects in ND images." Computer [42] Alom, Md Zahangir, et al. "Inception Recurrent Convolutional Neural
Vision, 2001. ICCV 2001. Proceedings. Eighth IEEE International Network for Object Recognition." arXiv preprint
Conference on. Vol. 1. IEEE, 2001. arXiv:1704.07709 (2017).
[20] Litjens, Geert, et al. "A survey on deep learning in medical image [43] Alom, Md Zahangir, et al. "Improved Inception-Residual Convolutional
analysis." arXiv preprint arXiv:1702.05747 (2017). Neural Network for Object Recognition." arXiv preprint
[21] Greenspan, Hayit, Bram van Ginneken, and Ronald M. Summers. "Guest arXiv:1712.09888 (2017).
editorial deep learning in medical imaging: Overview and future promise [44] Staal, Joes, et al. "Ridge-based vessel segmentation in color images of the
of an exciting new technique." IEEE Transactions on Medical retina." IEEE transactions on medical imaging23.4 (2004): 501-509.
Imaging 35.5 (2016): 1153-1159. [45] Hoover, A. D., Valentina Kouznetsova, and Michael Goldbaum.
[22] Havaei, Mohammad, et al. "Brain tumor segmentation with deep neural "Locating blood vessels in retinal images by piecewise threshold probing
networks." Medical image analysis 35 (2017): 18-31. of a matched filter response." IEEE Transactions on Medical
[23] G. Brostow, J. Fauqueur, and R. Cipolla, “Semantic object classes in imaging 19.3 (2000): 203-210.
video: A high-definition ground truth database,” PRL, vol. 30(2), pp. 88– [46] Zhao, Yitian, et al. "Automated vessel segmentation using infinite
97, 2009. [23] S. Song, S. P. perimeter active contour model with hybrid region information with
[24] Lichtenberg, and J. Xiao, “Sun rgb-d: A rgb-d scene understanding application to retinal images." IEEE transactions on medical
benchmark suite,” in Proceedings of the IEEE Conference on Computer imaging 34.9 (2015): 1797-1807.
Vision and Pattern Recognition, pp. 567–576, 2015. [47] Soares, João VB, et al. "Retinal vessel segmentation using the 2-D Gabor
[25] Kistler, Michael, et al. "The virtual skeleton database: an open access wavelet and supervised classification." IEEE Transactions on medical
repository for biomedical research and collaboration." Journal of medical Imaging 25.9 (2006): 1214-1222.
Internet research 15.11 (2013). [48] Fraz, Muhammad Moazam, et al. "Blood vessel segmentation
[26] He, Kaiming, et al. "Identity mappings in deep residual methodologies in retinal images–a survey." Computer methods and
networks." European Conference on Computer Vision. Springer programs in biomedicine 108.1 (2012): 407-433.
International Publishing, 2016. [49] https://challenge2017.isic-archive.com
[27] S. Zheng, S. Jayasumana, B. Romera-Paredes, V. Vineet, Z. Su, D. Du, [50] https://www.kaggle.com/kmader/finding-lungs-in-ct-data/data.
C. Huang, and P. H. Torr, “Conditional random fields as recurrent neural [51] Dice, Lee R. "Measures of the amount of ecologic association between
networks,” in Proceedings of the IEEE International Conference on species." Ecology 26.3 (1945): 297-302.
Computer Vision, pp. 1529–1537, 2015. [52] Jaccard, Paul. "The distribution of the flora in the alpine zone." New
[28] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, phytologist 11.2 (1912): 37-50.
“Semantic image segmentation with deep convolutional nets and fully [53] Cheng, Erkang, et al. "Discriminative vessel segmentation in retinal
connected CRFs,” in ICLR, 2015. images by fusing context-aware hybrid features." Machine vision and
[29] Kendall, Alex, Vijay Badrinarayanan, and Roberto Cipolla. "Bayesian applications 25.7 (2014): 1779-1792.
segnet: Model uncertainty in deep convolutional encoder-decoder [54] Azzopardi, George, et al. "Trainable COSFIRE filters for vessel
architectures for scene understanding." arXiv preprint delineation with application to retinal images." Medical image
arXiv:1511.02680 (2015). analysis 19.1 (2015): 46-57.
[30] Zhang, Zhengxin, Qingjie Liu, and Yunhong Wang. "Road Extraction by [55] Roychowdhury, Sohini, Dara D. Koozekanani, and Keshab K. Parhi.
Deep Residual U-Net." arXiv preprint arXiv:1711.10684 (2017). Blood vessel segmentation of fundus images by major vessel extraction
[31] Li, Ruirui, et al. "DeepUNet: A Deep Fully Convolutional Network for and subimage classification." IEEE journal of biomedical and health
Pixel-level Sea-Land Segmentation." arXiv preprint informatics 19.3 (2015): 1118-1128.
arXiv:1709.00201 (2017). [56] Liskowski, Paweł, and Krzysztof Krawiec. "Segmenting Retinal Blood
[32] Kayalibay, Baris, Grady Jensen, and Patrick van der Smagt. "CNN-based Vessels With Deep Neural Networks." IEEE transactions on medical
Segmentation of Medical Imaging Data." arXiv preprint imaging 35.11 (2016): 2369-2380.
arXiv:1701.03056 (2017). [57] Li, Qiaoliang, et al. "A cross-modality learning approach for vessel
[33] Drozdzal, Michal, et al. "The importance of skip connections in segmentation in retinal images." IEEE transactions on medical
biomedical image segmentation." International Workshop on Large- imaging 35.1 (2016): 109-118.
Scale Annotation of Biomedical Data and Expert Label Synthesis. [58] Marín, Diego, et al. "A new supervised method for blood vessel
Springer International Publishing, 2016. segmentation in retinal images by using gray- level and moment
[34] Chen, Hao, et al. "Dcan: Deep contour-aware networks for accurate gland invariants-based features." IEEE Transactions on medical imaging 30.1
segmentation." Proceedings of the IEEE conference on Computer Vision (2011): 146-158.
and Pattern Recognition. 2016. [59] Fraz, Muhammad Moazam, et al. "An ensemble classification-based
[35] McKinley, Richard, et al. "Nabla-net: A Deep Dag-Like Convolutional approach applied to retinal blood vessel segmentation." IEEE
Architecture for Biomedical Image Segmentation." International Transactions on Biomedical Engineering 59.9 (2012): 2538-2548.
Workshop on Brainlesion: Glioma, Multiple Sclerosis, Stroke and [60] Fraz, Muhammad Moazam, et al. "Delineation of blood vessels in
Traumatic Brain Injuries. Springer, Cham, 2016. pediatric retinal images using decision trees-based ensemble
[36] Q. Dou, L. Yu, H. Chen, Y. Jin, X. Yang, J. Qin, and P.-A. Heng, “3D classification." International journal of computer assisted radiology and
deeply supervised network for automated segmentation of volumetric surgery 9.5 (2014): 795-811.
medical images,” Medical Image Analysis, vol. 41, pp. 40–54, 2017. [61] Burdick, Jack, et al. "Rethinking Skin Lesion Segmentation in a
[37] Li, Wenqi, et al. "On the Compactness, Efficiency, and Representation of Convolutional Classifier." Journal of digital imaging (2017): 1-6.
3D Convolutional Networks: Brain Parcellation as a Pretext [62] Codella, Noel CF, et al. "Skin lesion analysis toward melanoma detection:
Task." International Conference on Information Processing in Medical A challenge at the 2017 international symposium on biomedical imaging
Imaging. Springer, Cham, 2017. (isbi), hosted by the international skin imaging collaboration (isic)." arXiv
[38] Kamnitsas, Konstantinos, et al. "Efficient multi-scale 3D CNN with fully preprint arXiv:1710.05006 (2017).
connected CRF for accurate brain lesion segmentation." Medical image [63] Li, Yuexiang, and Linlin Shen. "Skin Lesion Analysis Towards
analysis 36 (2017): 61-78. Melanoma Detection Using Deep Learning Network." arXiv preprint
[39] Roth, Holger R., et al. "Deeporgan: Multi-level deep convolutional arXiv:1703.00577 (2017).
networks for automated pancreas segmentation." International [64] Hsu, Roy Chaoming, et al. "Contour extraction in medical images using
Conference on Medical Image Computing and Computer-Assisted initial boundary pixel selection and segmental contour
Intervention. Springer, Cham, 2015. following." Multidimensional Systems and Signal Processing 23.4
[40] Chen, Hao, et al. "Voxresnet: Deep voxelwise residual networks for (2012): 469-498.
volumetric brain segmentation." arXiv preprint [65] Alom, Md Zahangir, et al. "The History Began from AlexNet: A
arXiv:1608.05895 (2016). Comprehensive Survey on Deep Learning Approaches." arXiv preprint
arXiv:1803.01164 (2018).