PMP

Progress in Medical Physics 30(2), June 2019

Review Article https://doi.org/10.14316/pmp.2019.30.2.49
eISSN 2508-4453

Medical Image Analysis Using Artificial Intelligence

Hyun Jin Yoon1,2, Young Jin Jeong1,2, Hyun Kang2, Ji Eun Jeong1, Do-Young Kang1,2
1
Department of Nuclear Medicine, Dong-A University Medical Center, Dong-A University College of Medicine, 2Institute of Convergence
Bio-Health, Dong-A University, Busan, Korea

Received 2 May 2019 Purpose: Automated analytical systems have begun to emerge as a database system that enables
Revised 21 May 2019 the scanning of medical images to be performed on computers and the construction of big data.
Accepted 21 May 2019 Deep-learning artificial intelligence (AI) architectures have been developed and applied to medical
images, making high-precision diagnosis possible.
Corresponding author Materials and Methods: For diagnosis, the medical images need to be labeled and standardized.
Do-Young Kang After pre-processing the data and entering them into the deep-learning architecture, the final
([email protected]) diagnosis results can be obtained quickly and accurately. To solve the problem of overfitting
Tel: 82-51-240-5630 because of an insufficient amount of labeled data, data augmentation is performed through
Fax: 82-51-242-7237 rotation, using left and right flips to artificially increase the amount of data. Because various deep-
learning architectures have been developed and publicized over the past few years, the results of
the diagnosis can be obtained by entering a medical image.
Results: Classification and regression are performed by a supervised machine-learning method
and clustering and generation are performed by an unsupervised machine-learning method. When
the convolutional neural network (CNN) method is applied to the deep-learning layer, feature
extraction can be used to classify diseases very efficiently and thus to diagnose various diseases.
Conclusions: AI, using a deep-learning architecture, has expertise in medical image analysis of the
nerves, retina, lungs, digital pathology, breast, heart, abdomen, and musculo-skeletal system.

Keywords: Artificial Intelligence (AI), Medical images, Deep-learning, Machine-learning,

Convolutional Neural Network (CNN), Big data

Introduction cient model for image analysis is the ‘Convolutional Neural

Network’ (CNN), a key block in the configuration of deep
1. Recent trends in AI research networks. CNN involves developing various optimization
algorithms such as LeNet, AlexNet, ZF Net, GoogleNet, VG-
It is well-established that automated analytical systems GNet and ResNet. CNN is a powerful feature extractor with
have begun to emerge as a database system capable of a deep layer that can extract features from images.1,2) Deep
scanning medical images with computers and configur- learning algorithms, especially convolutional networks, are
ing big data . With the development of back-propagation rapidly emerging as a methodology for analyzing medical
deep-learning artificial intelligence (AI) architectures, images. There has been remarkable growth in medical im-
medical imaging has begun to be applied and it has been aging analysis of the nerves, retina, lung, digital pathology,
reported that accurate diagnosis is possible. The most effi- breast, heart, abdomen, and musculo-skeletal system us-

Copyright © 2019 Korean Society of Medical Physics

CCThis is an Open-Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-
nc/4.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
50 Hyun Jin Yoon, et al：Medical Image Analysis Using AI

ing the AI methods that make up the deep-learning archi- approach for detecting cartilage lesions in the knee joint
tecture. with MR images.9) PET images with low spatial resolution
overestimated the volume due to the partial volume effect,
2. The need for AI in medical imaging but the optimal volume was extracted using the Artificial
neural networks (ANN) algorithm.10) Through the develop-
Finding non-invasive and quantitative assessment tech- ment of AI algorithms, the diagnostic performance of vari-
niques for early detection of Alzheimer’s disease [AD] ous medical images has been improved, and it is expected
is fundamentally important for early treatment. Tumor to be continuously introduced into medical image diagno-
detection, classification, and quantitative assessment in sis systems because it has a higher diagnostic performance
positron emission tomography (PET) imaging are impor- index than any other quantitative analysis methods so far.
tant for early diagnosis and treatment planning. A number
of techniques have been proposed for segmenting medi- Medical Image Diagnosis Using AI
cal image data through quantitative assessment. However,
some quantitative methods of evaluating medical images For the diagnosis of medical imaging, there should be
are inaccurate and require considerable computation time labeled and standardized big data, as shown in Fig. 1. AI
to analyze large amounts of data. Analytical methods using is generally a concept that includes machine-learning
AI algorithms can improve diagnostic accuracy and save and deep-learning, and can be analyzed in various ways,
time. depending on the characteristics of images stored in the
database. The data is pre-processed in the same order as
3. AI research field in Fig. 2, entered into the deep-learning architecture, and
final diagnosis results are obtained. The hyper-parameter
The deep-learning technology, known as the AI method, is changed and the deep-learning result is checked and
dramatically improves diagnostic performance by au- modified to optimize the parameter. In order to solve the
tomatically extracting features of complex and precise problem of over-fitting due to lack of the sufficient labeled
3,4)
medical images and comparing their differences. Image data, data augmentation is performed through rotation, left
analysis by AI algorithm is superior to the traditional im- and right flip, and GAN.
age analysis method. Automatically classifying skin lesions
using images is a challenging task because the skin lesion 1. Data sets
shapes differ a lot, but they have successfully categorized
skin cancer successfully.5) When evaluating diabetic retinal Although medical images are stored in the PACS system,
fundus photographs, deep-learning using Inception-v3 ar- they are almost useless data for AI analysis. In order to use
chitecture enabled diagnosis of diabetic retinopathy with medical images for AI, labeling, standardization, bounding
6)
high sensitivity and specificity. Early detection of pulmo- box, and segmentation are required. These high quality im-
nary nodules on a chest by computed tomography (CT) ages require manual work, which can be time-consuming
scan was performed using the Lung Image Database Con- and can vary, depending on the skill of the person. Stan-
sortium (LIDC) and the Image Database Resource Initia- dardized and anonymized well-labeled databases are very
tive (IDID) database for the diagnosis of lung cancer, and important for developing and testing AI algorithms that re-
extracted successfully using the network back propagation quire big data. There has been an attempt to build a large-
7)
algorithm. CNN (Convolutional Neural Network)-type scale database system through these complex processes,
deep learning model, trained for large mammography le- and to improve diagnostic performance through competi-
8)
sions, is superior in performance to existing CAD system. tion. These attempts have been quite successful and ap-
An automated detection system has been successfully pear to be most effective in increasing diagnostic accuracy.
established by studying the feasibility of a deep learning In recent years, an anonymous medical image database

www.ksmp.or.kr
 Progress in Medical Physics Vol. 30, No. 2, June 2019 51

Data mining, big data, big data analysis, anonymize, labeling DB system

A technique which enables machines to mimic human behavior

Supervised: classification, regression

Unsupervised: clustering, generation

CNN, RNN - AlexNet, VGG, GoogLeNet, ResNet

Diagnosis

Deep learning (DL)

Machine learning (ML)

Artificial intelligence (AI)

Data Fig. 1. Medical image diagnosis me­

thod using artificial intelligence (AI).

dation did not show a significant difference from the analy-

Data
Modeling sis of 9 clinical experts at the Stanford University Medical
Labeling
Standardzation Deep learning
Training Center.12) The leader-board reports the average AUC (0.917)
and
Bounding box architectures of current (Jan 09, 2019) abnormal detection, ACL tears,
Verification Diagnosis
Segmentation
CNN and meniscal tear work (https://stanfordmlgroup.github.
Hyper- Reporting
Augmentation RNN
Patch RBM
parameter system io/competitions/mrnet).
Optimization
extraction Batch To screen for severe pneumothorax, a total of 13,292
Over fitting
normalization
Under fitting
Anonymization Activation frontal chest X-rays were visually labeled by a radiologist
Database Pooling
system to create a database.13) Candidate images for analysis were
identified and stored in a clinical PACS system in the form
Fig. 2. Flow chart of medical imaging AI diagnosis system development. of ‘dicom’. These well-defined data sets can be used to
train and evaluate different network architectures. Build-
has been released and researchers are freely approaching ing such a well-defined database requires a lot of time and
and presenting research results. A database of F-18 FDG effort. The National Institute of Health (NIH) has opened
PET images from participants of the Alzheimer’s disease the Chest X-ray14 set, so that it is publicly available (avail-
neuro-imaging initiative (ADNI) is being developed to able at https://nihcc.app.box.com/v/ChestXray-NIHCC).
study early detection of AD patients. F-18 FDG PET images The NIH allows students to learn by using well-organized
of the ADNI database were used to distinguish AD patients and labeled data sets, and opens up a database for anyone
with 88.64% specificity, 87.70% sensitivity, and 88.24% ac- to challenge and develop optimized algorithms. Lung Im-
curacy.11) Magnetic resonance imaging (MRI) of the knee age Database Consortium (LIDC) and Image Database
is the preferred method for diagnosing a knee injury, but Resource Initiative (IDRI) databases were created and
the analysis is time consuming and the likelihood of diag- released for the diagnosis of lung cancer. For the diagnosis
nostic errors is high. MRNet, a CNN, and logistic regression of skin cancer, the database can be opened (https://isic-
analysis were performed as a dataset of 1,370 knee MRI archive.com) to participate in the competition.5) The MR
scans performed at the Stanford University Medical Center knee database is also available and anyone interested in
(https://stanfordmlgroup.github.io). The results of the vali- this field can participate in AI analysis (https://stanfordml-

www.ksmp.or.kr
52 Hyun Jin Yoon, et al：Medical Image Analysis Using AI

group.github.io/competitions/mrnet). Most problems with When applying an AI algorithm through a normalized im-
large data sets and public challenge data sets are listed at age, the accuracy is generally increased. However, there is a
http://www.grand-challenge.org. lack of logical explanations and studies on why normalized
In order to construct a more efficient, accurate and images increase accuracy. Theoretical studies are needed,
useful database, the Department of Nuclear Medicine at as well as empirical studies on the optimal pre-processing
Dong-A University Medical Center established a database method for image analysis using AI.
system called SortDB with IRM (http://www.irm.kr/). The The accuracy of the quantitative analysis has a consider-
feature of SortDB is some software that allows researchers able impact on the quality of the image transformation by
to download a desired dicom file in a specified file format pre-processing. In particular, the accuracy of AI analysis
(nii, jpg, gif etc.) by entering a manageable list of Excel varies considerably, depending on the quality of medical
files. To apply an AI algorithm, big data is needed. It is an images for applying the algorithm. The image transfor-
automated database generation program that stores and mation of the dicom file is linearly transformed using the
manages data stored in small and medium sized hospitals return value of getOuputData function of a library.15) Bits
in the necessary format. A database is constructed with the number 8 and the SUV values were linearly converted to
system configuration shown in Fig. 3, shortening the pre- 0-255 scales. By using this method of linear transformation,
processing time for creating a standardized data set. the image representing the SUV value can be reproduced
as a jpg file with distortion. It is necessary to study the im-
2. Pre-processing of medical images age conversion methods so that medical images will suffer
only minimal deformation in the conversion process. It is
For quantitative analysis, an image was normalized using necessary to maintain the consistency of AI learning using
a standard affine model with 12 parameters using SPM5 image-aware CNNs by down-sampling commonly used
14)
software. In order to use quantitative evaluation using images.13)
machine learning, a pre-processing method, including
the normalization process of image data must necessarily 3. Data augmentation, model and parameter
be included. Since the accuracy of the normalized image optimization
depends on the standard template or the variety of the pro-
cess, there are considerable variations among researchers. Generally, PET images do not have a large number of

a b
(3) Analysis integrated with
sandwich Dicom

DB (excel etc.)

Matching list
Grouped data nii etc.

SortDB

Output format order

Various output figure

(1) Nuclear medicine images (2) Sandwich (4) Web-based image display with
image screening results

nii, jpg, gif, png etc.

Fig. 3. Developed SortDB ver1.0 for DICOM, hdr/img, nii image file viewer, delineation, cropping, and file conversion function. The
environment (a) and the functional flowchart (b) of SortDB ver1.0 are displayed. Dicom PET images can be converted to multi-frame nii
files, jpg, gif and save, and the preprocessing process for normalization can be performed quickly and conveniently.

www.ksmp.or.kr
 Progress in Medical Physics Vol. 30, No. 2, June 2019 53

data, making it difficult to construct a database of big data. Machine Learning

In this case, data augmentation is performed through im-
age inversion, enlargement/reduction, shearing, and rota- One of the most successful methods is the PCA method,
tion to randomly distribute data to perform AI analysis. Too which is based on the eigen-image approach.22) After ana-
many hyper-parameter optimizations are some of the most lyzing the images through PCA, SVM was used as a clas-
important difficulties in achieving optimal accuracy. This sifier to distinguish the boundary between normal and
can be done using the open source Future Gadget Labora- demented patients and to improve the accuracy by diag-
tory (https://github.com/Kaixhin/FGLab) framework.13) nosing and analyzing the initial dementia. PCA was applied
Evaluation of the validity of deep-learning can be done by to SPECT images and the possibility of classification was
calculation of the area under the receiver operating char- greatly improved, and the results were presented through
acteristic curve (AUC), sensitivity, specificity, and positive 3-D scattering.23) Quantitative evaluation and classifica-
predictive value (PPV) of the entire validation set, and can tion is one of the most efficient and logical methods. The
be done using scikit-learn (https://scikit-learn.org). Re- accuracy of classification of AD by applying PCA-SVM to
ceiver operating characteristic (ROC) curves can be drawn SPECT and PET images was 96.7% and 89.52%, respective-
using matplotlib (https://matplotlib.org). ly.24) Using Gaussian mixture models (GMMs), a method
Currently, CNN algorithms are implemented in vari- of automatically selecting the region of interest (ROI) of a
ous ways. The recent trend is to use Keras (version 2.0.3, three-dimensional functional brain image containing high
https://keras.io/) deep-learning library on TensorFlow and low activation region information was presented.25)
(version 1.2.1, Google) to implement a model of convo- Based on classification using eigenvector decomposition
luted neural networks. Deep-learning architectures for AI and SVM, including concept extraction by representative
analysis are VGG16/19, Xception, Inception, and ResNet vector separation using principal component analysis, we
16-19)
models, which are the best performing algorithms. The find that (PCA)/independent component analysis (ICA)
optimal algorithm for image classification may be an al- for early AD diagnosis by brain image analysis and pattern
gorithm that can reflect the feature characteristics of each disease discrimination CAD is a very useful direction.11)
medical image. Currently, the location of such features is
estimated through the activation map. However, at present, Application of AI Medical Image Analysis
only the experience of repeated execution of each medi-
cal image is the method of finding the optimal model. The 1. AI in Brain Imaging
AlexNet algorithm, which won the ILSRVC competition
in 201220) to classify the chest X-ray image into 5 classes, In brain research using AI, many studies have been con-
was used and successfully classified with 92.10% accu- ducted in the field of Alzheimer’s disease classification, an-
racy.21) AlexNet consists of 5 convolution layers and 3 fully- atomical segmentation of brain regions, and tumor detec-
connected (FC) layers, and the last FC layer uses a softmax tion. Alzheimer’s disease (AD)/mild cognitive impairment
21)
activation function. The researcher down-sampled the (MCI)/HC classification was successfully performed by
dicom file and converted it into a 256×256 pixel PNG file using the Gaussian Restricted Boltzmann Machine (RBM)
and used it as the input image of AlexNet. The mini-batch to find feature expressions in volume patches of MRI and
size is 128, the number of training iterations is 100, the PET images.26) The 3-D convolution neural network in AD
adaptive learning rate is 1×10−3 for Adam optimizer, and classification is superior to other algorithm classifiers.27,28)
the momentum is 0.5. The ReLU activation functions were It automatically segments the magnetic resonance (MR)
used in the previous step of the max-pooling layers and the images of the human brain using CNN.29) Segmentation of
L2 regularization was adjusted to 1×10−4. striatum segmentation was performed using deep CNN,
and the results were compared with those of FreeSurfer.30)
In the brain area, manual segments are time-consuming,

www.ksmp.or.kr
54 Hyun Jin Yoon, et al：Medical Image Analysis Using AI

with individual differences occurring, while automatic seg- cancer is the detection and classification of tumor lesions,
mentation has significant difficulties in performing in com- the detection and classification of micro-calcifications,
plex structures. Twenty-five deep-layers called the ‘voxel- and risk-scoring work, which can be effectively analyzed
wise residual network’ (VoxResNet) were developed and by CNN or RBM methods. For the measurement of breast
31)
successfully segmented automatically. To demonstrate density, CNNs for feature extraction were used,39) and a
end-to-end nonlinear mapping from MR images to CT im- modified region proposal CNN (R-CNN) has been used for
ages, a 3-D fully convolutional neural network (FCN) was localization. It has been reported that U-net is used for seg-
employed and verified in a real pelvic CT/MRI data set.32) mentation breast and fibro-glandular tissue (FGT) in MRI
Input and output improved performance by using two- in a dataset and accurate breast density calculation results
volume CNNs, and excellent performance was observed by are observed.40) It has been reported that a short-term risk
evaluating the input and output forms in the MRI and PET assessment model has been developed by achieving a pre-
images of the Neuroimaging Initiative (ADNI) database.33) dictive accuracy of 71.4% by calculating the risk score by
implementing the mammographic X-ray image as a risk
2. AI in Chest Imaging prediction module MLP (multiple layer perception) classi-
fier.41)
By introducing the multiple-instance learning (MIL)
framework, a de-convolutional neural network is con- 4. AI in Cardiac Imaging
34)
structed to generate the heat map of suspicious regions. A
unique set of radiologic datasets of publicly available chest Cardiac artificial research fields include left ventricle
X-rays and their reports were used to find and report 17 segmentation, slice classification, image quality assess-
unique patterns by applying CNN algorithms.35) It has been ment, automated calcium scoring and coronary centerline
reported that the presence of interstitial patterns have been tracking, and super-resolution. 2-D and 3-D CNN tech-
found by applying a segmentation-based label propagation niques are mainly used for classification, and deep-learn-
method to a dataset of interstitial lung disease,36) and it has ing techniques such as U-net segmentation algorithm are
been reported that lung texture patterns are classified using used for segmentation. The high-resolution 3-D volume in
37)
CNN. A method for classifying frontal and lateral chest X- the 2-D image stack has been reconstructed using a novel
ray images using deep-learning methods and automating image super-resolution (SR) approach.42) The image qual-
metadata annotations has been reported.38) A new method ity is superior to the SR method because the CNN model is
of using a three-dimensional (3-D) CNN for false positive computationally efficient, but SR-CNN is advantageous in
reduction in automatic pulmonary nodule detection in a image segmentation and motion tracking.43) Using multi-
Volumetric Computed Tomography (CT) scan has been stream CNN (3 views), it has been reported that low-dose
proposed. 3-D CNN is able to enter more spatial informa- chest CT can be identified with high accuracy by deep
tion and extract more representative features through a learning when the region of interest is considered as a cor-
hierarchical architecture, trained with 3-D samples. The onary artery calcification candidate over 130 HU.44) Coro-
proposed algorithm has achieved high CPM (Competition nary calcium in gated cardiac CT angiography (CCTA) was
Metric) scores, has been extensively tested in the LUNA16 detected using 3-D CNN and multi-stream 2-D CNN.45)
Challenge, and can be applied to 3-D PET images.
5. AI in Musculo-skeletal Imaging
3. AI in Breast Imaging
Musculo-skeletal images are analyzed by deep-learning
Since most mammograms are 2-D and the number of algorithms for segmentation and identification of bone,
data is large, AI images can be successfully analyzed using joint, and associated soft tissue abnormalities. 3-D CNN
deep-learning in natural images. The discovery of breast architecture has been developed to automatically perform

www.ksmp.or.kr
 Progress in Medical Physics Vol. 30, No. 2, June 2019 55

supervised segmentation of vertebral bodies (VBs) from sis. However, more research is needed to see if synthetic
3-D magnetic resonance (MR) spine images, and a ‘Dice’ images can be used for artificial intelligence learning to de-
23)
similarity coefficient of 93.4% has been reached. Auto- termine clinical diagnostics that require rigorous accuracy.
matic spine recognition, including spine location identifi- Increasing data through the GAN is still controversial, but it
cation and multiple image naming, requires large amounts is a field of research that is needed. It is necessary to make
of image data and is difficult to recognize due to the variety the medical image stored in the PACS system into the im-
of spine shapes and postures. By using a deep-learning age of the previous step for AI analysis. Research is needed
architecture called Transformed Deep Convolution Net- to automatically generate standard images so that medical
work (TDCN), the posture of the spine was automatically images can be used directly in deep-learning.
46)
corrected to process the image. A CNN regression has AI analysis of medical images requires labeled, standard-
a limitation in that its computation time is long and the ized and optimized images. There is a clear difference in
capture range is small for implementing an intensity-based the accuracy of the final classification between pre-pro-
2-D/3-D registration technology. It is reported that highly cessing and non-pre-processing medical images. However,
accurate real-time 2-D/3-D registration is possible, even in since the pre-processed image has noise different from that
47)
a greatly enlarged capture range. Several deep-learning of the original image, it is necessary to study the effect of
methods have been developed for automatically evaluat- the resulting image on the accuracy. Deep-learning archi-
ing the age of skeletal bones using X-ray imaging and their tecture can have various forms, and new excellent architec-
performance has been verified by showing an average dis- ture for image analysis is being released every year through
48)
crepancy of about 0.8 years. competition. However, it is only through experience that
the most accurate result of any architecture can be used for
Discussion each medical image. Therefore, it is necessary for people
to apply various deep-learning architectures to each medi-
Diagnosis using AI is quickly performed and its accuracy cal image and then to discuss each other’s experiences to
is very high. AI diagnosis is becoming an important tech- find the optimal method. Even if one deep-learning model
nology for future diagnostic systems. However, AI diagnosis is determined, there are many hyper-parameters in it. The
needs to be supplemented in several aspects. AI learning composition of parameter combinations is often depen-
using deep-learning architecture requires big data. Howev- dent on user experience. Optimization of Hyper-parame-
er, most medical images are technical and man-powered, ters is usually done through a time-consuming grid search
making it difficult to build big data systems. It is also time- and random search, but Bayesian optimization and genetic
consuming to create a database of standardized and la- algorithm optimization can be used efficiently.
beled images of medical images. Most people are building Medical imaging diagnostics using deep-learning archi-
databases by manually pre-processing all medical images tecture have reached expert levels in the areas of neurons,
for AI application. Performing data augmentation through retina, lung, digital pathology, breast, heart, abdomen, and
rotation, left/right flip, and up/down reversal of a medi- musculo-skeletal system. However, when other hospital
cal image has a positive effect on the accuracy of learning. protocols apply different medical images, their accuracy
Data augmentation using GAN is being applied in various is significantly reduced and new optimization parameters
21,49)
areas of the medical field. In liver lesion classification must be found. Even with some resolution and noise
using CT images, it was reported that the accuracy of 7.1% changes in medical imaging, AI diagnostics are very frag-
was increased when the number of data using GAN was in- ile. There is also a lack of logical explanations as to which
creased.21) In chest pathology classification using X-ray im- process to diagnose. Although there is an effort to maintain
49)
age, accuracy was reported to be increased by 21.23%. It consistency through data harmonization, there is the pos-
has been reported that synthesized images using GAN can sibility that the overall quality of the image will deteriorate.
be a method of data augmentation in medical image analy- It is necessary to study the change of the accuracy by using

www.ksmp.or.kr
56 Hyun Jin Yoon, et al：Medical Image Analysis Using AI

the data whose resolution is decreased. It is necessary to 3. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;
develop algorithms that can recognize and generalize im- 521:436.
ages of medical images with different resolutions or noise, 4. Zeiler MD, Fergus R. Visualizing and understanding con-
but it will take a considerable amount of time. Neverthe- volutional networks. 2014;818-833.
less, the diagnosis of medical images by the deep-learning 5. Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level
architecture developed up till now is almost at expert diag- classification of skin cancer with deep neural networks.
nosis level. In addition, many data that have not yet been Nature. 2017;542:115.
analyzed can be discovered and studied with AI diagnosis, 6. Gulshan V, Peng L, Coram M, et al.Development and vali-
and this can accurately and quickly diagnose and improve dation of a deep learning algorithm for detection of dia-
the quality of medical care dramatically. betic retinopathy in retinal fundus photographs. JAMA.
2016;316:2402-2410.
Acknowledgements 7. Golan R, Jacob C, Denzinger J. Lung nodule detection in
CT images using deep convolutional neural networks.
This research was supported by the project at Institute 2016;243-250.
of Convergence Bio-Health, Dong-A University funded 8. Kooi T, Litjens G, Van Ginneken B, et al. Large scale deep
by Busan Institute of S&T Evaluation and Planning. The learning for computer aided detection of mammographic
authors would like to thank Dr. Adrian Ankiewicz of the lesions. Med Image Anal. 2017;35:303-312.
Australian National University (Australia) for helpful com- 9. Liu F, Zhou Z, Samsonov A, et al. Deep learning approach
ments on the manuscript. for evaluating knee MR images: Achieving high diagnos-
tic performance for cartilage lesion detection. Radiology.
Conflicts of Interest 2018;289:160-169.
10. Sharif MS, Abbod M, Amira A, Zaidi H. Artificial neural
The authors have nothing to disclose. network-based system for PET volume segmentation.
Journal of Biomedical Imaging. 2010:4.
Availability of Data and Materials 11. Illán I, Górriz JM, Ramírez J, et al. 18F-FDG PET imaging
analysis for computer aided Alzheimer’s diagnosis. Inf Sci.
All relevant data are within the paper and its Supporting 2011;181:903-916.
Information files. 12. Bien N, Rajpurkar P, Ball RL, et al. Deep-learning-assisted
diagnosis for knee magnetic resonance imaging: Develop-
Ethics Approval and Consent to Participate ment and retrospective validation of MRNet. PLoS Medi-
cine. 2018;15:e1002699.
The study was approved by the institutional review board 13. Taylor AG, Mielke C, Mongan J. Automated detection of
(IRB approval number; DAUHIRB-17-108). moderate and large pneumothorax on frontal chest X-rays
using deep convolutional neural networks: A retrospective
References study. PLoS Medicine. 2018;15:e1002697.
14. Woods RP, Grafton ST, Holmes CJ, Cherry SR, Mazziotta
1. Farooq A, Anwar S, Awais M, Alnowami M. Artificial intel- JC. Automated image registration: I. general methods and
ligence based smart diagnosis of alzheimer’s disease and intrasubject, intramodality validation. J Comput Assist
mild cognitive impairment. 2017;1-4. Tomogr. 1998;22:139-152.
2. Vieira S, Pinaya WHL, Mechelli A. Using deep learning to 15. DCMTK V364. Https://support.dcmtk.org/docs/classDi-
investigate the neuroimaging correlates of psychiatric and comImage.html#ac1b5118cbae9e797aa55940fcd60258e.
neurological disorders: Methods and applications. Neuro- 2019.
science & Biobehavioral Reviews. 2017;74:58-75. 16. Simonyan K, Zisserman A. Very deep convolutional net-

www.ksmp.or.kr
 Progress in Medical Physics Vol. 30, No. 2, June 2019 57

works for large-scale image recognition. ArXiv Preprint Methods. 2016;274:146-153.

arXiv:1409.1556 (2014). 31. Chen H, Dou Q, Yu L, Qin J, Heng P. VoxResNet: Deep vox-
17. Szegedy C, Liu W, Jia Y, et al. Going deeper with convolu- elwise residual networks for brain segmentation from 3D
tions. 2015;1-9. MR images. Neuroimage. 2018;170:446-455.
18. He K, Zhang X, Ren S, Sun J. Deep residual learning for im- 32. Nie D, Cao X, Gao Y, Wang L, Shen D. Estimating CT im-
age recognition. 2016;770-778. age from MRI data using 3D fully convolutional networks.
19. Chollet F. Xception: Deep learning with depthwise sepa- Anonymous : Springer, (2016), pp. 170-178.
rable convolutions. 2017;1251-1258. 33. Li R, Zhang W, Suk H, et al. Deep learning based imaging
20. Krizhevsky A, Sutskever I, Hinton GE. Imagenet classifica- data completion for improved brain disease diagnosis.
tion with deep convolutional neural networks. 2012;1097- 2014;305-312.
1105. 34. Kim H, Hwang S. Deconvolutional feature stacking for
21. Salehinejad H, Valaee S, Dowdell T, Colak E, Barfett J. weakly-supervised semantic segmentation. ArXiv Pre-
Generalization of deep neural networks for chest pathol- print arXiv:1602.04984 (2016).
ogy classification in x-rays using generative adversarial 35. Shin H, Roberts K, Lu L, Demner-Fushman D, Yao J, Sum-
networks. 2018;990-994. mers RM. Learning to read chest x-rays: Recurrent neural
22. Turk M, Pentland A. Eigenfaces for recognition. J Cogn cascade model for automated image annotation. 2016;
Neurosci. 1991;3:71-86. 2497-2506.
23. Korez R, Likar B, Pernuš F, Vrtovec T. Model-based seg- 36. Gao M, Xu Z, Lu L, et al. Segmentation label propagation
mentation of vertebral bodies from MR images with 3D using deep convolutional neural networks and dense con-
CNNs. 2016;433-441. ditional random field. 2016;1265-1268.
24. López M, Ramírez J, Górriz JM, et al. Principal component 37. Gao M, Bagci U, Lu L, et al. Holistic classification of CT at-
analysis-based techniques and supervised classification tenuation patterns for interstitial lung diseases via deep
schemes for the early detection of alzheimer’s disease. convolutional neural networks. Computer Methods in
Neurocomputing. 2011;74:1260-1271. Biomechanics and Biomedical Engineering: Imaging &
25. Górriz J, Lassl A, Ramírez J, Salas-Gonzalez D, Puntonet C, Visualization. 2018;6:1-6.
Lang E. Automatic selection of ROIs in functional imaging 38. Rajkomar A, Lingam S, Taylor AG, Blum M, Mongan J.
using gaussian mixture models. Neurosci Lett. 2009;460: High-throughput classification of radiographs using deep
108-111. convolutional neural networks. J Digital Imaging. 2017;30:
26. Suk H, Lee S, Shen D, Alzheimer’s Disease Neuroimaging 95-101.
Initiative. Hierarchical feature representation and multi- 39. Fonseca P, Mendoza J, Wainer J, et al. Automatic breast
modal fusion with deep learning for AD/MCI diagnosis. density classification using a convolutional neural net-
Neuroimage. 2014;101:569-582. work architecture search procedure. 2015;9414:941428.
27. Payan A, Montana G. Predicting alzheimer’s disease: A 40. Dalmış MU, Litjens G, Holland K, et al. Using deep learn-
neuroimaging study with 3D convolutional neural net- ing to segment breast and fibroglandular tissue in MRI
works. ArXiv Preprint arXiv:1502.02506 (2015). volumes. Med Phys. 2017;44:533-546.
28. Hosseini-Asl E, Gimel’farb G, El-Baz A. Alzheimer’s dis- 41. Qiu Y, Wang Y, Yan S, et al. An initial investigation on de-
ease diagnostics by a deeply supervised adaptable 3D veloping a new method to predict short-term breast cancer
convolutional network. ArXiv Preprint arXiv:1607.00556 risk based on deep learning technology. 2016;9785:978521.
(2016). 42. Avendi M, Kheradvar A, Jafarkhani H. A combined deep-
29. de Brebisson A, Montana G. Deep neural networks for learning and deformable-model approach to fully auto-
anatomical brain segmentation. 2015;20-28. matic segmentation of the left ventricle in cardiac MRI.
30. Choi H, Jin KH. Fast and robust segmentation of the stria- Med Image Anal. 2016;30:108-119.
tum using deep convolutional neural networks. J Neurosci 43. Oktay O, Bai W, Lee M, et al. Multi-input cardiac image

www.ksmp.or.kr
58 Hyun Jin Yoon, et al：Medical Image Analysis Using AI

super-resolution using convolutional neural networks. deep convolution network. Comput Med Imaging Graph-
2016;246-254. ics. 2016;51:11-19
44. Lessmann N, Išgum I, Setio AA, et al. Deep convolutional 47. Miao S, Wang ZJ, Liao R. A CNN regression approach for
neural networks for automatic coronary calcium scoring real-time 2D/3D registration. IEEE Trans Med Imaging.
in a screening study with low-dose chest CT. 2016;9785: 2016;35:1352-1363.
978511. 48. Spampinato C, Palazzo S, Giordano D, Aldinucci M, Leon-
45. Wolterink JM, Leiner T, de Vos BD, van Hamersvelt RW, Vi- ardi R. Deep learning for automated skeletal bone age as-
ergever MA, Išgum I. Automatic coronary artery calcium sessment in X-ray images. Med Image Anal. 2017;36:41-51.
scoring in cardiac CT angiography using paired convolu- 49. Frid-Adar M, Diamant I, Klang E, Amitai M, Goldberger J,
tional neural networks. Med Image Anal. 2016;34:123-136. Greenspan H. GAN-based synthetic medical image aug-
46. Cai Y, Landis M, Laidley DT, Kornecki A, Lum A, Li S. mentation for increased CNN performance in liver lesion
Multi-modal vertebrae recognition using transformed classification. Neurocomputing. 2018;321:321-331.

www.ksmp.or.kr