Professional Documents
Culture Documents
Advances in Deep Learning Techniques For Medical Image Analysis
Advances in Deep Learning Techniques For Medical Image Analysis
Abstract—Deep learning is contributing to the high level text summarization, machine translation, game playing etc.
of services to the healthcare sector. As the digital medical Many companies like Google’s Deep Mind, Apple, IBM, Mi-
data is increasing exponentially with time, early detection and crosoft, NVIDIA, etc. are using the deep learning techniques
prediction of diseases are becoming more efficient because of
the deep learning techniques which reduce the fatality rate to for building up new technologies. Horus Technology with its
a great extent. The main focus of this paper is to provide the fascinating innovation helps the blind people to see. They are
comprehensive review of deep learning in the domain of medical developing a wearable device that uses computer vision, deep
image processing and analysis. We have demonstrated the use of learning, and GPUs to understand the surrounding environ-
new deep learning architectures in oncology for the prediction ment and describe it to users [1] so that they get the insight
of different types of cancer like the brain, lung, skin, etc. The
state-of-the-art architectures effectively carry out analysis of 2D of the things around them. Along with other applications,
and 3D medical images to make the diagnosis of patients faster deep learning is deployed at the front lines of healthcare
and more accurate. The use of popular approaches in machine and has produced the influential results by analyzing huge
learning such as ensemble and transfer learning with fine-tuning electronic medical data for the treatment of the diseases.
of parameters improve the performance of the deep neural net- Researchers showed that they could predict heart failure nine
works in medical image analysis. The existing deep networks urge
the new image classification network called Capsule Network months before the traditional techniques [1]. The future of
(CapsNet) to make the classification and detection comparatively personalized medicine is expected to be accomplished by
better. The equivariance characteristics of CapsNet make it more deep learning techniques. The researchers in Imperial College
influential as it discourages the effect of any structural invariance London are working to automatically provide the assessment
of an input image on the network. of brain damage very fast [43]. Enlitic, the San Francisco-
Keywords—cancer, convolutional neural network, 3D CNN,
capsule network, transfer learning, ensemble learning, deep
based startup is using deep learning to build solutions and
learning the state-of-the-art clinical decision support products [44].
Arterys is the AI assistance to the radiologist and provides
quantification and visualization of the heart flow in the body
I. I NTRODUCTION
using MRI machine [1], [45]. Researchers at the University
The automated imitation of the brain is an emerging technol- of Toronto is working on the cancer-causing mutations using
ogy and has driven large focus and attention of the researchers deep learning [46]. Deep Genomics are using deep learning to
and big research organizations towards itself. Deep Learning understand the variations of the genes that cause diseases [1].
is the subfield of machine learning which uses techniques Deep learning techniques are used for analyzing X-Rays,
inspired by the learning ability of the human brain. The deep CT Scans and MRIs images which has improved detection,
neural network is the neural network with many layers and diagnosis, and treatment of disease. Cancer is a deadly disease
the architecture of these deep nets is a little complicated but and the number of patients suffering from cancer is increasing
is computationally stronger than any other machine learning rapidly. Indian Council of Medical Research (ICMR) stated in
methods like Linear Regression, Logistic Regression, Random 2016 that the total number of new cancer cases is expected
forest, k-nearest neighbors, Support Vector Machine (SVM), to reach nearly 17.3 lakh in 2020 [2]. The early detection of
etc. Deep neural networks have the hierarchical architecture such deadly disease can reduce the fatality rate and the deep
where each layer categorizes some information, makes modi- learning methods have proved to be beneficial in the early
fications and finally passes this information to the next layer. detection of diseases. The deep nets in medical diagnosis are
These networks are trained on many levels of concepts, ideas more efficient than the previous image processing techniques.
and abstractions ranging from simple to complex ones. Deep One of the state-of-the-art architecture used in deep learning
learning has set its flag in many fields like Natural Language image processing is a Convolution Neural Network (CNN or
Processing, Computer Vision, Prediction analysis, etc. It has ConvNet). The CNN is very effective in the areas of image
emerged as the powerful tool and produces phenomenal results recognition and classification problems. This technique of deep
in many applications like image processing, object detection, learning has provided remarkable results in many competitions
Architecture Layer Improvement Top-5 error (Classification) Training time GPU Parameters Winner
AlexNet 8 ReLU was used first time 16.4 6 days 2 60 Million 2012
Replaced filters used in AlexNet
VGG 19 7.3 2-3 weeks 4 138 Million 2014
with multiple 3x3 filters
GoogLeNet 22 Inception module and avg. pooling 6.7 2 weeks 8 25 Million 2014
ResNet 152 Residual modules 3.5 2-3 weeks 8 60 Million 2015
has used the Threshold approach and Marker-Controlled Wa- interest. The most popular CNN architecture, U-net is used
tershed Segmentation and for feature extraction, Binarisation for the segmentation of 2D biomedical images [13]. The
and Masking approaches are used. Similarly, there are many architecture consists of the equal number of the downsampling
other methods that can be used for enhancement, segmentation layers to capture features and upsampling layers for accurate
and feature extraction such as Histogram equalization, Adap- localization. The information of the downsampling layers
tive Histogram Equalization, CLAHE, Histogram of oriented is transferred to the upsampling layers by means of skip
gradients (HOG), Local binary patterns (LBP), etc. Abhishek connections. The U-net takes the entire images in one forward
et al. [8] have presented the Statistical technique for the feature pass and results in the segmentation map. Furthermore, V-
extraction and then classified a brain tumour by analyzing net which is the variant of U-net is used for the volumetric
Brain MRI images using associative rules. The authors have segmentation of the image. It is the fully convolutional neural
taken the handicraft textual features as an input to the system network and performs segmentation of 3D medical images
for the classification of a brain tumour. [14]. The high- level deep learning pipeline is implemented
by NiftyNet [15] for the number of medical image applica-
B. State-of-the-art architecture for classification
tions such as segmentation, image generation, regression, and
AlexNet, VGG, GoogLeNet and ResNet are the popular representation. The main components of the NiftyNet includes
architectures that were introduced for the image classification. data loading, data augmentation, deep network architectures,
Krizhevsky et al. [9] designed the first deep Convolutional loss functions and the evaluation metrics. It enables the rapid
Neural Network called AlexNet which is the 8 layer architec- development of the deep learning solutions for the medical
ture composed of 5 convolutional layers and 3 fully connected image applications. Besides these architectures, DenseNet [16]
layers. Then the University of Oxford presented their archi- and W-net [17] is also implemented to carry out the task
tecture called VGG which is the 19 layer architecture and of segmentation. DenseNet architecture is an extension of
made the improvements over AlexNet where the authors used ResNet. It consists of the dense block where each layer takes
multiple small size kernels to detect more complex features feature maps of all previous layers as input. The concatenation
[10]. To utilize the computing resources Google introduced of feature maps improves the efficiency and encourages reuse
their model called GoogLeNet/inception. It cited the problem of features. For segmentation purpose, upsampling layers are
of computing efficiency and tried to design inception module added to DensNet but only the feature maps of the previous
that is efficient in the amount of computing. The inception layer are upsampled. Like U-net, DenseNet also has skip
modules are stacked on each other forming the 22 layer archi- connections that transfer high-resolution information between
tectures [11]. To overcome the problem of vanishing gradient upsampling and downsampling layers. In the W-net, two U-
and training error that most deep neural networks faced during net architectures are bridged to fully use the feature maps.
backpropagation, Kaiming et al. [12] introduced ResNet. It is The features are fused together using concatenation in the
a 152 layer architecture formed of residual module stacked bridge method and eventually improves the performance of the
over each other. All these architectures were introduced in network. Two activations functions are used in the architecture
ImageNet LSVRC for the classification of 1.2 million images which is Exponential liner unit (ELU) and RELU. In the skip
into 1000 classification and emerged as winners as shown connection method, addition is used instead of concatenation.
in Table I. They have huge memory and high computation The avoidance of concatenation was done to remove redun-
power so the availability of requirements for these models is dancy and reduce the learning burden.
an important concern especially during training. The accuracy
of the models is improved with the addition of layers making D. Automatic extraction of features
the architecture computationally intensive so there is a trade-
off between accuracy and computation. The extraction of distinct features is very difficult and need
to be carefully designed so that the chances of missing out any
C. Segmentation in medical images distinct features is reduced. The deep learning techniques for
Segmentation is an essential step in medical image analysis. medical image analysis eliminates the overhead of manually
It isolates the region of interest and significantly improves selecting features and hence improves the classification and
the performance of image classification. Several variations of the performance of the system. CNN extracts the features
CNN is used for the precise segmentation of the object of of input images with different lighting condition better than
other classification methods [18]. Deep learning method like detection (DCNN). DCNN is trained to detect lung nodules in
Faster Region-based Convolutional Neural Network (Faster subvolumes of CT Scan images and used it to predict the
R-CNN) [19] is used to train networks for the detection of location and boundary of lung nodules in unprocessed CT
cancer in lungs. The network is formed of two modules- Scan images. J. Ma et al. [30] have proposed a hybrid CNN
Region Proposal Network depicts the region of interest and for the thyroid nodule diagnosis. It is the combination of 2
ROI classifier recognizes whether ROIs are nodule or not. pre-trained CNN models, CNN 1 and CNN 2 with different
The deconvolution layer is used in the network to recover convolution and fully connected layer. CNN 1 capture the
more fine-grained features of nodule because of its small ultra-fine low-level features and CNN 2 capture the complex
size. RPN takes feature layer as input and outputs region of features of the thyroid nodule. The feature maps of both the
interest. Q. Song et al. [20] have taken 3 deep learning methods networks are fused and given as input to the softmax classifier
CNN, Deep Neural Network (DNN) and space autoencoder to to detect thyroid nodule. N. Kumar et al. [31] have proposed
classify nodules as malignant and benign on the LIDC-IDRI the deep learning architecture for the detection of prostate
dataset. Among the three networks, CNN was successful in cancer. The authors have used two CNN models, one for the
selecting the most distinct features of lung nodules in CT detection of nuclei and the second CNN model to classify
Scan images for classification of cancer and takes lead with an them. Nima et al. [32] have taken 3 specialities in medical
accuracy of 84.16%. In [40] Hakan et al. have used two deep imaging application (cardiology, radiology, gastroenterology)
learning architectures VGG and Resnet inspired by VGG16 and have done classification, detection, and segmentation
and ResNet18 to detect the oral and cervix cancer. The authors and measured the performance of trained deep CNN from
have shown the variation of results with different dataset scratch and pre-trained CNN which is fine-tuned in a layer-
along with different architecture. The model is trained on wise manner. The better accuracy and performance of the
CerviSCAN and Herlev dataset for cervix cancer detection and deep learning techniques is not limited only to the standard
Oral Dataset 1 and Oral Dataset 2 for oral cancer detection. architectures. Even newly designed deep learning frameworks
On cervical dataset, VGG and ResNet showed the accuracy yield captivated results, thus favours the use of these new
of 84.20% and 84.45% respectively. On Herlev dataset, VGG architectures in many medical applications. In [41] CNN is
and ResNet showed the accuracy of 86.56% and 86.45% designed for the classification of H&E stained breast biopsy
respectively. For Oral Dataset 1 VGG and ResNet showed the images into 4 and 2 classes. The four classes in which images
accuracy of 80.66% and 78.34% respectively and Oral Dataset are classified are normal, benign lesion, in situ carcinoma and
2 VGG gained accuracy of 80.83% and ResNet 82.39% invasive carcinoma and the two classes are carcinoma and non-
respectively. carcinoma. The SVM is used as classifier where the features
extracted by CNN are given as input and classified into benign
E. Classification of cancer or malignant.
Many new architectures were designed which worked very
well for medical images. K. J. Geras et al. [25] have proposed F. Analysis of 3D volume medical images
the Multiview Deep Convolutional Neural Network (MV- CT scan and MRI produces hundreds of images for a single
DCN) that takes the HD medical images for the classification patient and analyzing all these images one by one is time-
of the breast cancer. The model is trained on 866,000 im- consuming. Therefore, 3D Computer Aided Design (CAD)
ages and has classified images as BI-RADS 1 (”Assessment technologies are used which take these CT Scan image slices
is incomplete”), BI-RADS 2 (”Negative”) and BI-RADS 3 and stack them into a concise 3D area. 3D-Convolution Neural
(”Benign findings”). S. Pereira et al. [26] have designed Networks also performed tremendously well for 3D volume
an automated system based on CNN for segmentation of a medical images. In the volumetric medical image analysis, 3D
tumour in the brain using MRI images. They have classified CNN has shown better results than 2D CNN [37]. Kingsley
brain tumour into Low-Grade Gliomas (LGG) and High- et al. [33] mentioned 3D CNN based on ResNet-101 for lung
Grade Gliomas (HGG) and have designed separate models cancer detection in 3D CT scans and placed at the 41st position
for each. For HGG and LGG, the authors have used 11 in the KAGGLE DATA SCIENCE BOWL 2017 competition.
and 9 layer architecture respectively. The proposed model The framework of the architecture is divided into four separate
is the winner of the Brain Tumor Segmentation Challenge neural networks - nodule detector, malignancy detector, nodule
2013 (BRATS 2013). S. Liu et al. [27] have proposed deep classifier, patient classifier. The ResNet-101 is modified for
learning architecture (XmasNet) based on convolution neural nodule detection and malignant detection and ResNet-18 is
network inspired by VGG for classification of prostate cancer modified for nodule classification. In [34] B. Erden et al.
lesions, using 3D multiparametric MRI image data provided have designed a 3D-CNN for the detection of a brain tumour.
by PROSTATEx challenge. P. Rao et al. [28] have trained The authors have used U-Net for the segmentation of images
CNN for lung cancer screening on Lung Images Database and then fed it as input to 3D CNN. Albert et al. [35] have
Consortium (LIDC) dataset. The proposed CanNet contains used modified U-Net to segment lung tissue trained on dataset
2 convolution layers followed by pooling layer, a dropout LUNA16. The U-Net produces many false positives so the
layer and a fully connected layer. Rotem et al. [29] have pro- segmented output of U-Net is fed into 3D convolution neural
posed Deep Convolutional Neural Network for Lung Nodule network. Two CNN models, Vanilla 3D CNN and GoogLeNet
based on 3D CNN are used to classify CT Scan as positive or on the three tasks, the first is carcinoma which is the prevalent
negative for lung cancer. type of skin cancer. The second is the most dangerous type
The deep learning is used to learn about the fundamen- of skin cancer called melanoma and the third is dermoscopy
tal process of tumour growth. In [36] the deep network is images of melanoma. The pre-trained model is retrained with
modelled to learn about cell invasion and mass effect for the their ISIC dataset and the parameters are fine-tuned across all
prediction of tumour growth. They have shown their results on layers. Transfer learning with fine-tuning reduces the training
pancreatic tumour dataset. The idea of their model is to predict time and improves the performance of a model.
whether the voxel in the current time point will have a tumour The ensemble networks were also proposed to maximize
or not at the next time point. The input is the image patches the performance of the network. In [23] the performance of
that represent cell invasion and expansive growth information GoogLeNet is improved by an ensemble of networks for pre-
derived from multimodal imaging data. The output is either processing and decision fusion to classify microcalcification of
1 (input patch will have a tumour at next time point) or 0 a breast. In contrast with standalone models, ensemble models
(input patch will not have a tumour at next point of time). show better accuracy and are highly preferable. Codella et al.
The methods of fusing invasion and expansion such as Late [24] have proposed ensemble models including deep residual
Fusion, Early Fusion and End to End Fusion are studied to networks, CNN, and fully convolutional U-Net architectures
take an advantage of the invasion and expansion information. to segment the skin lesion, analyze the detected area and its
surrounding tissue for recognition of melanoma.
G. Comparison of deep learning with traditional techniques
I. Capsule Network
Deep learning techniques show favourable results when
compared with traditional image analysis techniques. Harshita Convolutional Neural Networks are translation and spatial
et al. [38] have proposed 9 layer CNN architecture for two invariant. When the object in an image is altered in position
applications called cancer classification and necrosis detection or rotation it can still classify that object but does not predict
of stomach cancer from H&E stained images. The authors have any additional information of alternation. Hilton et al. [42]
compared their proposed CNN architecture with traditional proposed the implementation of the idea of CapsNet. A
methods - LBP histograms, Gabor filter-bank responses, Gray CapsNet is basically a neural network that performs inverse
histograms, HSV histograms, Gray level co-occurrence matrix, graphics and it is composed of many capsules as shown in
and RGB histograms followed by machine learning method Fig. 2. A capsule is a function that predicts the presence
called the random forest. The proposed model is comparatively and instantiation parameters of a particular object at a given
analyzed with standard architecture, ALexNet, for classifica- location. A capsule is a group of neurons whose output is
tion problems. RGB histograms followed by random forest the activity vector that represents the instantiation parameter
among other traditional methods achieved the highest classifi- of the entity such as an object or object part. Length of the
cation accuracy of 76.41%, AlexNet achieved 73.28% and the vector represents the probability of the existence of that entity
proposed architecture achieved 69.90%. For necrosis detection, and its orientation represent that parameter. The authors have
the proposed architecture achieved the overall highest accuracy trained their network on the MNIST dataset. They have used
of 81.44%. In [39] the efficiency of the existing prostate an iterative ‘routing by agreement’ mechanism which means
CAD was compared with the proposed Deep Learning based the capsule at the low level will send its output only to the
Prostate-CAD (CADDL ). CADDL attained 86% detection neurons at the high level which has an activity vector with
rate whereas CAD attained 80% detection rate. a big scalar product. By big scalar product we mean the
probability of having an accurate output at the last layer will
H. Transfer learning and Ensemble learning used in medical be through these neurons at a high level. The CapsNet got
image analysis the low test error (0.25%) on a 3 layer network which was
Transfer learning is an optimal solution when there is previously achieved by deep networks with many layers.
limited data available for training, where pre-trained weights
of the standard deep network architectures are used as initial-
ization weights. H.Chougrad et al. [21] have shown transfer
learning is the good approach to start the learning process
by reusing the pre-trained weights of a pre-trained model
(Inception-V3) as the initialization weights for their proposed
model. The loaded weights were fine-tuned so that the model
better fits the new Breast Cancer Digital Repository (BCDR- Fig. 2: Capsule Network by Hinton et al. [42] in 2017
F03) dataset and hence improves the performance of the
network for the better detection of breast cancer. Transfer
learning with exponential decay of learning rate yielded an IV. C HARACTERISTICS OF NEW DEEP LEARNING
accuracy of 97.50% and AUC= 0.96. A. Esteva et al. [22] have ARCHITECTURES FOR DETECTION OF CANCER
also used pre-trained weights of the Inception-V3 model for Different deep learning architectures were used for the anal-
classification of skin cancer. The authors have got their results ysis of medical images for the detection of the various cancers
TABLE II: Overview of Deep learning models for the detection of cancer
where each architecture takes the different size of images as pre-trained models for better classification and reducing the
input and the performance of the architecture is measured by training time. Regularization methods help to overcome the
different performance metrics. The architectures are trained on problem of overfitting and tuning of hyperparameters improve
large well-known datasets evaluating the phenomenal results the performance of the deep neural networks without making
as shown in Table II. any kind of changes in the architecture which is very effective
and feasible. The CNN models have worked a lot in medical
V. C ONCLUSION image analysis and performed comparatively better than the
The analysis of medical images such as X-Rays, CT Scans traditional image processing techniques. CapsNet has achieved
and MRIs is a difficult task because of the increase in the desirable accuracy by only the 3 layer architecture which
digital medical data every year which requires high potential is usually achieved by deep layer networks. We expect that
analysis techniques. Deep learning is in fame because of its medical image analysis will improve more with the use of Cap-
phenomenal classification of millions of images. It came in sNet. Deep learning is revolutionizing the healthcare with its
collaboration with the medical imaging and performed excep- extraordinary capabilities making the diagnosis and detection
tionally well. We have highlighted the use of deep learning more accurate and faster. All these emerging technologies and
techniques for the detection of cancer. The state-of-the-art such new fascinating advancements in medical sciences contribute
as ResNet, GoogLeNet or VGG are used for medical image to the better health.
analysis. Also, new architectures were designed that performed R EFERENCES
better for the classification of cancer. 3D deep networks were [1] NVIDIA - Deep learning in Medicine. [Online]. Available:
introduced for 3D volumetric medical image analysis detecting http://www.nvidia.com/object/deep-learning-in-medicine.html
the most discriminating features that classify the number of [2] Over 17 lakh new cancer cases in India by 2020: ICMR. [Online].
Available: icmr.nic.in/icmrsql/archive/2016/7.pdf
diseases. [3] Y. LeCun, L. Bottou, Y. Bengio and P. Haffner, “Gradient-Based
The ensemble of CNN models is also used to improve the Learning Applied to Document Recognition,” Proceedings of IEEE, vol.
predictive power of the model for better prediction of cancer. 86, pp. 2278-2324.
[4] Leslie N. Smith, “Cyclical Learning Rates for Training Neural Net-
Transfer learning with fine-tuning of parameters is incorpo- works,” 2017 IEEE Winter Conference on Applications of Computer
rated in medical image analysis increasing the performance of Vision (WACV), pp. 464–472, 2017.
[5] J.Wang, L.Perez, “The Effectiveness of Data Augmentation in Image [26] S. Pereira, A. Pinto, V. Alves, and C. A. Silva “Brain Tumor Segmen-
Classification using Deep Learning,” Computing Research Repository tation Using Convolutional Neural Networks in MRI Images,” IEEE
(CoRR) - arXiv, 2017. Transaction, vol. 35, 2016.
[6] N. Srivastava, G. Hinton, A. Krizhevsky and I. Sutskever and R. [27] S. Liu, H. Zheng, Y. Feng, W. Li, “Prostate Cancer Diagnosis using
Salakhutdinov, “Dropout: A Simple Way to Prevent Neural Networks Deep Learning with 3D Multiparametric MRI,” Computing Research
from Overfitting,” Journal of Machine Learning Research (JMLR) vol. Repository (CoRR) - arXiv, 2017.
15, pp. 19291958, 2014. [28] P. Rao, N. A. Pereira, and R. Srinivasan, “Convolutional Neural Net-
[7] M. S. Al-tarawneh, “Lung Cancer Detection Using Image Processing works for Lung Cancer Screening in Computed Tomography (CT)
Techniques,” Leonardo Electronic Journal of Practices and Technologies Scans,” 2016 2nd International Conference on Contemporary Computing
(LEJPT) no. 20, pp. 147–158, 2012. and Informatics (IC3I), pp. 489–493, 2016.
[8] Abhishek S. Sambyal, Asha T., “Knowledge Abstraction from Textural [29] C. Jacob, R. Golan and J. Denzinger, “Lung Nodule Detection in CT
Features of Brain MRI Images for Diagnosing Brain Tumor using Images using Deep Convolutional Neural Networks,” 2016 International
Statistical Techniques and Associative Classification,” 2016 International Joint Conference on Neural Networks (IJCNN), pp. 243–250, 2016.
Conference on Systems in Medicine and Biology, IIT Kharagpur. [30] J. Ma, Fa Wu, J. Zhu, D. Xu, D. Kong, “A Pre-trained Convolutional
[9] A. Krizhevsky, I. Sutskever and G. E. Hinton, “ImageNet Classification Neural Network-Based Method for Thyroid Nodule Diagnosis,” Ultra-
with Deep Convolutional Neural Networks,” NIPS’12 Proceedings of sonics, vol. 73, pp. 221-230, 2017.
the 25th International Conference on Neural Information Processing [31] N. Kumar, R. Verma, A. Arora, A. Kumar, S. Gupta, A. Sethi, and P. H.
Systems, vol. 1, pp. 1097–1105. Gann, “Convolutional Neural Networks for Prostate Cancer Recurrence
[10] K. Simonyan and A. Zisserman, “Very Deep Convolutional Network for Prediction,” Medical Imaging 2017, SPIE, vol. 10140, 2017.
Large-Scale Image Recognition,” International Conference on Learning [32] N. Tajbakhsh, J. Y. Sin, S. R. Gurudu, R. T. Hurst, C. B. Kendall, M. B.
Representations (ICLR) 2015. Gotway, J. Liang “Convolutional Neural Networks for Medical Image
[11] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, Analysis: Full Training or Fine Tuning?,” IEEE Transactions on Medical
V. Vanhoucke, A. Rabinovich, “Going deeper with convolutions,”2015 Imaging, vol. 35, no. 5, pp. 1299–1312, 2016.
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), [33] K. Kuan, M. Ravaut, G. Manek, H. Chen, J. Lin, B. Nazir, C. Chen,
pp.1–9, 2015. T. Chiang Howe, Z. Zeng and V. Chandrasekhar, “Deep Learning for
[12] K. He, X. Zhang, S. Ren, J. Sun, “Deep Residual Learning for Image Lung Cancer Detection: Tackling the Kaggle Data Science Bowl 2017
Recognition,” 2016 IEEE Conference on Computer Vision and Pattern Challenge,” Computing Research Repository (CoRR) - arXiv, 2017.
Recognition (CVPR), pp.770-778, 2016. [34] B. Erden, N. Gamboa, S. Wood, “3D Convolutional Neural Network for
[13] O. Ronneberger, P. Fischer, and T. Brox,“U-Net: Convolutional Net- Brain Tumor Segmentation,” Stanford Report, 2017.
works for Biomedical Image Segmentation,” MICCAI 2015, pp. [35] A. Chon, N. Balachandar and P. Lu, “Deep Convolutional Neural
234?241, 2015. Networks for Lung Cancer Detection,” Stanford Report, 2017.
[14] F. Milletari, N. Navab, and S. A. Ahmadi,“V-Net: Fully Convolutional [36] L. Zhang, L. Lu, R. M. Summers, E. Kebebew, and J. Yao, “Convolu-
Networks for Volumetric Medical Image Segmentation,” 2016 Fourth tional Invasion and Expansion Networks for Tumor Growth Prediction,”
International Conference on 3D Vision, pp.565-571, 2016. IEEE Transactions on Medical Imaging, 2017.
[15] E. Gibson, W. Li, C. Sudre, L. Fidon, D. I. Shakir, G. Wang Z. Eaton- [37] X. Huang, J. Shan and V. Vaidya, “Lung nodule detection in CT
Rosen, R. Gray, T. Doel, Y. Hu, T. Whyntie, P. Nachev, M. Modat D. C. using 3D convolutional neural networks,” 2017 IEEE 14th International
Barratt, S. Ourselin, M. Jorge Cardoso and T. Vercauteren, “NiftyNet: Symposium on Biomedical Imaging (ISBI 2017), pp. 379–383, 2017
a deep-learning platform for medical imaging,” Computer Methods and [38] H. Sharma, N. Zerbe, I. Klemper, O. Hellwich, P. Hufnagl, “Deep
Programs in Biomedicine vol. 158, pp. 113 - 122, 2018. convolutional neural networks for automatic classification of gastric
[16] S. Jgou, M. Drozdzal, D. Vzquez, A. Romero, and Y. Bengio,“The One carcinoma using whole slide images in digital histopathology,” Com-
Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic puterized medical imaging and graphics: the official journal of the
Segmentation,” arXiv:1611.09326v2 [cs.CV],2016. Computerized Medical Imaging Society, vol. 61, pp. 2–13, 2017.
[17] W. Chen, Y. Zhang, J. He, Y. Qiao, Y. Chen, H. Shi, X. [39] Y. K. Tsehay, N. S. Lay, H. R. Roth, X. Wang, J. T. Kwak, B. I.Turkbey,
Tang, “W-net: Bridged U-net for 2D Medical Image Segmentation,” Peter A. Pinto b, B. J. Wood, and R. M. Summers,“Convolutional
arXiv:1807.04459v1 [cs.CV] 12 Jul 2018 Neural Network Based Deep-learning Architecture for Prostate Cancer
[18] G. Yang, H. Jing, “Multiple Convolutional Neural Network for Feature Detection on Multiparametric Magnetic Resonance Images,” Proc. SPIE
Extraction,” International Conference on Intelligent Computing (ICIC) 10134, Medical Imaging 2017: Computer-Aided Diagnosis, 2017.
2015. [40] H. Wieslander, G. Forslid, E. Bengtsson, C. Whlby, Jan-Michal Hirsch,
[19] J. Ding, A. Li, Z. Hu, and L. Wang, “Accurate Pulmonary Nodule C. R. Stark, S. K. Sadanandan,“Deep Convolutional Neural Networks
Detection in Computed Tomography Images Using Deep Convolutional For Detecting Cellular Changes Due To Malignancy,” The IEEE Inter-
Neural Networks,” Computing Research Repository (CoRR) - arXiv, national Conference on Computer Vision (ICCV), 2017.
2017. [41] T. Arajo, G. Aresta, E. Castro, J. Rouco, P. Aguiar, C. Eloy, A. Polnia,
[20] Q. Song, L. Zhao, X. Luo, and X. Dou, “Using Deep Learning for A. Campilho, “Classification of breast cancer histology images using
Classification of Lung Nodules on Computed Tomography Images,” Convolutional Neural Networks,” PLoS ONE, 2017.
Journal of Healthcare Engineering, vol. 2017, 2017. [42] S. Sabour, C. V Nov, and G. E. Hinton, “Dynamic Routing Between
[21] H. Chougrad, H. Zouaki, O. Alheyane “Convolutional Neural Net- Capsules,” 31st Conference on Neural Information Processing Systems
works for Breast Cancer Screening: Transfer Learning with Exponential (NIPS), vol. abs/1710.09829, 2017.
Decay,”- arXiv, 2017. [43] NVIDIA - How GPUs Help Assess for Brain Damage. [Online].
[22] A. Esteva, B. Kuprel, R. A. Novoa, J.Ko, S. M. Swetter, H. M. Blau & Available: https://blogs.nvidia.com/blog/2016/03/18/brain-damage/
S. Thrun, “Dermatologist-level classification of skin cancer with deep [44] Enlitic homepage. [Online]. Available: https://www.enlitic.com/
neural networks,” Nature 542, pp. 115–118, 2017. [45] Arterys homepage. [Online]. Available: https://arterys.com/
[23] E. Sert, S. Ertekin, U. Halici, “Ensemble of Convolutional Neural [46] NVIDIA - University of Toronto Wins $200k Compute the Cure.
Networks for Classification of Breast Microcalcification from Mam- [Online]. Available: https://blogs.nvidia.com/blog/2015/11/25/compute-
mograms,” 2017 39th Annual International Conference of the IEEE the-cure-3/
Engineering in Medicine and Biology Society (EMBC), pp. 689–692,
2017.
[24] N. C. F. Codella, Q. B. Nguyen, S. Pankanti, D. Gutman, B. Helba,
A. Halpern, J. R. Smith, “Deep learning ensembles for melanoma
recognition in dermoscopy images,” Computing Research Repository
(CoRR), vol. abs/1610.04662, 2016.
[25] K. J. Geras, S.Wolfson, Y. Shen, S. Gene Kim, L. Moy, K. Cho, “High-
Resolution Breast Cancer Screening with Multi-View Deep Convolu-
tional Neural Networks,” Computing Research Repository (CoRR) -
arXiv, 2017.