A Tour of Unsupervised Deep Learning for

Medical Image Analysis

Khalid Raza* and Nripendra Kumar Singh
Department of Computer Science, Jamia Millia Islamia, New Delhi
[email protected]

December 13, 2018

Abstract

Interpretation of medical images for diagnosis and treatment of complex disease from high-
dimensional and heterogeneous data remains a key challenge in transforming healthcare. In the last
few years, both supervised and unsupervised deep learning achieved promising results in the area of
medical imaging and image analysis. Unlike supervised learning which is biased towards how it is
being supervised and manual efforts to create class label for the algorithm, unsupervised learning
derive insights directly from the data itself, group the data and help to make data driven decisions
without any external bias. This review systematically presents various unsupervised models applied
to medical image analysis, including autoencoders and its several variants, Restricted Boltzmann
machines, Deep belief networks, Deep Boltzmann machine and Generative adversarial network.
Future research opportunities and challenges of unsupervised techniques for medical image analysis
have also been discussed.

Keywords: Unsupervised learning; medical image analysis; autoencoders; restricted Boltzmann

machine; Deep belief network

1. Introduction

Medical imaging techniques, including magnetic resonance imaging (MRI), positron

emission tomography (PET), computed tomography (CT), mammography, ultrasound, X-ray
and digital pathology images, are frequently used diagnostic system for the early detection,
diagnosis, and treatment of various complex diseases (Wani & Raza, 2018). In the clinics, the
images are mostly interpreted by human experts such as radiologists and physicians. Because
of major variations in pathology and the potential fatigue of human experts, scientists and
doctors have started using computer-assisted interventions. The advancement in machine
learning techniques, including deep learning, and availability of computing infrastructure
through cloud computing, have given fuel to the field of computer-assisted medical image
analysis and computer-assisted diagnosis (CAD). Deep learning is about learning
representations, i.e, learning intermediate concept or features which are important to capture
dependencies from input variables to output variables in supervised learning, or between
subsets of variables in unsupervised learning. Both supervised and unsupervised machine
learning approaches are widely applied in medical image analysis; each of them has their own
pros and cons. Some of widely used supervised (deep) learning algorithms are Feedforward
Neural Network (FFNN), Recurrent Neural Network (RNN), Convolutional Neural Network
(CNN), Support Vector Machine (SVM) and so on (Jabeen et al., 2018). There are many

1
scenarios where human supervisions are unavailable, inadequate or biased and therefore,
supervised learning algorithm cannot be directly used. Unsupervised learning algorithms,
including its deep architecture, give a big hope with lots of advantages and have been widely
applied in several areas of medical and engineering problems including medical image
analysis.

This chapter presents unsupervised deep learning models, its applications to medical image
analysis, list of software tools/packages and benchmark datasets; and discusses opportunities
and future challenges in the area.

2. Why Unsupervised Learning?

In the majority of machine learning projects, the workflow is designed in a supervised way,
where the algorithm is guided by us what to do and what not to! In such supervised
architecture the potential of the algorithms are limited in three ways, (i) A huge manual effort
to create labels and (ii) Biases related to how it is being supervised, which probabilities the
algorithm to think for other corner cases during problem solving, and (iii) Reduce the
scalability of target function at hand.

To intelligently solve these issues, unsupervised machine learning algorithm can be used.
Unsupervised machine learning algorithms not only derives insights directly from the data
and group the data, but also uses these insights for data-driven decisions making. Also,
unsupervised models are more robust in the sense that they act as a base for several different
complex tasks where these can be utilized as the holy grail of learning and classification. In
fact, the classification is not the only task that we do; rather, other tasks such as compression,
dimensionality reduction, denoising, super resolution and some degree of decision making
are also performed. Therefore, it is rather more useful to construct a model without knowing
what tasks will be at hand and we will use representation (or model) for. In a nutshell, we can
think of unsupervised learning as preparation (preprocessing) step for supervised learning
tasks, where unsupervised learning of representation may allow better generalization of a
classifier (Jabeen et al., 2018).

3. Taxonomy of Unsupervised Learning Tasks

In unsupervised learning, we group the unlabeled data set on the basis of underlying hidden
features. By grouping data through unsupervised learning, at least we learn something about
raw data.

3.1 Density estimation

Density estimation is one of the popular categories of unsupervised learning which discovers
the intrinsic feature and structure of large and complex unlabeled data set via another non-
probabilistic approach. Density estimation is a non-parametric method which doesn’t possess
much restriction and distributional assumption unlike parametric estimation.

2
Estimation of univariate or multivariate density function without any prior functional
assumptions get almost limitless function from data. There are some widely used non-
parametric methods of estimation.

3.1.1 Kernel density estimation

Kernel density estimation (KDE) uses statistical model to produce a probabilistic distribution
that resembles an observed variable as a random variable. Basically, KDE is used for data
smoothing, exploratory data analysis and visualization. A large number of kernels have been
proposed, namely normal Gaussian mixture model and multivariate Gaussian mixture model.
Some of the advantages of Kernel density estimation are:

o No need for model specification (data speaks itself).

o Estimation converges to any density, shape with sufficient sample.
o Easily generalizes to higher dimension.
o Densities are multivariate and multimodal with irregular cluster shape.

3.1.2 Histogram density estimation

Histogram based technique mainly adds smoothness of the density curve of reconstruction
which can be optimized by kernel parameters and closely related to KNN density estimation
algorithm (Bishop et al., 2006).

3.2 Dimensionality reduction

Why dimensionality reduction? There has been a tremendous increase in deployment of

sensors and various imaging technique’s which are being used in industry and medical
diagnosis continuously record data and store it to be analyzed later. Lots of redundancy or
noises are present initially when data are captured. For example, let us take a case of a patient
having bone fracture. Initially doctors suggested for X-ray images which is a 2D/3D imaging,
but when they do not find it helpful in diagnosis, then a CT scan and/or MRI (magnetic
resonance imaging) may be taken which gives more detailed results for further right
diagnosis. Now assume that an analyst sits with all this data to analyze and identified all
important variables/dimensions which have most significant information’s and left all
unwanted parts of data. This is the problem of high unwanted dimension removal and needs
treatment of dimension reduction. Dimension reduction is the process of reducing higher
dimension data set into a lesser dimension, ensuring that final reduced data must convey
equivalent information concisely.

Let’s look at figure shown below. It shows two-dimensional x and y which are measurement
of several objects in cm (x1) and inches (y1), if you continue to use both dimensions in
machine learning problems it will introduce lots of noise in the system. So, it is better to just
use one-dimension (z1) and they will convey similar information.

3
 Fig 1. Representation of data in two dimensional and one dimensional space

There are some common methods to perform dimensionality reduction:

S.1.2 Factor analysis
Some variables in given data are highly correlated. These variables can be grouped on the
basis of their correlations. This means a particular group can have highly correlated variable,
but have low correlation with variables of other groups. Each group represents single inherent
construct or factor. As compared to data having large number of dimensions, these factors are
small in number, while these factors are difficult to find. There are two methods for doing
factor analysis: (i) Exploratory Factor Analysis, (ii) Confirmatory Factor Analysis

3.2.2 Principal component analysis

A set of variables, which are linear combination of the original set of variables, performs
higher dimensional space mapped to lower dimensions in such a way that variance of data in
lower dimensional space is maximized. These new set of variables is known as principle
components.

Let’s consider a situation of two-dimensional data set, there can be only two principal
components, first principal component is the most possible variation of original data and
second principal component is orthogonal to the first principal component, as shown in Fig.
2.

4
 Fig. 2 Principle components on a two-dimensional data

In practice, a simple principal component analysis (PCA) can be used to construct the
covariance or correlation matrix of the data and compute the eigenvectors. The eigenvectors
correspond to the largest eigenvalues (principal component) are used to reconstruct a large
fraction of variance of original data. As a result, it is left with lesser number of eigenvector
and original space has been reduced. There might be chances of loss of data, but it is retained
by most important eigenvectors.

Consider a matrix U(m) which stored empirical mean of input matrix R,

( ) ∑ ( ) ( ) ( )

Calculate a normalized matrix X, X = (R – Ue), where e is a unitary vector matrix of size N.

Finally, the mean square error (E2) is calculated in which smallest eigenvalues are removed,

( ( )) ∑ ∑ (2)

The trace(A) is the sum of all eigenvalues. Simple PCA is not capable of constructing
nonlinear mapping, however, can implement nonlinear classification by using kernel
techniques.

3.2.3 Kernel PCA

Kernel PCA is a nonlinear extension of conventional PCA, which is designed for
dimensionality reduction of nonlinear subspaces depending on magnitude of input data and
problem setup. In medical image analysis, hybrid kernel PCA is frequently used to get better
results in unsupervised deep learning training model. Fischer et al. (2017) proposed an
unsupervised deep learning illumination invariant kernel PCA, which is applied to each patch
of respiratory signal extraction from X-ray fluoroscopy images leading to a set of low-
dimensional embedding.

5
A kernel PCA comprised a kernel matrix K and kernel function k(.) is a Mercer kernel
(Minh et al. 2006), defined as Ki j = k(x(i) , x (j)),such that k(.) return dot product of feature
space. Now mapping of an eigenvalue of the kernel matrix, the Eigen decomposition and
respected eigenvectors are computed as,

λ{i}e{i} = K e{i} (3)

*+ ( )
( ) *+
∑ ( ) (4)

where λ{i} is eigenvalues and e{i} is eigenvectors of K; T is the number of training sample x to
the principal component “i”. Fischer et al. (2017) analyzed different methods like PCA,
KPCA and Multi-Resolution PCA to Diaphragm tracking correlation coefficient between
different versions of the same sequence and agreed that Multi-Resolution PCA produce the
best result among most of the parameters. Principal component analysis network (PCANet) is
a simple network architecture and one of the benchmark frameworks (Chan et al. 2015) for
the unsupervised deep learning in recent time. However, Shi et al. (2017) propose an
encoding as C-RBH-PCANet which is improved PCANet to effectively integrate the color
pattern extraction and random binary hashing method for learning feature from color
histopathological images.

3.3 Clustering
Clustering is an unsupervised classification of unlabeled data (patterns, data item or feature
vectors) into similar groups (clusters) (Fig. 3). Cluster analysis is explanatory in nature to
find structure in data (Jain, 2008). Some model of clustering includes semi-supervised
clustering, ensemble clustering, simultaneous feature selection and large-scale data clustering
were emerging as a hybrid-clustering. It involves analysis of multivariate data and applied in
various scientific domains where clustering technique is utilized, such as machine learning,
image analysis, bioinformatics, pattern recognition, computer vision and so on.

Fig. 3 An illustration of data clustering

6
Clustering algorithm is broadly divided into two groups: hierarchal clustering and partitional
clustering, as described below.

3.3.1 Hierarchical clustering

Hierarchical clustering algorithms find clusters recursively (using previously established

cluster). These algorithms can be either in the agglomerative mode (bottom-up) in which
begin with each element as a separate cluster, merge the most similar pair of clusters
successively into large clusters, or in divisive (top-down) mode which begin with all elements
in one cluster, recursively divide into smaller clusters. A hierarchical clustering algorithm
yields a dendrogram representing group of patterns and similarity level (Jain et al., 1999). A
detailed discussion can be found in Jain et al. (1999).

3.3.2 Partitional (k-means) clustering

One of the most popular partitioning clustering algorithms is k-means. In spite of several
clustering algorithms published in over 50 years, k-means is still widely used (Jain, 2010).
The most frequently used functions in partitional clustering is squared error criterion, which
applied to isolate and compact clusters. Let X = {xi: i = 1, 2, 3, …. N} be the set of n d-
dimensional elements clustered into set of K clusters as C = {ck : k = 1, 2, 3,…K}. To find
partitions, squared error between empirical mean of a cluster and elements in the cluster is
minimized. Let µk be the mean of the cluster (ck), the squared error between mean and
elements in a cluster is defined as:

( ) ∑‖ ‖ ( )

The main objective of K-means is to minimize the sum of squared error for all k clusters
(Drineas et al., 1999).

( ) ∑ ∑‖ ‖ ( )

Minimizing objective function is an NP-hard problem even for k = 2. Thus, k-means is a

greedy algorithm and it can only be expected to converge to local minima.

4. Unsupervised deep learning models

This section introduces a formal introduction of unsupervised deep learning concepts,
models and architectures that are used in medical image analysis. The unsupervised deep
learning models can be roughly classified as shown in Fig. 4.

7
 Autoencoders
(AEs) &
Variants

Generative Restricted
adversarial Boltzmann
networks Machines
(GANs) Unsupervised (RBMs)
deep learning
models

Deep
Deep Belief
Bolzmann
Networks
Machines
(DBNs)
(DBMs)

Fig. 4 Unsupervised deep learning models

4.1 Auto-encoders and its variants

In the literature, autoencoders and its several variants are reported and are being extensively
applied in medical image analysis.

4.1.1 Autoencoders and Stacked autoencoder

Autoencoders (AEs) (Bourlard et al., 1988) are simple unsupervised learning model consist
single-layer neural network that transforms the input into a latent or compressed
representation by minimizing the reconstruction errors between input and output values of the
network. By constraining the dimension of latent representation (may be from different input)
it is possible to discover relevant pattern from the data. AEs framework defines a feature to
extract function with specific parameters (Bengio et al., 2013). Basically, AEs are trained
with specific function fƟ is called encoder and h = fƟ(x) is feature vector or representation
from input x, another parameterized function gƟ called decoder, producing input space back
from feature space. In short, basic AEs are trained to minimize reconstruction error in finding
a value of parameter , given by,

( ) ∑ . 〖( 〗 ( ))/ ( )

This minimization optionally followed by a non-linearity (most commonly used for encoder
and decoder) as given by,

8
 ( ) ( ) ( )

( ) ( ) ( )

where Sf and Sg are encoder and decoder activation function (normally, sigmoid, hyperbolic
tangent or an identity function), respectively; parameters of model = {W, b, W’, d}, where
W and W’ are encoder decoder weight matrices, and b and d are encoder and decoder bias
vector, respectively. Moreover, regularization or sparsity constraints may be applied in order
to boost the discovery process. In case, hidden layer has the same input as the input layer,
and no any non-linearity is added, the model would simply learn an identity function. Fig.
5(a) illustrates the basic structure of AE.

Stacked autoencoders (SAEs) are constructed by organizing AEs on top of each other also
known as deep AEs. SAEs consist of multiple AEs stacked into multiple layers where the
output of each layer is wired to the inputs of the successive layers Fig. 5(b). To obtain good
parameters, SAE uses greedy layer-wise training. The benefit of SAE is that it can enjoy the
benefits of deep network, which has greater expressive power. Furthermore, it usually
captures a useful hierarchical grouping of the input (Shin et al., 2013).

4.1.2 Denoising autoencoder

Denoising autoencoder (DAEs) is another variant of the auto-encoder. Denoising investigated

as a training criterion for learning to constitute better higher-level representation and extract
useful features (Vincent et al. 2010). DAEs prevent the model from learning a trivial solution
(Litjens G. et al., 2017) where the model is trained to reconstruct a clean input from the
corrupted version from noise or another corruption. his is done by corrupting the initial
input x into x by using a stochastic function x ~ qD x x . The corrupted input is then
mapped to a hidden representation y = fƟ(x) = s(Wx + b) and reconstruct z = gƟ’ (y). A
schematic representation of DAE is shown in Fig.5(c). Parameter and are initialized
randomly and trained using stochastic gradient descent in order to minimize average
reconstruction error. The denoising autoencoders continue minimizing same reconstruction
loss between clean X and reconstruction from Y. This continues maximizing a lower bound
on the mutual information between input x and representation y, and difference is obtained by
applying mapping fƟ to a corrupted input. Hence, such learning is cleverer than the identity,
and it extracts features useful for denoising.

Stack denoising autoencoder (SDAE) is a deep network utilizing the power of DAE (Bengio
et al., 2007; Vincent et al., 2010) and RBMs in the deep belief network (Hinton &
Salakhutdinov, 2006; Hinton et al., 2006).

4.1.3 Sparse autoencoder

The limitation of autoencoders to have only small numbers of hidden units can be overcome
by adding a sparsity constraint, where a large number of hidden units can be introduced
usually more than one input. The aim of sparse autoencoder (SAE) is to make a large number
of neurons to have low average output so that neurons may be inactive most of the time.

9
Sparsity can be achieved by introducing a loss function during training or manually zeroing
few strongest hidden unit activations. A schematic representation of SAE is shown in Fig.
5(d).

If the activation function of hidden neurons is aj, the average activation function of each
hidden neuron j is given by

∑[ ] ( )

The objective of sparsity constraints is to minimize so that , where is a sparsity

constraint very close to 0 such as 0.05.

To enforce sparsity constraints, a penalty term is added to cost function which penalizes ̂ ,
de-weighting significantly from . The penalty term is the Kullback-Leibler (KL) divergence
between Bernoulli random variables, can be calculated as (Ng, 2013; Makhzani & Frey,
2013),

∑ ( ̂) ( )

where is number of neurons in the hidden layers,and index is summing over the hidden
units in the network.

( ̂) ( ) ( )
̂ ̂

The property of penalty function is that ( ̂) , if ̂ , otherwise it increases

gradually as ̂ diverses for .

The k-sparse autoencoder (Makhzani & Frey 2013) is a form of sparse AE where k neurons
having the highest activation function are chosen and the rest is ignored. The advantage of k-
sparse AE is that they allow better exploration on a data set in terms of percentage activation
of the network. The advantage of SAE is the sparsity constraints which penalize the cost
function and as a result degrees of freedom is reduced. Hence, it regularizes and maintains
the complexity of the network by preventing over-fitting.

4.1.4 Convolutional autoencoder

The most popular and widely used network model in deep unsupervised architecture is
stacked AE. Stacked AE requires layer-wise pre-training. When layers go deeper during the
pre-training process, it may be time consuming and tedious because of stacked AE is built
with fully connected layers. Li et al. (2017) propose first trial to train convolutional directly
an end-to-end manner without pre-training. Guo et al. (2017) suggested convolutional
autoencoder (CAE) that is beneficial to learn feature for images and preserving the local
structure of data and avoid distortion of feature space. A general architecture of CAE is
depicted in Fig. 5(c).

10
Fig. 5 (a)-(g) Diagrams showing networks of autoencoders and its different variants

11
4.1.5 Variational autoencoder
Another variant of autoencoder, called variational autoencoder (VAE), was introduced as a
generative model (Kingma &Welling, 2013). A general architecture of VAE is given in Fig. 4(f).
VAEs utilize the strategy of deriving a lower bond estimator from the directed graphical models
with continuous distribution of latent variables. he generative parameter θ in the decoder
(generative model) assist the learning process of the variational parameter, ϕ as encoder in the
variational approximation model. VAEs apply the variational approach to latent representation,
learning as additional loss component training estimators, known as stochastic gradient
variational Bayes (SGVB) and Autoencoding variational Bayes (AEVB) (Kingma & Welling,
2013). It Optimizes the parameter ϕ and θ for probabilistic encoder qϕ(z|x), which is an
approximation to the generative model pθ(x, z), where z is latent variable and x is continuous or
discrete variable. Its aim is to maximize the probability of each x in the training data set under
entire generative process. However, alternative configuration of generative latent variable
modeling rises to give deep generative models (DGMs) instead of existing assumption of
symmetric Gaussian posterior (Partaourides at el., 2017).

4.1.6 Contractive autoencoder

Rifai (2011) presented a novel approach for training deterministic autoencoder. Contractive
autoencoder is additional of explicit regularizer in the objective function that enables the model to
learn a function having slight variations of input values. This additional regularizer corresponds
to the squared Forbenius norm of the Jacobian matrix of given activation with respect to the
input. The contractive autoencoder is obtained with the regularization term in following equation
yield final objective function,
( ) ∑. ( ( ( )) ‖ ( )‖ / ( )

The difference between contractive AE and DAE stated by (Vincent et al., 2010) as contractive
AE explicitly encourage robustness of representation, whereas DAE stressed on the robustness of
reconstruction this property make sense of contractive AE a better choice than DAEs to learn
useful feature extraction. Table 1 presents a summary of autoencoders and its variants, and Table
2 presents its applications for medical image analysis.

Table 1.Summary of autoencoders and its variants

Types Descriptions References
Autoencoder One of the simplest form which aims to learn Ballard (1987);
a representation (encoding) for a set of data. Bourlard & Kamp (1988)
Stacking autoencoder An autoencoder having multiple layers where Zabalza et al. (2016)
the outputs of each layers are given as inputs
of the successive layer.
Sparse autoencoder Encourages hidden units to be zero or near to Goodfellow et al. (2009)
zero
Denosing autoencoder Capable to predict true inputs from noisy LeCun & Gallinari, (1987);
data Vincent et al. (2008)
Convolutional autoencoder Learn feature, preserve the local structure of Guo et al. (2017)
data and avoid distortion of feature space
Variational autoencoder A generative model utilizing strategy of Kingma & Welling (2013)
deriving a lower bond estimator from
directed graphical models with continuous
distribution of latent variables.
Contractive autoencoder Forces encoder to take small derivatives Rifai et al. (2011)

12
Table 2 Applications of autoencoders and its variants for medical image analysis.
[Abbreviations: H&E: hematoxylin and eosin staining; AD: Alzheimer’s disease; MCI: Mild cognitive
impairment; fMRI: Functional magnetic resonance imaging; sMRI: Structural magnetic resonance imaging; rs-
fMRI: Resting-state fMRI; DBN: Deep belief network; RBM: Restricted Boltzmann machine]

Method Task Image type Remarks References

SAE AD/MCI classification MRI SAE accompanied by supervised fine Suk & Shen (2013)
tuning
SAE AD/MCI/HC MRI & PET Extraction of latent features on a huge Suk et al. (2013a)
classification set of features obtained from MRI and
PET images using SAE
SAE AD/MCI/HC MRI SAE used to pre-train 3D CNN Payan & Montana
classification (2015)
SAE MCI/HC classification fMRI SAE used for feature extraction, HMM Suk et al. (2016)
as a generative model on top
SAE Hippocampus MRI SAE used for representation learning Guo et al. (2014)
segmentation and measure target/atlas patch
SAE Visual pathway MRI SAE used to learn appearance features Mansoor et al. (2016)
segmentation to steer the shape model for
segmentation
SAE Denoising DCE-MRI MRI Uses an ensemble of denoising SAE Benou et al. (2016)
(pre-trained with RBMs). Denoising
contrast-enhanced MRI sequences using
expert DNNs (pre trained with RBMs)
SSAE Nucleus detection Digital detection of nuclei on breast cancer Xu et al. (2016)
pathology digital histopathological images.
image
SAE Stain normalization Digital SAE is applied to classify tissues and Janowczyk et al.
pathology their subsequent histogram (2017)
image Matching
SAE Density classification Mammography Unsupervised CNN with SAE to learn Kallenberg et al.
features from unlabeled data for breast (2016)
texture and density classification
SAE Lesion classification MRI Learn to extract features from multi- Zhu et al. (2017)
parametric MRI data, subsequently
creates a hierarchical classification to
detect prostate cancer.
SAE Detection of Heart, MRI SAE used for acquisition of spatio- Shin et al. (2013)
kidney and liver temporal features on 2D along with time
location DCE-MRI
SAE Cell segmentation Digital Learning spatial relationships Hatipoglu, N. 2017
pathology
image
SAE Segmentation right MRI SAE applied to obtain an initial right Avendi, M. 2017
ventricle in cardiac ventricle segmentation.
MRI
SDAE Cell segmentation Digital The SDAE trained with data and their Su. H. at el 2018
pathology structured labels for cell segmentation
image
SSAE AD MRI SSAE for early detection of Liu et al. (2014)
Alzheimer’s disease from brain MRI
SDAE Breast lesion Ultrasound and Stacked Denoising AE for Diagnosis of Cheng et al. (2016)
CT breast nodules and lesions
SDAE Patient clinical events Patient clinical SDAE for an unsupervised early Miotto et al. (2016)
history prediction of patients e future clinical
events and disease.
SDAE -- CT/MRI Multi-modal SDAE used to pre-train the Cheng et al. (2018)
DNN.
DCAE Modeling task fMRI tfMRI Deep Convolutional AE to model Huang et al. (2018)
tfMRI.
CAE AD/MCI/HC fMRI CAE used to pre-train 3D CNN. Hosseini-Asl et al.
classification (2016)
CAE Nucleus detection Digital Sparse CAE to detect and encode nuclei Hou et al. (2019)
pathology and feature extraction from tissue
image section images.

13
4.2. Restricted Boltzmann Machines
Restricted Boltzmann Machines (RBMs) are a variant of Markov Random Field (MRF),
constitute of single layer undirected graphical model with an input layer or visible layer x = (x 1,
x2...... xN) and a hidden layer h = {h1, h2, …. HM}. The connection between nodes/units are
bidirectional, so each given input vector x can take the latent feature representation h and vice-
versa. An RBM is a generative model which learns probability distribution over the given input
space and generates new data point (Yoo, et al. 2014). Illustration of a typical RBM is shown in
Fig. 6(a). In fact, RBMs are restricted version of Boltzmann machines where neurons must form
an arrangement of bipartite graphs. Due to this restriction, pairs of nodes belonging to each of the
visible and hidden nodes have a symmetric connection between them, and nodes within a group
have no internal connections.. This restriction makes RBM more efficient training algorithm than
the general case of Boltzmann machine. Hinton et al. (2010) proposed a practical guide to train
RBMs.

RBMs have been utilized in various aspects of medical image analysis such as detection of
variations in Alzheimer disease (Brosch, et al. 2013), image segmentation (Yoo et al. 2014),
dimensionality reduction (Cheng et al. 2016), feature learning (Pereira et al. 2018) and so on. A
brief account for the application of RMBs in medical image analysis is shown in Table 3.
Table 3. Applications of RBM for medical image analysis

Method Task Image type Remarks References

RBM AD MRI Uses a large dataset of MRI to rule Brosch et al. (2013)
out the mode of variations in AD
brains.
RBM Multiple sclerosis 3DMRI Uses multi-channel 3D MR images Yoo et al. (2014)
lesions of multiple sclerosis (MS) lesion
for MS segmentation
RBM AD/MCI/HC MRI, PET DBMs on multimodal images from Suk et al. (2014)
classification MRI and PET scans for disease
classification.
RBM Mass detection in breast Mammography RBM based method for Cao et al. (2015)
cancer oversampling and semi-supervised
learning to solve classification of
imbalanced data with a few labeled
samples
RBM fMRI blind source fMRI RBM used for both internal and Huang et al. (2016)
separation functional interaction-induced latent
source detection
RBM Vertebrae localization CT, MRI RBMs to locate the exact position of Cai et al. (2016b)
the vertebrae.
RBM Benign/Malignant Ultrasound Shear wave elsatrography for class Zhang et al. (2016a)
classification indication of benign and malignant
mammary gland tumors using RBM.
RBM Tongue contour Ultrasound Analysis of tongue motion during Jaumard-Hakoun et al.
extraction speech, using auto encoders in (2016)
combination with RBM.
CRBM Lung tissue CT Discriminative and generative Van Tulder & de Bruijne
classification and learning by CRBM to develop filters (2016)
airway detection for data training as well
classification.
RBM Cardiac arrhythmia ECG Achieves average recognition Mathews et al. (2018)
classification accuracy for ventricular and
supraventricular ectopic beats
(93.63% and 95.57%, respectively)
for Cardiac arrhythmia
classification.
RBM Brain lesion MRI RBM is used for feature learning, Pereira et al. (2018)
segmentation and a Random Forest as a classifier.

14
 Fig. 6 (a)-(d) Diagrams showing various unsupervised network models

4.3. Deep Belief Networks

Deep Belief Networks (DBN) is a kind of neural network proposed by Bengio (2009). It is a
greedy layer-wise unsupervised learning algorithm with several layers of hidden variables
(Hinton et al., 2016). Layer-wise unsupervised training (Bengio 2007) help the optimization
and weight initialization for better generalization. In fact, DBN is a hybrid single
probabilistic generative model, like a typical RBM. In order to construct a deep architecture
like SAEs where AEs layers are replaced by RBMs, DBN has one lowest visible layer v,
representing state of input data vector and a series of hidden layers h1, h2, h3, . . . hL. When
multiple RBMs are stacked hierarchically, an undirected generative model is formed by top
two layers and directed generative model is formed by lower layers. Fig. 6(b), illustrates the

15
structure of DBN. The following function in DBN represents the joint distribution of visible
unit v, hidden layers hl (l = 1, 2…. L :

) ( ) ( ( )
P( (∏ )) (14)

Hinton at el. (2006a) applied layer-wise training procedure, where lower layers learns low-
level features and subsequently higher layers learns high-level features (Hinton at el. 1995).
DBN are used to extract features from fMRI images (Plis et al., 2014), temporal ultrasound
(Azizi et al. 2016), classify Autism spectrum disorders (Aghdam et al. 2018), and so on.
Some of the applications of DBNs are presented in Table 4.

Table 4. Applications of DBNs for medical image analysis

Method Task Image type Remarks References

DBN AD/HC classification MRI DBNs with convolutional RBMs Brosch & Tam (2013)
for manifold learning
DBN, Manifold Learning MRI DBM along with convolutional Brosch et al. (2014)
Convolutional RBM layers to efficiently train
RBM DBMs in order to detect
morphological changes in brain in
normal as well as disease
conditions
DBN MRI Evaluation of DBN to estimate Plis et al. (2014)
brain networks in neurocognitive
disorders like Huntington’s
disease and Schizophrenia
DBN AD/MCI/HC MRI A group of voting schemes Ortiz et al. (2016)
classification clubbed using an SVM to better
classify AD and MCI from brain’s
3D gray mater images.
DBN Left ventricle Ultrasound DBN assisted system exploiting Carneiro et al. (2012);
segmentation non-rigid registration, landmarks Carneiro &
and patches to maneuver multi Nascimento (2013)
atlas segmentation.
DBN Schizophrenia/NH MRI Characterizing differences in Pinaya et al. (2016)
classification morphology of various brain
regions in schizophrenia using
DBN and supervised fine tuning.
DBN Lesion classification Ultrasound Training DBN to extract features Azizi et al. (2016)
from prostate ultrasonography
images to classify benign and
malignant lesions.
DBN Left ventricle MRI The combination of DBN and level Ngo et al. (2017)
segmentation set method to yield automated
segmentation of the left ventricle
from cardiac cine MRI
DBN Cardiac arrhythmia ECG Achieves average recognition Mathews et al. (2018)
classification accuracy of ventricular and
supraventricular ectopic beats
(93.63% and 95.57%, respectively)
for cardiac arrhythmia
classification.
DBN Autism spectrum rs-fMRI, Classifies Autism spectrum Aghdam et al. (2018)
disorders classification sMRI disorders (ASDs) in children using
rs-fMRI and sMRI data on the
basis of Random Neural Network
clustering.

16
4.4. Deep Boltzmann Machine

Deep Boltzmann machine (DBM) is a robust deep learning model proposed by Salakhutdinov
et al. (2009) and Salakhutdinov et al. (2012). They stacked multiple RBMs in a hierarchal
manner to handle ambiguous input robustly. Fig. 6(c) represents the architecture of DBM as a
composite model of RBMs which clearly shows how DBM differ from DBN. Unlike DBNs,
DBMs form undirected generative model combining information from both lower and upper
layers which improves the representation power of DBMs. Training of layer-wise greedy
algorithm for DBM (Salakhutdinov et al., 2015; Goodfellow et al., 2013b) is calculated by
modifying in procedure of DBN.

Recently, a three-layer DBM was presented by Salakhutdinov et al. (2015) and Dinggang et
al. (2017). In this three-layer DBM, to learn parameters * + the values of
neighbour layer(s) and probability of visible and hidden units are computed using logistic
sigmoidal function. The derivative of log likelihood of the observation ( ) with respect to
the model parameter( ) is computed as,

( ) , ( ) - , ( ) - ( )
()

Where [.] denote data-dependent obtained from visible units and [.] denote
data-independence obtained from the model. Some of the applications of DBMs are shown in
Table 5.

Table 5. Applications of DBMs for medical image analysis

Method Task Image type Remarks References

DBM Heart motion tracking MRI Using three-layered Deep Wu, et al., (2018)
Boltzmann Machine to guide
frame-by-frame heart
segmentation during radiation
therapy of cancer patient on cine
MRI images.
DBN AD/HC classification MRI DBN combined with Brosch & Tam
convolutional RBMs for (2013)
manifold learning.
RBM Breast-image MRI Restricted Boltzmann machine Nahid et al., (2018)
classification with backpropagation have been
used for histopathological breast-
image classification
DBM Medical image retrieval Multi digital DBM based multi model learning Cao, et al., (2014)
image to learn joint density model.

4.5. Generative Adversarial Network (GAN)

Generative Adversarial Network (GAN) (Goodfellow, et al. 2014) is one of recent promising
technique for building flexible deep generative unsupervised architecture. Goodfellow et al.
(2014) proposed two models generative model G and Discriminative model D, where G
capture data distribution (pg)over real data t, and D estimates the probability of a sample
coming from training data (m) not from G. In every iteration, backpropagation generator and
discriminator competing with each other. The training procedure the probability of D is

17
maximized. This framework functions like a mini-max two-player game. The value function
V(G, D) establishes following two-player mini-max game is given by,

⏟⏟ ( ) , ( )- ( )0 . ( ( ))/1 ( )

Where D(t) represents the probability of t from data m and pdata is distribution of real-world
data. This model seems to be stable and improved as pg = pdata. A typical architecture of GAN
is depicted in Fig. 6(d). In fact, these two adversaries, Generator and Discriminator,
continuously battle during the processing of training. GAN have been applied to generate
samples of photorealistic images to visualize new designs. Some of the applications of GAN
for medical image analysis are presented in Table 6.

Table 6. Applications of GAN for medical image analysis

Method Task Image Remarks References

type
GAN Synthesis of retinal Retinal MI-GAN generates precise Iqbal & Ali (2018)
images images segmented images for the
application of supervised
learning of retinal images.
GAN Chest X-ray X-ray GAN used to produce Canas, et al, (2018)
photorealistic images which
retain pathological quality
Dual GAN- Segmentation of --- Improve GAN using dual- Bi, et al., (2018)
FCN regions of interest path adversarial learning for
(ROIs) Fully Convolutional Network
based image segmentation
GAN Simulation of B- Ultrasound Conditional generative Hu et al., (2017)
mode ultrasound adversarial networks used to
images simulate ultrasound images at
given 3D spatial locations.
GAN Treatment of PET Multi-channel generative Bi, et al., (2017)
lymphomas and lung adversarial networks used to
cancer synthesize PET data.

5. List of software tools/packages and benchmark datasets

A plethora of software tools and packages implementing unsupervised learning models (as
discussed in the paper) has been developed and made available to the research community
and data analysts. Some of the tools/packages and medical images benchmark datasets are
listed in Table 7 and Table 8, respectively.

18
 Table 7. List of software tools/packages for unsupervised learning models
S. No. Tools/ Packages Models/ Description Language URL
Name Methods /Technology
1. deeplearning4 Autoencoders Deep learning APIs for Java Java https://deeplearning4j.org/
j having an implementation of
several deep learning techniques.
2. unsup under Autoencoder, A scientific computing Lua https://github.com/torch/to
torch7 etc. framework with good support for rch7
machine learning algorithms that
puts GPUs first. Unsup package
provides few unsupervised
learning algorithms such as
autoencoders, clustering, etc.
3. DeepPy Autoencoders MIT licensed deep learning Python https://github.com/andersb
framework that runs on CPU or ll/deeppy
GPUs and implements http://andersbll.github.io/d
autoencoders, in addition to other eeppy-website/
supervised learning algorithms.
4. SAENET.train Stacked Build a stacked autoencoder in R R package https://rdrr.io/cran/SAEN
autoencoder environment for pre-training of ET/man/SAENET.train.ht
feed-forward NN and dimension ml
reduction of features.
5. kdsb17 Convolutional Gaussian Mixture Convolutional Python, https://github.com/alegonz
autoencoder Autoencoder (GMCAE) used for Keras, /kdsb17
CT lung scan using Tensor-flow-
Keras/TensorFlow gpu
6. autoencoder Deep Training a deep autoencoder for Matlab http://www.cs.toronto.edu/
autoencoder MNIST digits datasets ~hinton/code/Autoencoder
_Code.tar
7. H2O Deep Parallelized implementations of R package https://cran.r-
autoencoder many supervised and project.org/web/packages/
unsupervised machine learning h2o/
algorithms, including GLM,
GBM, RF, DNN, K-Means,
PCA,Deep AE, etc.
8. dbn DBN Deep belief network pre-train in R package https://rdrr.io/github/Timo
unsupervised manner with stacks Matzen/RBM/src/R/DBN.
of RBM, which in return fine- R
tuned DBN.
9. darch DBN, RBM Restricted Boltzmann machine, R package https://github.com/maddin
deep belief network 79/darch
implementation
10. deepnet DBN, RBM, Implementation of RBM, DBN, R package https://cran.r-
deep deep stacked autoencoders project.org/web/packages/
autoencoders deepnet/
11. Vulpes DBN DBN and other deep learning Visual https://github.com/fsproje
implementation in F#. Studio cts/Vulpes
12. pydbm DBM/ RBM RBM/DBM are implemented in Python https://pypi.org/project/py
python for pre-learning or dbm/
dimension reduction
13. RBM RBM Simple RBM implementation in Python https://github.com/echen/r
Python estricted-boltzmann-
machines
14. xRBM RBM and its Implementation of RBM and its Python https://github.com/omimo/
variants variants in Tensorflow xRBM
15. DCGAN.torch GAN Unsupervised representation Lua https://github.com/soumit
learning using Deep h/dcgan.torch
Convolutional GAN
16. pix2pix GAN Conditional Adversarial Networks Linux Shell https://github.com/phillipi/
for Image-to-image translation Script pix2pix
synthesizing from the image.
17. ebgan GAN Energy-based GAN equivalent to Python https://github.com/eriklind
probabilistic GANs produces high ernoren/PyTorch-
resolution images. GAN/tree/master/impleme
ntations/ebgan

19
 Table 8. List of benchmark medical image datasets
[Abbreviations. ADNI: Alzheimer’s Disease Neuroimaging Initiative; ABIDE: Autism Brain Imaging Data Exchange; DICOM: Digital
Imaging and Communications in Medicine; BCDR: Breast Cancer Digital Repository; CIVM: Center for in Vivo Microscopy; DDSM:
Digital Database for Screening Mammography; DRIVE: Digital Retinal Images for Vessel Extraction; IDA: Image & Data Archive; ISDIS:
International Society for Digital Imaging of the Skin; NBIA: National Biomedical Imaging Archive; OASIS: Open Access Series of
Imaging Studies; TCGA: The Cancer Genome Atlas; TCIA: The Cancer Imaging Archive]

S. No. Data set Modalities Medical Accessibility URL

condition
1. ABIDE MRI Autism spectrum Open access http://fcon_1000.projects.nitrc.org/indi/abide/
disorder
2. ADNI MRI Alzheimer’s Paid http://adni.loni.usc.edu/data-samples/access-
disease data/
3. BCDR Mammography Breast cancer Open access https://bcdr.eu/

4. CIVM 3D-MRM Histology of the Limited http://www.civm.duhs.duke.edu/devatlas/

Embryonic and access
Neonatal Mouse
5. DDSM Mammography Breast cancer Open access http://marathon.csee.usf.edu/Mammography/
Database.html
6. DermNet Photo A huge database Limited http://www.dermnet.com/
dermatology of various skin access
diseases
7. DICOM MRI, CT, etc. A variety of Open access https://www.dicomlibrary.com
medical images,
videos and signal
files
8. DRIVE 2D color Retinal blood Open access http://www.isi.uu.nl/Research/Databases/DRI
images of vessel VE/download.php
retina segmentation to
study diabetic
retinopathy
9. IDA An online Open access https://ida.loni.usc.edu/
resource for
neuroscience
images
10. ISDIS Dermoscopy, Skin disease Paid https://isdis.org/
telemedicine,
spectroscopy
etc.
11. MedPix Variety of Online database Open access https://medpix.nlm.nih.gov
imaging data of medical
images, teaching
cases, and clinical
topics
12. NBIA CT, PT, MRI, A database of the Limited/ https://imaging.nci.nih.gov/
etc. National Cancer open access
Institute proving
medical images of
various conditions
and anatomical
sites.
13. OASIS MRI and PET Normal aging or Open access http://www.oasis-brains.org/
mild to moderate
Alzheimer's
Disease
14. TCIA Collection of Multimodal image Limited/ http://www.cancerimagingarchive.net/
MRI, CT etc. archive for open access
various types of
cancer
15. TCGA Histopathology Histopathology Open https://cancergenome.nih.gov/
slide images slide images from
sample portions of
various types of
cancers

20
6. Discussion, opportunities and challenges

Medical imaging and diagnostic techniques are one of the most widely used for early
detection, diagnosis and treatment of complex diseases. After significant advancement in
machine learning and deep learning (both supervised and unsupervised), there is a paradigm
shift from the manual interpretation of medical images by human experts such as radiologists
and physicians to an automated analysis and interpretation, called computer-assisted
diagnosis (CAD). As unsupervised learning algorithms can derive insights directly from data,
use them for data-driven decisions making, and are more robust, hence they can be utilized as
the holy grail of learning and classification problems. Furthermore, these models are also
utilized for other important tasks including compression, dimensionality reduction, denoising,
super resolution and some degree of decision making.

Unsupervised learning and CAD, both being in its infancy, researchers and practitioners have
much opportunity in this area. Some of them are: (i) Allow us to perform exploratory analysis
of data (ii) Allow to be used as preprocessing for supervised algorithm, when it is used to
generate a new representation of data which ensure learning accuracy and reduces memory
time overheads. (iii) Recent development of cloud computing, GPU-based computing,
parallel computing and its cheaper cost allow big data processing, image analysis and execute
complex deep learning algorithms very easily.

Some of the challenges and research directions are:

(i) Difficult to evaluate whether algorithm has learned anything useful: Due to lack of
label in unsupervised learning, it is nearly impossible to quantify its accuracy. For instance,
how can we access whether K-means algorithm found the right clusters? In this direction,
there is a need to develop algorithms which can give an objective performance measure in
unsupervised learning.

(ii) Difficult to select right algorithm and hardware: Selection of right algorithm for a
particular type of medical image analysis is not a trivial task because performances of the
algorithm are highly dependent on the types of data. Similarly, hardware requirement also
varies from problem to problem.

(iii) Will unsupervised learning work for me? It is mostly asked question, but its answer
totally depends on the problem at hand. In image segmentation problem, clustering algorithm
will only work if the images do fit into naturals groups.

(iv) Not a common choice for medical image analysis: Unsupervised learning is not a
common choice for medical image analysis. However, from literature it is revealed that these
(autoencoders and its variants, DBN, RBM, etc.) are mostly used to learn the hierarchy level
of features for classification tasks. It is expected that unsupervised learning will play pivotal
role in solving complex medical imaging problems which are not only scalable to large
amount of unlabeled data, but also suitable for performing unsupervised and supervised
learning tasks simultaneously (Yi et al., 2018).

21
(v) Development of patient-specific anatomical and organ model: Anatomical skeletons
play crucial role in understanding diseases and pathology. Patient-specific anatomical model
is frequently used for surgery and interventions. They help to plan procedure, perform
measurement for device surging, and predict the outcome of post-surgery complexities.
Hence, the algorithm needs to be developed to construct patient-specific anatomical and
organ model from medical images.

(vi) Heterogeneous image data: In the last two to three decades, more emphasis was given
to well-defined medical image analysis applications, where developed algorithms were
validated on well-defined types of images with well-defined acquisition protocol. The
algorithms are required, which can work on more heterogeneous data.

(vii) Semantic segmentation of images: Semantic segmentation is task of complete scene

understanding, leading to knowledge inference from imagery. Scene understanding is a core
of computer vision problems which has several applications, including human-computer
interaction, self-driving vehicles, virtual reality, and medical image analysis. The semantic
segment of medical images with acceptable accuracy is still challenging.

(viii) Medical video transmission: Enabling 3D video in recently adopted telemedicine and
U-healthcare applications result in more natural viewing conditions and better diagnosis.
Also, remote surgery can be benefited from 3D video because of additional dimensions of
depth. However, it is crucial to transmit data-hungry 3D medical video stream in real-time
through limited bandwidth channels. Hence, efficient encoding and decoding techniques for
3D video data transmission is required.

(ix) Need extensive inter-organizational collaborations: Inter-professional and inter-

organizational collaboration is important for better functioning of the health care system,
eliminating some of the pitfalls such as limited resources, lack of expertise, aging
populations, and combat chronic diseases (Karam et al., 2017). Medical image based CAD
needs extensive inter-organizational collaborations among doctors, radiologists, medical
image analysts, and computational data analysts.

(x) Need to capitalize big medical imaging market: According to IHS Markit report
(https://technology.ihs.com.), medical imaging market has total global revenue of $21.2
billion in 2016, which is forecasted to touch $24.0 billion by 2020. According to WHO,
global population will rise from 12% to 22% from 2015 to 2050. Population aging lead to
increased rate of chronic diseases globally and hence there is a need to capitalize a big
medical imaging market worldwide.

(xi) Black-box and its acceptance by health professionals: Machine learning algorithms
are boon which solves the problems earlier thought to be unsolvable, however, it suffers from
being “black-box”, i.e., how output arrives from the model is very complicated to interpret.
Particularly, deep learning models are almost non-interpretable and but still being used for
complex medical image analysis. Hence, its acceptance by health professionals is still
questionable.

22
(xii) Will technology replace radiologists? For the processing of medical images, deep
learning algorithms help select and extract important features and construct new ones, leading
to new representation of images, not seen before. For image interpretation side, deep learning
helps identify, classify, quantify disease patterns, allow measure predictive targets, and make
predictive models, and so on. So, will technology “replace radiologists”, or migrate to
“virtual radiologist assistant” in near future? Hence, following slogan is quite relevant in this
context: “Embrace it, it will make you stronger; reject it, it may make you irrelevant”.

In a nutshell, unsupervised learning is very much open topic where researchers can make
contributions by developing a new unsupervised method to train how network (e.g. Solve a
puzzle, generate image patterns, image patch comparison, etc.) and re-thinking of creating a
great unsupervised feature representation, (e.g. What is the object and what is the
background?), nearly analogous to the human visual system.

7. Conclusion
Medical imaging is one of the important techniques for early detection, diagnosis and
treatment of complex diseases. Interpretation of medical images is usually performed by
human experts such as radiologists and physicians. After the success of machine learning
techniques, including deep learning, availability of cheap computing infrastructure through
cloud computing, there has been a paradigm shift in the field of computer-assisted diagnosis
(CAD). Both supervised and unsupervised machine learning approaches are widely applied in
medical image analysis, each of them with their own pros and cons. Due to the fact that
human supervisions are not always available or inadequate or biased, therefore, unsupervised
learning algorithms, including its deep architecture, give a big hope with lots of advantages.

Unsupervised learning algorithms derive insights directly from data, and use them for data-
driven decisions making. Unsupervised models are more robust and they can be utilized as
the holy grail of learning and classification problems. These models are also used for other
tasks including compression, dimensionality reduction, denoising, super resolution and some
degree of decision making. Therefore, it is better to construct a model without knowing what
tasks will be at hand and we would use representation (or model) for. In a nutshell, we can
think of unsupervised learning as preparation (preprocessing) step for supervised learning
tasks, where unsupervised learning of representation may allow better generalization of a
classifier.

Acknowledgements
Authors would like to thank Ms. Sahar Qazi, Ms. Almas Jabeen, and Mr. Nisar Wani for
necessary support.

Conflict of Interest Statement

Authors declare that there is no any conflict of interest in the publication of this manuscript.

23
References
Aghdam, M. A., Sharifi, A., & Pedram, M. M. (2018). Combination of rs-fMRI and sMRI Data to Discriminate
Autism Spectrum Disorders in Young Children Using Deep Belief Network. Journal of digital imaging,
1-9. https://doi.org/10.1007/s10278-018-0093-8
Avendi, M. R., Kheradvar, A., &Jafarkhani, H. (2017). Automatic segmentation of the right ventricle from
cardiac MRI using a learning-based approach. Magnetic Resonance in Medicine, 78(6), 2439–2448.
https://doi.org/10.1002/mrm.26631
Azizi, S., Imani, F., Ghavidel, S., Tahmasebi, A., Kwak, J.T., Xu, S., Turkbey, B., Choyke, P., Pinto, P., Wood,
B., Mousavi, P., Abolmaesumi, P. (2016). Detection of prostate cancer using temporal sequences of
ultrasound data: a large clinical feasibility study. Int. J. Comput. Assist. Radiol. Surg. 11 (6), 947–956.
https://doi.org/10.1007/s11548-016-1395-2
Ballard, D. H. (1987). Modular Learning in Neural Networks. In AAAI (pp. 279-284).
Bengio Y, Courville A, Vincent P. (2013). Representation learning: a review and new perspectives. IEEE Trans
Pattern Anal Mach Intell. 35:1798–828. https://doi.org/10.1109/TPAMI.2013.50
Bengio, Y. (2009). Learning Deep Architectures for AI. Foundations and Trends in Machine Learning, 2(1), 1-
127. https://doi.org/10.1561/2200000006
Bengio, Y., Lamblin, P. , Popovici, D. , Larochelle, H. (2007). Greedy layer-wise training of deep networks. In:
Proceedings of the Advances in Neural Information Processing Systems, pp. 153–160.
Benou, A., Veksler, R. , Friedman, A. , Raviv, T.R. (2016). De-noising of contrast-enhanced MRI sequences by
an ensemble of expert deep neural networks. In: Deep Learning and Data Labeling for Medical
Applications (pp. 95-110). Springer, Cham. https://doi.org/10.1007/978-3-319-46976-8_11
Bi, L., Feng, D., Kim, J. (2018). Dual-Path Adversarial Learning for Fully Convolutional Network (FCN)-Based
Medical Image Segmentation, Visual Computer, 34(6-8), 1043-1052. https://doi.org/10.1007/s00371-
018-1519-5
Bi, L., Kim, J., Kumar, A., Feng, D., Fulham, M. (2017). Synthesis of positron emission tomography (PET)
images via multi-channel generative adversarial networks (GANs). Lecture Notes in Computer Science,
10555, pp. 43-51. https://doi.org/10.1007/978-3-319-67564-0_5
Bishop CM. (2006). Pattern Recognition and Machine Learning (Information Science and Statistics), 1st edn.
Springer, New York..
Bourlard H, Kamp Y. (1988). Auto-association by multilayer perceptrons and singular value decomposition.
Biological Cybernetics, 59, 291–94. https://doi.org/10.1007/BF00332918
Brosch, T., Tam, R. (2013). Manifold learning of brain MRIs by deep learning. In: Proceedings of the Medical
Image Computing and Computer-Assisted Interven- tion. In: Lecture Notes in Computer Science, 8150,
pp. 633–640. https://doi.org/10.1007/978- 3- 642-40763-5_78
Brosch, T., Yoo, Y., Li, D. K. B., Traboulsee, A., Tam, R. (2014). Modeling the variability in brain morphology
and lesion distribution in multiple sclerosis by deep learning. In: Med Image Comput Comput Assist
Interv. Lecture Notes in Computer Science, 8674 (pp. 462–469).
Cai, Y., Landis, M., Laidley, D. T., Kornecki, A., Lum, A., Li, S. (2016b). Multi-modal vertebrae recognition
using transformed deep convolution network. Comput Med Imaging Graph, 51, 11–19.
Canas, K., Liu, X., Ubiera, B., Liu, Y. (2018). Scalable biomedical image synthesis with GAN. ACM
International Conference Proceeding Series, art. no. a95. https://doi.org/10.1145/3219104.3229261
Cao, P., Liu, X., Bao, H., Yang, J., & Zhao, D. (2015). Restricted Boltzmann machines based oversampling and
semi-supervised learning for false positive reduction in breast CAD. Bio-medical materials and
engineering, 26(s1), S1541-S1547. https://doi.org/10.3233/BME-151453
Cao, Y., Steffey, S., He, J., Xiao, D., Tao, C., Chen, P., Müller, H. (2014). Medical image retrieval: A
multimodal approach. Cancer Informatics, 125-136. https://doi.org/10.4137/CIN.S14053

24
Carneiro, G., Nascimento, J.C. (2013). Combining multiple dynamic models and deep learning architectures for
tracking the left ventricle endocardium in ultrasound data. IEEE Trans. Pattern Anal. Mach. Intell. 35,
2592–2607. https://doi.org/10.1109/TPAMI. 2013.96
Carneiro, G., Nascimento, J.C., Freitas, A. (2012). The segmentation of the left ven- tricle of the heart from
ultrasound data using deep learning architectures and derivative-based search methods. IEEE
Transactions on Image Processing, 21(3), 968–982. https://doi.org/10.1109/TIP.2011.2169273
Chan, T. H., Jia, K., Gao, S., Lu, J., Zeng, Z., & Ma, Y. (2015). PCANet: A simple deep learning baseline for
image classification?. IEEE Transactions on Image Processing, 24(12), 5017-5032.
https://doi.org/10.1109/TIP.2015.2475625
Cheng J-Z, Ni D, Chou Y-H, et al. (2016). Computer-aided diagnosis with deep learning architecture:
applications to breast lesions in US images and pulmonary nodules in CT scans. Scientific Reports, 6,
24454. https://doi.org/10.1038/srep24454
Cheng, Li Zhang & Yefeng Zheng. (2018). Deep similarity learning for multimodal medical images, Computer
Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 6:3, 248-252.
https://doi.org/10.1080/21681163.2015.1135299
Dinggang S. & Wu, Guorong& Suk, Heung-Il. (2017). Deep Learning in Medical Image Analysis. Annual
review of biomedical engineering, 19. https://doi.org/10.1146/annurev-bioeng-071516-044442
Drineas, Petros & Frieze, Alan & Kannan, Ravindran & Vempala, Santosh & Vinay, V. (1999). Clustering in
Large Graphs and Matrices. In Proceedings of the 10th ACM-SIAM Symposium on Discrete
Algorithms(pp. 291-299).
Fischer, P. Pohl, T. Faranesh, A., Maier, A. and Hornegger, J. (2017). Unsupervised Learning for Robust
Respiratory Signal Estimation From X-Ray Fluoroscopy, IEEE Transactions on Medical Imaging, 36(4),
865-877. https://doi.org/10.1109/TMI.2016.2609888
Gallinari, Y. LeCun, S. Thiria, and F. Fogelman-Soulie. Memoires associatives distribuees. In Proceedings of
COGNITIVA 87, Paris, La Villette, 1987
Goodfellow, I. J., Mirza, M., Courville, A., and Bengio, Y. (2013b). Multi-prediction deep Boltzmann
machines. In NIPS’2013.
Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, SherjilOzair, Aaron Courville,
and YoshuaBengio. (2014). Generative adversarial nets. In Advances in Neural Information Processing
Systems. Curran Associates, 2672–2680.
Goodfellow, Quoc Le, Andrew Saxe, and Andrew Ng. (2009). Measuring invariances in deep networks. In
Yoshua Bengio, Dale Schuurmans, Christopher Williams, John Lafferty, and Aron Culotta, editors,
Advances in Neural Information Processing Systems 22 NIPS’09 , pages 646–654.
Guo, X., Liu, X., Zhu, E., & Yin, J. (2017, November). Deep clustering with convolutional autoencoders.
In International Conference on Neural Information Processing (pp. 373-382). Springer, Cham.
Guo, Y., Wu, G., Commander, L. A., Szary, S., Jewells, V., Lin, W., & Shent, D. (2014). Segmenting
hippocampus from infant brains by sparse patch matching with deep-learned features. Medical image
computing and computer-assisted intervention : MICCAI ... International Conference on Medical Image
Computing and Computer-Assisted Intervention, 17(Pt 2), 308-15.
Hatipoglu, N., &Bilgin, G. (2017). Cell segmentation in histopathological images with deep learning algorithms
by utilizing spatial relationships. Medical & Biological Engineering & Computing, 55(10), 1829–1848.
https://doi.org/10.1007/s11517-017-1630-1
Hinton G, Dayan P, Frey B,NealR. (1995). he “wake–sleep” algorithm for unsupervised neural networks.
Science, 268:1158–61. https://doi.org/10.1126/science.7761831
Hinton GE, Salakhutdinov RR. (2006). Reducing the dimensionality of data with neural networks. Science
313:504–7. https://doi.org/10.1126/science.1127647
Hinton, G. , 2010. A practical guide to training restricted boltzmann machines. Momentum 9 (1), 926

25
Hinton, G.E., Osindero, S., Teh, Y.-W., 2006a. A fast learning algorithm for deep belief nets. Neural Comput.
18, 1527–1554. https://doi.org/10.1162/neco.2006.18.7.1527
Hosseini-Asl, E., Gimel’farb, G., El-Baz, A. (2016). Alzheimer’s disease diagnostics by a deeply supervised
adaptable 3D convolutional network. arxiv: 1607.00556 .
Hou, L., Nguyen, V., Kanevsky, A. B., Samaras, D., Kurc, T. M., Zhao, T., ... & Saltz, J. H. (2019). Sparse
Autoencoder for Unsupervised Nucleus Detection and Representation in Histopathology Images. Pattern
Recognition, 86: 188-200. https://doi.org/10.1016/j.patcog.2018.09.007
Hu, Y., Gibson, E., Lee, L.-L., Xie, W., Barratt, D.C., Vercauteren, T., Noble, J.A. (2017). Freehand ultrasound
image simulation with spatially-conditioned generative adversarial networks. Lecture Notes in Computer
Science, 10555 (pp. 105-115). https://doi.org/10.1007/978-3-319-67564-0_11
Huang, H., Hu, X., Han, J., Lv, J., Liu, N., Guo, L., Liu, T., 2016. Latent source mining in FMRI data via deep
neural network. In: Proceedings of the IEEE International Symposium on Biomedical Imaging, pp. 638–
641. https://doi.org/10.1109/ISBI.2016.7493348
Huang, H., Hu, X., Zhao, Y., Makkie, M., Dong, Q., Zhao, S., ... & Liu, T. (2018). Modeling task fMRI data via
deep convolutional autoencoder. IEEE transactions on medical imaging, 37(7), 1551-1561.
Iqbal, T., Ali, H. Generative Adversarial Network for Medical Images (MI-GAN) (2018) Journal of Medical
Systems, 42 (11), art. no. 231. https://doi.org/10.1007/s10916-018-1072-9
Jabeen, A., Ahmad, N., & Raza, K. (2018). Machine Learning-Based State-of-the-Art Methods for the
Classification of RNA-Seq Data. In Classification in BioApps (pp. 133-172). Springer, Cham.
https://doi.org/10.1007/978-3-319-65981-7_6
Jain K. (2010). Data clustering: 50 years beyond K-means. Pattern Recogn. Lett. 31, 8 (June 2010), 651-
666. https://doi.org/10.1016/j.patrec.2009.09.011
Jain K., Karthik Nandakumar, and Abhishek Nagar. (2008). Biometric Template Security. EURASIP Journal on
Advances in Signal Processing Volume 2008, Article ID 579416, 17 pages.
https://doi.org/10.1155/2008/579416
Jain K., Murty, M & J. Flynn, Patrick. (1999). Data clustering: a review. ACM Comput Surv. ACM Comput.
Surv.. 31. 264-323. https://doi.org/10.1145/331499.331504
Janowczyk, A. , Basavanhally, A. , Madabhushi, A. (2017). Stain normalization using sparse autoencoders
(STANOSA): application to digital pathology. Comput. Med. Imaging Graph 57, 50–61.
https://doi.org/10.1016/j.compmedimag.2016.05.003
Jaumard-Hakoun, A., Xu, K., Roussel-Ragot, P., Dreyfus, G., Denby, B. (2016). Tongue contour extraction
from ultrasound images based on deep neural network. arxiv: 1605.05912 .
Junbo Zhao, Michael Mathieu and Yann LeCun, (2017). Energy-Based Generative Adversarial Networks. ICLR
2017, arXiv:1609.03126v4
Kallenberg, M., Petersen, K., Nielsen, M., Ng, A., Diao, P., Igel, C., Vachon, C., Hol- land, K., Karssemeijer,
N., Lillholm, M., 2016. Unsupervised deep learning ap- plied to breast density segmentation and
mammographic risk scoring. IEEE Trans. Med. Imaging 35, 1322–1331.
https://doi.org/10.1109/TMI.2016.2532122
Karam, M., Brault, I., Van Durme, T., & Macq, J. (2017). Comparing interprofessional and interorganizational
collaboration in healthcare: A systematic review of the qualitative research. International journal of
nursing studies.
Karim Armanious, Chenming Yang, Marc Fischer, Thomas K¨ustner, Konstantin Nikolaou, Sergios Gatidis, and
Bin Yang. MedGAN: Medical Image Translation using GANs. Journal Of Latex Class Files, Vol. 14, No.
8, August 2015.
Kingma and Max Welling. 2013. Auto-encoding variational bayes. CoRRabs/1312.6114 (2013). Retrieved from
http://arxiv.org/abs/1312.6114

26
Li, F., Qiao, H., Zhang, B., Xi, X. (2017). Discriminatively boosted image clustering with fully convolutional
auto-encoders. arXiv preprint arXiv:1703.07980.
Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A., Ciompi, F., Ghafoorian, M., ... & Sánchez, C. I. (2017). A
survey on deep learning in medical image analysis. Medical image analysis, 42, 60-88..
https://doi.org/10.1016/j.media.2017.07.005
Liu S, Liu S, Cai W, et al. Early diagnosis of Alzheimer’s disease with deep learning. In: International
Symposium on Biomedical Imaging, Beijing, China 2014, 1015–18.
Makhzani, A. & Frey, B. (2013). k-Sparse Autoencoders. arxiv: preprint: 1312.5663.
Mansoor, A., Cerrolaza, J., Idrees, R., Biggs, E., Alsharid, M., Avery, R., Linguraru, M.G., 2016. Deep learning
guided partitioned shape model for anterior visual path- way segmentation. IEEE Trans. Med. Imaging
35 (8), 1856–1865. https://doi.org/10.1109/TMI.2016.2535222
Mathews, S. M., Kambhamettu, C., & Barner, K. E. (2018). A novel application of deep learning for single-lead
ECG classification. Computers in biology and medicine, 99:53-62.
https://doi.org/10.1016/j.compbiomed.2018.05.013
Minh H.Q., Niyogi P., Yao Y. (2006). Mercer’s heorem, Feature Maps, and Smoothing. In: Lugosi G., Simon
H.U. (eds) Learning Theory. COLT 2006. Lecture Notes in Computer Science, vol 4005. Springer,
Berlin, Heidelberg. https://doi.org/10.1007/11776420_14
Miotto R, Li L, Kidd BA, et al. (2016). Deep patient: an unsupervised representation to predict the future of
patients from the electronic health records. Scientific Reports, 6:26094.
https://doi.org/10.1038/srep26094
Nahid, A.-A., Mikaelian, A., Kong, Y. (2018). Histopathological breast-image classification with restricted
Boltzmann machine along with backpropagation. Biomedical Research, 29(10), 2068-2077.
https://doi.org/10.4066/biomedicalresearch.29-17-3903
Ng, A. (2013). Sparse autoencoder lecture notes. Source: web.stanford.edu/class/cs294a/sparseAutoencoder.pdf
Ngo, T.A., Lu, Z., Carneiro, G. (2017). Combining deep learning and level set for the au- tomated segmentation
of the left ventricle of the heart from cardiac cine mag- netic resonance. Med. Image Anal. 35, 159–171.
https://doi.org/10.1016/j.media.2016.05.009
Ortiz, A., Munilla, J., Górriz, J.M., Ramírez, J. (2016). Ensembles of deep learning architectures for the early
diagnosis of the Alzheimer’s disease. International Journal of Neural Systems, 26(7), 1650025.
https://doi.org/10.1142/S0129065716500258
Partaourides, Harris; Chatzis, Sotirios P. (2017). Asymmetric deep generative models. Neurocomputing, 241,
90. https://doi.org/10.1016/j.neucom.2017.02.028
Payan, A., Montana, G. (2015). Predicting Alzheimer’s disease: a neuroimaging study with 3D convolutional
neural networks. arXiv preprint arXiv:1502.02506.
Pereira, S., Meier, R., McKinley, R., Wiest, R., Alves, V., Silva, C. A., & Reyes, M. (2018). Enhancing
interpretability of automatically extracted machine learning features: application to a RBM-Random
Forest system on brain lesion segmentation. Medical image analysis, 44, 228-244.
https://doi.org/10.1016/j.media.2017.12.009
Pinaya, W.H.L., Gadelha, A., Doyle, O.M., Noto, C., Zugman, A., Cordeiro, Q., Jack- owski, A.P., Bressan,
R.A., Sato, J.R., 2016. Using deep belief network modelling to characterize differences in brain
morphometry in schizophrenia. Nat. Sci. Rep. 6, 38897. https://doi.org/10.1038/srep38897
Plis, S.M., Hjelm, D.R., Salakhutdinov, R., Allen, E.A., Bockholt, H.J., Long, J.D., John- son, H.J., Paulsen,
J.S., Turner, J.A., Calhoun, V.D., 2014. Deep learning for neu- roimaging: a validation study. Front.
Neurosci. https://doi.org/10.3389/fnins.2014.00229
Rifai, Pascal Vincent, Xavier Muller, Xavier Glorot, and YoshuaBengio. 2011. Contractive auto-encoders:
explicit invariance during feature extraction. In Proceedings of the 28th International Conference on

27
 International Conference on Machine Learning (ICML'11), LiseGetoor and Tobias Scheffer (Eds.).
Omnipress, USA, 833-840.
Salakhutdinov R, and Geoffrey Hinton. 2009. Deep Boltzmann machines. In Artificial Intelligence and
Statistics.PMLR, 448–455.
Salakhutdinov R, and Geoffrey Hinton. 2012. An efficient learning procedure for deep Boltzmann machines.
Neural Computation 24, 8 (2012), 1967–2006.
Salakhutdinov R. 2015. Learning deep generative models. Annu. Rev. Stat. Appl. 2:361–85
Shi, J. Wu, Y. Li, Q. Zhang and S. Ying, "Histopathological Image Classification WithColor Pattern Random
Binary Hashing-Based PCANet and Matrix-Form Classifier," in IEEE Journal of Biomedical and Health
Informatics, vol. 21, no. 5, pp. 1327-1337, Sept. 2017. https://doi.org/10.1109/JBHI.2016.2602823
Shin, H. C., Orton, M. R., Collins, D. J., Doran, S. J., & Leach, M. O. (2013). Stacked autoencoders for
unsupervised feature learning and multiple organ detection in a pilot study using 4D patient data. IEEE
transactions on pattern analysis and machine intelligence, 35(8), 1930-1943.
https://doi.org/10.1109/TPAMI.2012.277
Su H., Xing F., Kong X., Xie Y., Zhang S., Yang L. (2018). Robust Cell Detection and Segmentation in
Histopathological Images Using Sparse Reconstruction and Stacked Denoising Autoencoders. Lecture
Notes in Computer Science, 9351. Springer, Cham. https://doi.org/10.1007/978-3-319-24574-4_46
Suk, H. I., Lee, S. W., Shen, D., (2013a). Latent feature representation with stacked auto-encoder for AD/MCI
diagnosis. Brain structure & function, 220(2), 841-59. https://doi.org/10.1007/s00429-013-0687-3
Suk, H.-I., Lee, S.-W., Shen, D. (2014). Hierarchical feature representation and multi- modal fusion with deep
learning for AD/MCI diagnosis. Neuroimage 101, 569–582.
https://doi.org/10.1016/j.neuroimage.2014.06.077
Suk, H.-I., Shen, D. (2013). Deep learning-based feature representation for AD/MCI classification. In:
Proceedings of the Medical Image Computing and Computer-Assisted Intervention. In: Lecture Notes in
Computer Science, 8150 (pp. 583–590). https://doi.org/10.1007/978-3-642-40763-5_72
Suk, H.-I., Wee, C.-Y., Lee, S.-W., Shen, D. (2016). State-space model with deep learn- ing for functional
dynamics estimation in resting-state FMRI. Neuroimage, 129, 292–307.
https://doi.org/10.1016/j.neuroimage.2016.01.005
Van Tulder, G., & de Bruijne, M. (2016). Combining generative and discriminative representation learning for
lung CT analysis with convolutional restricted boltzmann machines. IEEE transactions on medical
imaging, 35(5), 1262-1272. https://doi.org/10.1109/TMI.2016.2526687
Vincent P, Larochelle H, Lajoie I, (2010) Stacked denoising autoencoders: learning useful representations in a
deep network with a local denoising criterion. Journal of Machine Learning Research, 11, 3371-3408
Vincent, H. Larochelle, Y. Bengio, and P.A. Manzagol. Extracting and composing robust features with
denoising autoencoders. In W.W. Cohen, A. McCallum, and S.T. Roweis, editors, Proceedings of the
Twenty-fifth International Conference on Machine Learning ICML’08 , pages 1096–1103. ACM, 2008.
Wani, N., & Raza, K. (2018). Multiple Kernel-Learning Approach for Medical Image Analysis. In Soft
Computing Based Medical Image Analysis (pp. 31-47). https://doi.org/10.1016/B978-0-12-813087-
2.00002-6
Wu, J., Ruan, S., Mazur, T.R., Daniel, N., Lashmett, H., Ochoa, L., Zoberi, I., Lian, C., Gach, H.M., Mutic, S.,
Thomas, M., Anastasio, M.A., Li, H. (2018). Heart motion tracking on cine MRI based on a deep
Boltzmann machine-driven level set method. In Proceedings of International Symposium on Biomedical
Imaging (pp. 1153-1156). https://doi.org/10.1109/ISBI.2018.8363775
Xu, J., Xiang, L., Liu, Q., Gilmore, H., Wu, J., Tang, J., Madabhushi, A. (2016). Stacked sparse autoencoder
(SSAE) for nuclei detection on breast cancer histopathology images. IEEE Trans. Med. Imaging 35, 119–
130. https://doi.org/10.1109/TMI.2015.2458702

28
Yi, W., Tsang, K. K., Lam, S. K., Bai, X., Crowell, J. A., & Flores, E. A. (2018). Biological plausibility and
stochasticity in scalable VO2 active memristor neurons. Nature Communications, 9(1), 4661.
https://doi.org/10.1038/s41467-018-07052-w
Yoo, Y., Brosch, T., Traboulsee, A., Li, D. K., & Tam, R. (2014). Deep learning of image features from
unlabeled data for multiple sclerosis lesion segmentation. In International Workshop on Machine
Learning in Medical Imaging (pp. 117-124). Springer, Cham. https://doi.org/10.1007/978-3-319-10581-
9_15
Zabalza, J., Ren, J., Zheng, J., Zhao, H., Qing, C., Yang, Z., ... & Marshall, S. (2016). Novel segmented stacked
autoencoder for effective dimensionality reduction and feature extraction in hyperspectral
imaging. Neurocomputing, 185, 1-10. https://doi.org/10.1016/j.neucom.2015.11.044
Zhang, Q., Xiao, Y., Dai, W., Suo, J., Wang, C., Shi, J., Zheng, H., (2016a). Deep learning based classification
of breast tumors with shear-wave elastography. Ultrasonics 72, 150–157.
https://doi.org/10.1016/j.ultras.2016.08.004
Zhao, Wei & Jia, Zuchen & Wei, Xiaosong & Wang, Hai. (2018). An FPGA Implementation of a Convolutional
Auto-Encoder. Applied Sciences. 8. 504. https://doi.org/10.3390/app8040504
Zhu, Y., Wang, L., Liu, M., Qian, C., Yousuf, A., Oto, A., Shen, D. (2017). MRI Based prostate cancer
detection with high-level representation and hierarchical classification. Med. Phys. 44 (3), 1028–1039.
https://doi.org/10.1002/mp.12116

29