SlideShare a Scribd company logo
Deep learning application to medicine
Department of Nuclear Medicine, Seoul National University Hospital
Hongyoon Choi
CONTENTS
• Basic knowledge :
Linear regression to deep learning
• Overview of deep learning
• Real application to medical data
B a s i c k n o w l e d ge :
L i n e a r re g re s s i o n to d e e p l e a r n i n g
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Age prediction
No. of wrinkles
Age
Regression:
Predict continuous value
Supervised Learning:
“Right answers” given
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Lung tumor (Benign vs malignancy)
Tumor Size
Classification:
Predict discrete value
Supervised Learning:
“Right answers” given
0: Benign
1: Malignant
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Supervised Learning Unsupervised Learning
Regression
Classification
Clustering
Generative model
Semi-Supervised Learning
Reinforcement Learning
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Training Set
Learning Algorithm
hHeight Estimated
Weight
Training set
Height(X) Weight(Y)
170
155
180
175
190
160
.
.
.
71
49
80
63
91
52
.
.
.
h(X) = a0 + a1X
L i n e a r re g re s s i o n to d e e p l e a r n i n g
h(X) = a0 + a1X
a  Parameters
How to choose “a”?
Basic idea:
Input a0 and a1 to minimize “Some Target”
Some Target = Cost Function
J(a0, a1) = Mean of (h(X) – Y)2
L i n e a r re g re s s i o n to d e e p l e a r n i n g
0
1
2
3
0 1 2 3
Simply, a0 = 0 and only consider a1
0
2
4
6
-0.5 0 0.5 1 1.5 2 2.5
X
Y
a1
J(a1)
L i n e a r re g re s s i o n to d e e p l e a r n i n g
h(X) = a0 + a1X
J(a0, a1) = Mean of (h(X) – Y)2
Hypothesis
Cost Function
Target Minimize Cost
Parameters a0 and a1
L i n e a r re g re s s i o n to d e e p l e a r n i n g
a0 a1
J
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Height
Weight
a0
a1
Optimal points of a0 and a1
L i n e a r re g re s s i o n to d e e p l e a r n i n g
How to optimize?
Gradient Descent
• Basic algorithm for training deep learning model
• Briefly,
• Start any points of a0 and a1
• Iteratively changing a0 and a1 until reach minimum J(a0,a1)
L i n e a r re g re s s i o n to d e e p l e a r n i n g
a0 a1
J
L i n e a r re g re s s i o n to d e e p l e a r n i n g
aj  aj - 𝛂
𝝏
𝝏𝒂𝒋
𝑱(𝒂 𝟎, 𝒂 𝟏) (j= 0 and 1)
Repeat until convergence
Gradient Descent Algorithm
𝛂: Learning Rate
a1
0
2
4
6
-0.5 0 0.5 1 1.5 2 2.5
J(a1)
Positive slope
L i n e a r re g re s s i o n to d e e p l e a r n i n g
a1
0
2
4
6
-0.5 0 0.5 1 1.5 2 2.5
J(a1)
a1
0
2
4
6
-0.5 0 0.5 1 1.5 2 2.5
J(a1)
Small 𝛂 large 𝛂
L i n e a r re g re s s i o n to d e e p l e a r n i n g
• Multiple Variables
h(X) = a0 + a1X1 + a2X2+a3X3 …
Weight Prediction using Height, Waist circumference, Head circumference, …
J(ai) = Mean of (h(X) – Y)2
aj  aj - 𝛂
𝝏
𝝏𝒂𝒋
𝑱(𝒂)Repeat
L i n e a r re g re s s i o n to d e e p l e a r n i n g
0
2
4
6
-0.5 0 0.5 1 1.5 2 2.5
Single Parameter Two Parameter
>3 Parameters : n-dimensional parabolic shape
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Lung tumor (Benign vs malignancy)
Tumor Size
0: Benign
1: Malignant
h(x) = aX
h(X) = 1 if X > 3cm
h(X) = 0 if X < 3cm
Classification
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Lung tumor (Benign vs malignancy)
Tumor Size
0: Benign
1: Malignant
Classification
Borderline  value of 0~1
“Logistic Function”
=“Sigmoid Function”
What we want…
1) 0< h(X) <1
2) For borderline, h(X) ~ 0.5
ℎ 𝑋 =
1
1 + 𝑒−𝑍
Z = a0 + a1X (X: tumor size)
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Lung tumor (Benign vs malignancy)
Tumor Size
0: Benign
1: Malignant
Classification
Decision Boundary
ℎ 𝑋 =
1
1 + 𝑒−𝑍
Z = a0 + a1X (X: tumor size)0.5
3 cm
H(X) > 0.5  Malignant
H(X) < 0.5  Benign
Interpretation of h(X)
~ Probability of malignant
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Classification
x1
x2
1 2 3
1
2
3
Classification with two variables
x1
x2
1 2 3
1
2
3
Decision boundary
= Threshold
x1
x2
1 2 3
1
2
3Y=1
Y=0
Sigmoid
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Classification
Classification with multiple variables
x1
x2
1 2 3
1
2
3 ℎ 𝑋 =
1
1 + 𝑒−𝑍
Z = a0 + a1X1+a2X2
Linear regression h(X) = a0+a1x1+a2x2+…
Logistic classification h(X) = sig(Z) ,
Z = a0+a1x1+a2x2+…
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Classification
Tumor Size
0: Benign
1: Malignant
Tumor Size
0: Benign
1: Malignant
ℎ 𝑋 =
1
1 + 𝑒−𝑍 Z = a0 + a1X
Changing a0
Changing a1
How to optimize aj?
 Make convex cost function
J(a0, a1) = Mean of (h(X) – Y)2
For linear regression,
0
2
4
6
-0.5 0 0.5 1 1.5 2 2.5
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Classification
Cost function for logistic classification
J(a0, a1) = -log(h(X)) if Y = 1
-log(1-h(X)) if Y = 0
h(X)
J
10
(h(X) is 0~1)
Y=1
h(X)
J
10
(h(X) is 0~1)
Y=0
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Classification
Cost function for logistic classification
J(a0, a1) = -[Y log (h(X)) + (1-Y) log (1-h(X))]
Cost function for logistic = “Binary crossentropy”
aj  aj - 𝛂
𝝏
𝝏𝒂𝒋
𝑱(𝒂)Repeat
Optimization algorithm : Same with linear regression
L i n e a r re g re s s i o n to d e e p l e a r n i n g
• Logistic regression as a Perceptron
Lesion size
Circularity
Hounsfield unit
x1
x2
x3
Z
w1
w2
w3
b0
Activation function
Sigmoid(Z)
Output
1: Malignancy
0: Benign
Find optimized W for minimized error
 Gradient descent
L i n e a r re g re s s i o n to d e e p l e a r n i n g
• Perceptron vs neuron
L i n e a r re g re s s i o n to d e e p l e a r n i n g
x1
x2
x1
x2
1 2 3
1
2
3
Linear classification Nonlinear classification
• Limitation of single-layer perceptron
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Layer 3Layer 1 Layer 2
H(X)input
Output layer
a1
(2) = g(w11x1+w12x2+w13x3)
a2
(2) = g(w21x1+w22x2+w23x3)
a3
(2) = g(w31x1+w32x2+w33x3)
where g: activation function (sigmoid)
a1
(3)
= H(X)
= g(w21a1
(2)+w22a2
(2)+w23a3
(2))
Single-layer perceptron to neural network
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Non-linear classification example: XOR/XNOR
x1 and x2 are binary (0 or 1).
x1
x2
Y=1
Y=0
(XNOR Problem)
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Non-linear classification example: XOR
AND
X2
X1
+1 -30
20
20
H(X)
X2
X1
+1 -10
20
20
H(X)
OR
sigmoid(-10)~0
sigmoid(10)~1
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Non-linear classification example: XOR
AND
X2
X1
+1
a1
-30
20
20
(NOT x1) AND (NOT x2)
a2
10
-20
-20
+1
H(X)
-10
20
20
OR
X1 X2
0 0
0 1
1 0
1 1
a1 a2
0 1
0 0
0 0
1 0
1
0
0
1
H(X)
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Output
Lesion size
Circularity
Lesion size
Circularity
Malignancy Benign
Non-linear classification
L i n e a r re g re s s i o n to d e e p l e a r n i n g
H
O Output
How to calculate gradient descent ?
𝜕𝐽
𝜕𝑤2
=
𝜕𝐽
𝜕𝑂
𝜕𝑂
𝜕𝑍 𝑂
𝜕𝑍 𝑂
𝜕𝑤2
w2 𝑍 𝑜 = ෍ 𝑊2 𝐻
J = Cost (O, Y)
(O = h(X))
J= -[Y log (h(X)) + (1-Y) log (1-h(X))]
w1
𝜕𝐽
𝜕𝑤1
=
𝜕𝐽
𝜕𝑍ℎ
𝜕𝑍ℎ
𝜕𝑤1
H=sig(Zh) O=sig(ZO)
Zh=W1X ZO=W2H
“Back Propagation”
𝜕𝐽
𝜕𝑍ℎ
=
𝜕𝐽
𝜕𝑍 𝑂
𝜕𝑍 𝑂
𝜕𝐻
𝜕𝐻
𝜕𝑍ℎ
L i n e a r re g re s s i o n to d e e p l e a r n i n g
The era of artificial brain!
L i n e a r re g re s s i o n to d e e p l e a r n i n g
You see this:
But the camera sees this:
• Limitation of conventional neural network
L i n e a r re g re s s i o n to d e e p l e a r n i n g
• Limitation of conventional neural network
pixel 1
pixel 2
Cars
“Non”-Cars
50 x 50 pixel images→ 2500 pixels
(7500 if RGB)
pixel 1
pixel 2
= A point at
7500 dimensional axis
 7500 multivariable logistic regression
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Curse of dimensionality
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Curse of dimensionality
L i n e a r re g re s s i o n to d e e p l e a r n i n g
More dimensions
More sparse in data space
 Easy to overfit
L i n e a r re g re s s i o n to d e e p l e a r n i n g
More dimensions  More weights (Parameters to learn)
1990~2000
Better manual features instead of raw pixel value
+ Kernel-based learning
L i n e a r re g re s s i o n to d e e p l e a r n i n g
• Hard to learn deep layer
Problem of Vanishing Gradient
L i n e a r re g re s s i o n to d e e p l e a r n i n g
• Hard to learn deep layer
Slope~0
Slope~0
Output
Error Backpropagation
𝜕𝐽
𝜕𝑤2
=
𝜕𝐽
𝜕𝑂
𝜕𝑂
𝜕𝑍 𝑂
𝜕𝑍 𝑂
𝜕𝑤2
O=sig(ZO)
~0
L i n e a r re g re s s i o n to d e e p l e a r n i n g
Overcome Limitations &
To Deep Learning…
• Automatic feature extraction from raw data
• Learning deep-layered neural network
Algorithm Hardware Big Data
Overview of Deep Learning
Deep learning
• Nonlinear problem & multilayer perceptron
0 1 1 3 5 7 8 8
0 0 0 1 3 3 5 3
0 0 1 2 4 7 7 1
0 0 2 3 8 5 7 6
2 5 8 8 8 4 9 5
0 0 8 8 6 4 2 3
128x128
=16,384
Output
- Require good manual features
- Raw data  Too big.
- More layers?  Difficulty in learning
Deep learning
• MLP to Deep learning
- Require good manual features
- Raw data  Too big.
- More layers?  Difficulty in learning
• Automatic feature extraction from raw data
• New activation function
& Stochastic gradient descent
• Methods for reducing overfitting
Deep learning
• Train deep layer
Problem of Vanishing Gradient
Slope~0
Slope~0
Deep learning
• Train deep layer
Nonlinearity function
Sigmoid  ReLU , tanh, ELU, Leaky ReLU
Sigmoid
ReLU
Slope~0
Slope~0
Output
Error Backpropagation
Slope = 1
Deep learning
• Train big data
aj  aj - 𝛂
𝝏
𝝏𝒂𝒋
𝑱(𝒂)RepeatGradient Descent
50x50x3 ~7500 pixel data per image
100,000 images of cars and non-cars
Front :
- Estimate 100,000 h(X)
Back = update ‘a’
- Cost calculated by 100,000 cost(Y,h(X))
a1
0
2
4
6
-0.5 0 0.5 1 1.5 2 2.5
J(a1)
Small 𝛂
Per iteration
Deep learning
• Train big data
Training data
1 weight update
Training data
1 weight update
per mini-batch
Multiple weight updates
Gradient
Descent
Stochastic
Gradient
Descent
(=batch gradient)
(=minibatch stochastic gradient)
Deep learning
• Train big data
Faster and efficient
to reach
global minima
Deep learning
• Problem of overfitting
Apple
Not apple
Because it’s yellow
Not apple
Because it’s not round
Deep learning
• Problem of overfitting
Dropout
Deep learning
• Automatic feature extraction efficiently
– Convolutional Neural Network for Image
– Recurrent Neural Network for Sequential Data
Deep learning
• Convolutional Neural Network
• Local feature extraction
• Translational invariance
• Sparsity compared with fully-connected layer
Deep learning
• Convolutional Neural Network
• Local feature extraction
• Translational invariance
• Sparsity compared with fully-connected layer
Fully-connected layer Convolutional layer
Deep learning
• Convolutional Neural Network
Convolutional filter
0 1 1 3 5 7 8 8
0 0 0 1 3 3 5 3
0 0 1 2 4 7 7 1
0 0 2 3 8 5 7 6
2 5 8 8 8 4 9 5
0 0 8 8 6 4 2 3
Image Filter
1 0 1
0 1 0
1 0 1
2 6 12 22 27 28
2 5 15 16 30 24
11 17 24 29 33 24
15 19 32 28 27 27
Output
Number of feature maps
= number of convolutional filters
Instead
number of
nodes
Deep learning
identify line / some texture
identify head lights and wheels
identify Car!
Deep learning
• Convolutional Neural Network
Deep learning
• Convolutional Neural Network
Initial Data : 256 x 256
Ear Ear
Eye Eye
Nose Tail
Foot Foot
After convolutions and poolings
=Abstracted features
Deep learning
• Convolutional Neural Network
Abstracted Features Feature vectors
Multivariate
Logistic
Dimension 256x256x3 4x4x1024 4096 1000
Deep learning
• Convolutional Neural Network
ImageNet Challenge Results
28.2%
2010
25.8%
2011
16.4%
2012
Shallow model
AlexNet
11.7%
2013
6.7%
2014
3.57%
2015
GoogleNet
ResNet
8-layers
22-layers
152-layers
Deep learning
• Convolutional Neural Network
ImageNet Challenge Results
AlexNet
GoogleNet
ResNet
Deep learning
Pedestrian Car Motorcycle Truck
• Cf> Multiple output (instead of binary classification)
4 Output nodes,
instead of 1 node
Y = [ 1, 0 , 0, 0] for pedestrian
Y = [ 0, 1, 0, 0 ] for car
Y = [ 0, 0, 1, 0 ] for motorcycle
Y = [ 0, 0, 0, 1 ] for truck
Activation function:
Softmax, instead of sigmoid
Deep learning
• Recurrent Neural Network
Deep learning
• Recurrent Neural Network
Deep learning
• Recurrent Neural Network
Vision
Deep CNN
Language
Generating
RNN
“A group of people shopping
at an outdoor market.
There are many vegetables at
the fruit stands.”
Neural Machine Translation
Google Translate
Text, Music Generation
https://www.youtube.com/watch?v=A2gyidoFsoI
Combined with CNN : Image caption generation
Deep learning
• Current Concept of Deep learning
Deep layered
neural network
Output
+
Data type-specific
layers
Convolution
Recurrent
Modification for
training
+
ReLU activation
SGD training
Dropout
Batch normalization
Variable Cost Function
…
Deep learning
Nomenclature
Supervised
Learning
Unsupervised
Learning
Reinforcement
Learning
• Regression
• Classification
• Clustering
• Generative model
• Algorithms
react to environment
Machine Learning
Deep learning
Unsupervised learning
, particularly
generative model
Transfer learning One-shot learning
Bayesian modeling Mobile-friendly model
Manifold and non-
Euclidean data
Current trends of deep learning
High accuracy to various purposes/situations
Deep learning
Flexible and scalable deep learning
• Transfer learning
Car
ImageNet-based model as a feature extractor
Train only
last layer
Deep learning
Flexible and scalable deep learning
• Generative model
z~N(0,1)
G:
generator
Fake image
Real image
D:
Discriminator
1: real
0: fake
Can be composed with
convolutional, FC layers,
Batch normalization,
regularization, etc.
Various cost functions
/ combined cost functions
MSE, CE, Adversarial, KLD, etc.
Deep learning
Flexible and scalable deep learning
Ref. DeepMind, NIPS 2017
https://tykimos.github.io/
Deep learning in Medicine
Deep learning in Medicine
ROC curve
- better than dermatologists
Esteva, Andre, et al. Nature 2017
Deep learning in Medicine
Diabetic Retinopathy
 Better or equivalent
to ophthalmologists
Normal DM
Gulshan, Varun, et al. JAMA 2016
ChestXnet
 Equivalent/Superior to radiologists (?)
Pranav Rajpurkar, … Andrew Ng, 2017. Arxiv
DL for medical imaging:
Supervised learning using CNN
Deep learning in Medicine
FDA approve a device for diagnosing
diabetic retinopathy (2018.4)
AI-aided system (CT angiography
for large vessel occlusion)
Year of AI invasion to clinic
Deep learning in Medicine
DL for medical imaging:
Supervised learning using CNN
AD & NC
MCI-converter & non-converter
FDG and amyloid PET to predict future cognitive decline
Choi H and Jin KH Arxiv 2017
Deep learning in Medicine
https://adfdgpet.appspot.com
Online Demo
Input file
Web application
Output: likelihood for AD
& predicted cognitive score
Output:
Cognitive dysfunction-related map
p(Alzheimer|X)
Deep learning in Medicine
https://insight.lunit.io/
Web application
Deep learning in Medicine
Web application
https://modelderm.com
Han SS, et al. J Invest Derm 2018
Deep learning in Medicine
Laborious Work Replaced by DL:
Segmentation
Choi, H., & Jin, K. H. J Neurosci Methods 2016 de Brebisson, et al. CVPR 2015.
Deep learning in Medicine
Laborious Work Replaced by DL:
Detection
Liu Y, et al. Arxiv 2017
Deep learning in Medicine
Enhance Image Acquisition & Quality
Dahl et al. Arxiv 2017
Normal dose abdomen CT Low dose abdomen CT
Low dose abdomen CT+CNN
Chen H et al. Biomed Opt Exp 2017
Deep learning in Medicine
Image Generation
Cat
Cat
Common deep learning model Generative model
z = f(x)
where x: data, z: discriminative features
f: classifier model
x = g(z)
where x: data, z: latent
g: generation function
Deep learning in Medicine
Generative Adversarial Network
z~N(0,1)
G:
generator
Karras T, et al. Arxiv 2017
G:
generator
Isola P, et al. arxiv, 2016.
Generative Adversarial Network
Structural MR generation from PET
Florbetapir PET
Generator:
U-net
Skip connection
Generated MR
PETandgeneratedMRPETandrealMR
Discriminator Real or Fake
Generative Adversarial Networks
for MR generation
z G(z)
z & G(z)
z & x
Choi H and Lee DS, J Nucl Med 2017.
Deep learning in Medicine
RealMRI
Generated
MRI
18F-Florbetapir
PET
Deep learning in Medicine
Conditional Generation
Antipov G, Arxiv 2017
Encoder
Latent
features
Generator
+ Age
Latent
features
+ Age
VAE model for brain PET
generation
 Brain metabolism aging movie
Choi H,… Lee DS. Biorxiv 2017
Deep learning in Medicine
Conditional Generation
Choi H,… Lee DS. Biorxiv 2017
Estimating normal population distribution
Deep learning in Medicine
Deep learning in Medicine
Omics data
Quang D et al. NAR 2016
Predicting Function from
DNA sequences
Deep learning in Medicine
Omics data
Low risk group
High risk group
Deep learning-based risk score
Choi H and Na KJ, Biomed Res Int. 2017
Deep learning in Medicine
Disruptive Innovation: Raw medical & healthcare data
Diet + Previous Glucose Level
Future Glucose Level &
Scheduling Insulin
Sugar.iq from Medtronic
Deep learning in Medicine
Disruptive Innovation: Raw medical & healthcare data
Deep learning in Medicine
Disruptive Innovation: Raw medical & healthcare data
HTC DeepQ Tricoder
Predicting PVC from daily EKG
Diagnosis of otitis media
Deep learning in Medicine
Disruptive Innovation: Raw medical & healthcare data
Deep learning facilitates left-shifting

More Related Content

What's hot

Proximal Splitting and Optimal Transport
Proximal Splitting and Optimal TransportProximal Splitting and Optimal Transport
Proximal Splitting and Optimal Transport
Gabriel Peyré
 
Runtime Analysis of Population-based Evolutionary Algorithms
Runtime Analysis of Population-based Evolutionary AlgorithmsRuntime Analysis of Population-based Evolutionary Algorithms
Runtime Analysis of Population-based Evolutionary Algorithms
PK Lehre
 
Lecture 1: linear SVM in the primal
Lecture 1: linear SVM in the primalLecture 1: linear SVM in the primal
Lecture 1: linear SVM in the primal
Stéphane Canu
 
Relaxed Utility Maximization in Complete Markets
Relaxed Utility Maximization in Complete MarketsRelaxed Utility Maximization in Complete Markets
Relaxed Utility Maximization in Complete Marketsguasoni
 
Scientific Computing with Python Webinar 9/18/2009:Curve Fitting
Scientific Computing with Python Webinar 9/18/2009:Curve FittingScientific Computing with Python Webinar 9/18/2009:Curve Fitting
Scientific Computing with Python Webinar 9/18/2009:Curve Fitting
Enthought, Inc.
 
Options Portfolio Selection
Options Portfolio SelectionOptions Portfolio Selection
Options Portfolio Selection
guasoni
 
Actuarial Science Reference Sheet
Actuarial Science Reference SheetActuarial Science Reference Sheet
Actuarial Science Reference SheetDaniel Nolan
 
ABC: How Bayesian can it be?
ABC: How Bayesian can it be?ABC: How Bayesian can it be?
ABC: How Bayesian can it be?
Christian Robert
 
Modeling the Dynamics of SGD by Stochastic Differential Equation
Modeling the Dynamics of SGD by Stochastic Differential EquationModeling the Dynamics of SGD by Stochastic Differential Equation
Modeling the Dynamics of SGD by Stochastic Differential Equation
Mark Chang
 
ABC-Xian
ABC-XianABC-Xian
ABC-XianDeb Roy
 
Md2521102111
Md2521102111Md2521102111
Md2521102111
IJERA Editor
 
2D1431 Machine Learning
2D1431 Machine Learning2D1431 Machine Learning
2D1431 Machine Learningbutest
 
Stochastic calculus
Stochastic calculus Stochastic calculus
Stochastic calculus
Praveen Kumar Gangarapu
 
5 volumes and solids of revolution i x
5 volumes and solids of revolution i x5 volumes and solids of revolution i x
5 volumes and solids of revolution i x
math266
 
QMC: Operator Splitting Workshop, Boundedness of the Sequence if Iterates Gen...
QMC: Operator Splitting Workshop, Boundedness of the Sequence if Iterates Gen...QMC: Operator Splitting Workshop, Boundedness of the Sequence if Iterates Gen...
QMC: Operator Splitting Workshop, Boundedness of the Sequence if Iterates Gen...
The Statistical and Applied Mathematical Sciences Institute
 
Introduction to Stochastic calculus
Introduction to Stochastic calculusIntroduction to Stochastic calculus
Introduction to Stochastic calculus
Ashwin Rao
 
Lecture 2: linear SVM in the Dual
Lecture 2: linear SVM in the DualLecture 2: linear SVM in the Dual
Lecture 2: linear SVM in the Dual
Stéphane Canu
 
Introduction to Bayesian Methods
Introduction to Bayesian MethodsIntroduction to Bayesian Methods
Introduction to Bayesian Methods
Corey Chivers
 
Bayesian statistics using r intro
Bayesian statistics using r   introBayesian statistics using r   intro
Bayesian statistics using r intro
BayesLaplace1
 
Information in the Weights
Information in the WeightsInformation in the Weights
Information in the Weights
Mark Chang
 

What's hot (20)

Proximal Splitting and Optimal Transport
Proximal Splitting and Optimal TransportProximal Splitting and Optimal Transport
Proximal Splitting and Optimal Transport
 
Runtime Analysis of Population-based Evolutionary Algorithms
Runtime Analysis of Population-based Evolutionary AlgorithmsRuntime Analysis of Population-based Evolutionary Algorithms
Runtime Analysis of Population-based Evolutionary Algorithms
 
Lecture 1: linear SVM in the primal
Lecture 1: linear SVM in the primalLecture 1: linear SVM in the primal
Lecture 1: linear SVM in the primal
 
Relaxed Utility Maximization in Complete Markets
Relaxed Utility Maximization in Complete MarketsRelaxed Utility Maximization in Complete Markets
Relaxed Utility Maximization in Complete Markets
 
Scientific Computing with Python Webinar 9/18/2009:Curve Fitting
Scientific Computing with Python Webinar 9/18/2009:Curve FittingScientific Computing with Python Webinar 9/18/2009:Curve Fitting
Scientific Computing with Python Webinar 9/18/2009:Curve Fitting
 
Options Portfolio Selection
Options Portfolio SelectionOptions Portfolio Selection
Options Portfolio Selection
 
Actuarial Science Reference Sheet
Actuarial Science Reference SheetActuarial Science Reference Sheet
Actuarial Science Reference Sheet
 
ABC: How Bayesian can it be?
ABC: How Bayesian can it be?ABC: How Bayesian can it be?
ABC: How Bayesian can it be?
 
Modeling the Dynamics of SGD by Stochastic Differential Equation
Modeling the Dynamics of SGD by Stochastic Differential EquationModeling the Dynamics of SGD by Stochastic Differential Equation
Modeling the Dynamics of SGD by Stochastic Differential Equation
 
ABC-Xian
ABC-XianABC-Xian
ABC-Xian
 
Md2521102111
Md2521102111Md2521102111
Md2521102111
 
2D1431 Machine Learning
2D1431 Machine Learning2D1431 Machine Learning
2D1431 Machine Learning
 
Stochastic calculus
Stochastic calculus Stochastic calculus
Stochastic calculus
 
5 volumes and solids of revolution i x
5 volumes and solids of revolution i x5 volumes and solids of revolution i x
5 volumes and solids of revolution i x
 
QMC: Operator Splitting Workshop, Boundedness of the Sequence if Iterates Gen...
QMC: Operator Splitting Workshop, Boundedness of the Sequence if Iterates Gen...QMC: Operator Splitting Workshop, Boundedness of the Sequence if Iterates Gen...
QMC: Operator Splitting Workshop, Boundedness of the Sequence if Iterates Gen...
 
Introduction to Stochastic calculus
Introduction to Stochastic calculusIntroduction to Stochastic calculus
Introduction to Stochastic calculus
 
Lecture 2: linear SVM in the Dual
Lecture 2: linear SVM in the DualLecture 2: linear SVM in the Dual
Lecture 2: linear SVM in the Dual
 
Introduction to Bayesian Methods
Introduction to Bayesian MethodsIntroduction to Bayesian Methods
Introduction to Bayesian Methods
 
Bayesian statistics using r intro
Bayesian statistics using r   introBayesian statistics using r   intro
Bayesian statistics using r intro
 
Information in the Weights
Information in the WeightsInformation in the Weights
Information in the Weights
 

Similar to Basic deep learning & Deep learning application to medicine

ch3.ppt
ch3.pptch3.ppt
The world of loss function
The world of loss functionThe world of loss function
The world of loss function
홍배 김
 
Regression
RegressionRegression
Regression
Ncib Lotfi
 
Runtime Analysis of Population-based Evolutionary Algorithms
Runtime Analysis of Population-based Evolutionary AlgorithmsRuntime Analysis of Population-based Evolutionary Algorithms
Runtime Analysis of Population-based Evolutionary Algorithms
Per Kristian Lehre
 
MLHEP 2015: Introductory Lecture #1
MLHEP 2015: Introductory Lecture #1MLHEP 2015: Introductory Lecture #1
MLHEP 2015: Introductory Lecture #1
arogozhnikov
 
L1 intro2 supervised_learning
L1 intro2 supervised_learningL1 intro2 supervised_learning
L1 intro2 supervised_learning
Yogendra Singh
 
2 random variables notes 2p3
2 random variables notes 2p32 random variables notes 2p3
2 random variables notes 2p3
MuhannadSaleh
 
A Mathematically Derived Number of Resamplings for Noisy Optimization (GECCO2...
A Mathematically Derived Number of Resamplings for Noisy Optimization (GECCO2...A Mathematically Derived Number of Resamplings for Noisy Optimization (GECCO2...
A Mathematically Derived Number of Resamplings for Noisy Optimization (GECCO2...
Jialin LIU
 
Sums.of.equivalent.sequences.of.positive.operators.Taft.Beamer.Presentation
Sums.of.equivalent.sequences.of.positive.operators.Taft.Beamer.PresentationSums.of.equivalent.sequences.of.positive.operators.Taft.Beamer.Presentation
Sums.of.equivalent.sequences.of.positive.operators.Taft.Beamer.PresentationDragan (dragancn)
 
Categorical data analysis full lecture note PPT.pptx
Categorical data analysis full lecture note  PPT.pptxCategorical data analysis full lecture note  PPT.pptx
Categorical data analysis full lecture note PPT.pptx
MinilikDerseh1
 
A Proof of the Generalized Riemann Hypothesis
A Proof of the Generalized Riemann HypothesisA Proof of the Generalized Riemann Hypothesis
A Proof of the Generalized Riemann Hypothesis
Charaf Ech-Chatbi
 
Machine Learning lecture4(logistic regression)
Machine Learning lecture4(logistic regression)Machine Learning lecture4(logistic regression)
Machine Learning lecture4(logistic regression)
cairo university
 
A Proof of the Generalized Riemann Hypothesis
A Proof of the Generalized Riemann HypothesisA Proof of the Generalized Riemann Hypothesis
A Proof of the Generalized Riemann Hypothesis
Charaf Ech-Chatbi
 
Machine Learning lecture2(linear regression)
Machine Learning lecture2(linear regression)Machine Learning lecture2(linear regression)
Machine Learning lecture2(linear regression)
cairo university
 
columbus15_cattaneo.pdf
columbus15_cattaneo.pdfcolumbus15_cattaneo.pdf
columbus15_cattaneo.pdf
AhmadM65
 
Skiena algorithm 2007 lecture02 asymptotic notation
Skiena algorithm 2007 lecture02 asymptotic notationSkiena algorithm 2007 lecture02 asymptotic notation
Skiena algorithm 2007 lecture02 asymptotic notationzukun
 
UNIT I_5.pdf
UNIT I_5.pdfUNIT I_5.pdf
UNIT I_5.pdf
Muthukumar P
 
Algebric Functions.pdf
Algebric Functions.pdfAlgebric Functions.pdf
Algebric Functions.pdf
MamadArip
 

Similar to Basic deep learning & Deep learning application to medicine (20)

ch3.ppt
ch3.pptch3.ppt
ch3.ppt
 
The world of loss function
The world of loss functionThe world of loss function
The world of loss function
 
Regression
RegressionRegression
Regression
 
Runtime Analysis of Population-based Evolutionary Algorithms
Runtime Analysis of Population-based Evolutionary AlgorithmsRuntime Analysis of Population-based Evolutionary Algorithms
Runtime Analysis of Population-based Evolutionary Algorithms
 
MLHEP 2015: Introductory Lecture #1
MLHEP 2015: Introductory Lecture #1MLHEP 2015: Introductory Lecture #1
MLHEP 2015: Introductory Lecture #1
 
L1 intro2 supervised_learning
L1 intro2 supervised_learningL1 intro2 supervised_learning
L1 intro2 supervised_learning
 
2 random variables notes 2p3
2 random variables notes 2p32 random variables notes 2p3
2 random variables notes 2p3
 
A Mathematically Derived Number of Resamplings for Noisy Optimization (GECCO2...
A Mathematically Derived Number of Resamplings for Noisy Optimization (GECCO2...A Mathematically Derived Number of Resamplings for Noisy Optimization (GECCO2...
A Mathematically Derived Number of Resamplings for Noisy Optimization (GECCO2...
 
Sums.of.equivalent.sequences.of.positive.operators.Taft.Beamer.Presentation
Sums.of.equivalent.sequences.of.positive.operators.Taft.Beamer.PresentationSums.of.equivalent.sequences.of.positive.operators.Taft.Beamer.Presentation
Sums.of.equivalent.sequences.of.positive.operators.Taft.Beamer.Presentation
 
Categorical data analysis full lecture note PPT.pptx
Categorical data analysis full lecture note  PPT.pptxCategorical data analysis full lecture note  PPT.pptx
Categorical data analysis full lecture note PPT.pptx
 
A Proof of the Generalized Riemann Hypothesis
A Proof of the Generalized Riemann HypothesisA Proof of the Generalized Riemann Hypothesis
A Proof of the Generalized Riemann Hypothesis
 
Machine Learning lecture4(logistic regression)
Machine Learning lecture4(logistic regression)Machine Learning lecture4(logistic regression)
Machine Learning lecture4(logistic regression)
 
A Proof of the Generalized Riemann Hypothesis
A Proof of the Generalized Riemann HypothesisA Proof of the Generalized Riemann Hypothesis
A Proof of the Generalized Riemann Hypothesis
 
Machine Learning lecture2(linear regression)
Machine Learning lecture2(linear regression)Machine Learning lecture2(linear regression)
Machine Learning lecture2(linear regression)
 
CHAPTER 1.ppt
CHAPTER 1.pptCHAPTER 1.ppt
CHAPTER 1.ppt
 
columbus15_cattaneo.pdf
columbus15_cattaneo.pdfcolumbus15_cattaneo.pdf
columbus15_cattaneo.pdf
 
Skiena algorithm 2007 lecture02 asymptotic notation
Skiena algorithm 2007 lecture02 asymptotic notationSkiena algorithm 2007 lecture02 asymptotic notation
Skiena algorithm 2007 lecture02 asymptotic notation
 
Slides univ-van-amsterdam
Slides univ-van-amsterdamSlides univ-van-amsterdam
Slides univ-van-amsterdam
 
UNIT I_5.pdf
UNIT I_5.pdfUNIT I_5.pdf
UNIT I_5.pdf
 
Algebric Functions.pdf
Algebric Functions.pdfAlgebric Functions.pdf
Algebric Functions.pdf
 

More from Hongyoon Choi

How deep learning reshapes medicine
How deep learning reshapes medicineHow deep learning reshapes medicine
How deep learning reshapes medicine
Hongyoon Choi
 
Integrative analysis of medical imaging and omics
Integrative analysis of medical imaging and omicsIntegrative analysis of medical imaging and omics
Integrative analysis of medical imaging and omics
Hongyoon Choi
 
Deep learning application to medical imaging: Perspectives as a physician
Deep learning application to medical imaging: Perspectives as a physicianDeep learning application to medical imaging: Perspectives as a physician
Deep learning application to medical imaging: Perspectives as a physician
Hongyoon Choi
 
Neurobiology - Synthetic biology (2014)
Neurobiology - Synthetic biology (2014)Neurobiology - Synthetic biology (2014)
Neurobiology - Synthetic biology (2014)
Hongyoon Choi
 
2014 2nd neuralactivityimaging_pe_tand_cal2
2014 2nd neuralactivityimaging_pe_tand_cal22014 2nd neuralactivityimaging_pe_tand_cal2
2014 2nd neuralactivityimaging_pe_tand_cal2
Hongyoon Choi
 
2013 machine learning_choih
2013 machine learning_choih2013 machine learning_choih
2013 machine learning_choih
Hongyoon Choi
 

More from Hongyoon Choi (6)

How deep learning reshapes medicine
How deep learning reshapes medicineHow deep learning reshapes medicine
How deep learning reshapes medicine
 
Integrative analysis of medical imaging and omics
Integrative analysis of medical imaging and omicsIntegrative analysis of medical imaging and omics
Integrative analysis of medical imaging and omics
 
Deep learning application to medical imaging: Perspectives as a physician
Deep learning application to medical imaging: Perspectives as a physicianDeep learning application to medical imaging: Perspectives as a physician
Deep learning application to medical imaging: Perspectives as a physician
 
Neurobiology - Synthetic biology (2014)
Neurobiology - Synthetic biology (2014)Neurobiology - Synthetic biology (2014)
Neurobiology - Synthetic biology (2014)
 
2014 2nd neuralactivityimaging_pe_tand_cal2
2014 2nd neuralactivityimaging_pe_tand_cal22014 2nd neuralactivityimaging_pe_tand_cal2
2014 2nd neuralactivityimaging_pe_tand_cal2
 
2013 machine learning_choih
2013 machine learning_choih2013 machine learning_choih
2013 machine learning_choih
 

Recently uploaded

heat stroke and heat exhaustion in children
heat stroke and heat exhaustion in childrenheat stroke and heat exhaustion in children
heat stroke and heat exhaustion in children
SumeraAhmad5
 
micro teaching on communication m.sc nursing.pdf
micro teaching on communication m.sc nursing.pdfmicro teaching on communication m.sc nursing.pdf
micro teaching on communication m.sc nursing.pdf
Anurag Sharma
 
Flu Vaccine Alert in Bangalore Karnataka
Flu Vaccine Alert in Bangalore KarnatakaFlu Vaccine Alert in Bangalore Karnataka
Flu Vaccine Alert in Bangalore Karnataka
addon Scans
 
Tom Selleck Health: A Comprehensive Look at the Iconic Actor’s Wellness Journey
Tom Selleck Health: A Comprehensive Look at the Iconic Actor’s Wellness JourneyTom Selleck Health: A Comprehensive Look at the Iconic Actor’s Wellness Journey
Tom Selleck Health: A Comprehensive Look at the Iconic Actor’s Wellness Journey
greendigital
 
Non-respiratory Functions of the Lungs.pdf
Non-respiratory Functions of the Lungs.pdfNon-respiratory Functions of the Lungs.pdf
Non-respiratory Functions of the Lungs.pdf
MedicoseAcademics
 
Are There Any Natural Remedies To Treat Syphilis.pdf
Are There Any Natural Remedies To Treat Syphilis.pdfAre There Any Natural Remedies To Treat Syphilis.pdf
Are There Any Natural Remedies To Treat Syphilis.pdf
Little Cross Family Clinic
 
New Drug Discovery and Development .....
New Drug Discovery and Development .....New Drug Discovery and Development .....
New Drug Discovery and Development .....
NEHA GUPTA
 
Triangles of Neck and Clinical Correlation by Dr. RIG.pptx
Triangles of Neck and Clinical Correlation by Dr. RIG.pptxTriangles of Neck and Clinical Correlation by Dr. RIG.pptx
Triangles of Neck and Clinical Correlation by Dr. RIG.pptx
Dr. Rabia Inam Gandapore
 
Evaluation of antidepressant activity of clitoris ternatea in animals
Evaluation of antidepressant activity of clitoris ternatea in animalsEvaluation of antidepressant activity of clitoris ternatea in animals
Evaluation of antidepressant activity of clitoris ternatea in animals
Shweta
 
For Better Surat #ℂall #Girl Service ❤85270-49040❤ Surat #ℂall #Girls
For Better Surat #ℂall #Girl Service ❤85270-49040❤ Surat #ℂall #GirlsFor Better Surat #ℂall #Girl Service ❤85270-49040❤ Surat #ℂall #Girls
For Better Surat #ℂall #Girl Service ❤85270-49040❤ Surat #ℂall #Girls
Savita Shen $i11
 
Maxilla, Mandible & Hyoid Bone & Clinical Correlations by Dr. RIG.pptx
Maxilla, Mandible & Hyoid Bone & Clinical Correlations by Dr. RIG.pptxMaxilla, Mandible & Hyoid Bone & Clinical Correlations by Dr. RIG.pptx
Maxilla, Mandible & Hyoid Bone & Clinical Correlations by Dr. RIG.pptx
Dr. Rabia Inam Gandapore
 
How to Give Better Lectures: Some Tips for Doctors
How to Give Better Lectures: Some Tips for DoctorsHow to Give Better Lectures: Some Tips for Doctors
How to Give Better Lectures: Some Tips for Doctors
LanceCatedral
 
Phone Us ❤85270-49040❤ #ℂall #gIRLS In Surat By Surat @ℂall @Girls Hotel With...
Phone Us ❤85270-49040❤ #ℂall #gIRLS In Surat By Surat @ℂall @Girls Hotel With...Phone Us ❤85270-49040❤ #ℂall #gIRLS In Surat By Surat @ℂall @Girls Hotel With...
Phone Us ❤85270-49040❤ #ℂall #gIRLS In Surat By Surat @ℂall @Girls Hotel With...
Savita Shen $i11
 
263778731218 Abortion Clinic /Pills In Harare ,
263778731218 Abortion Clinic /Pills In Harare ,263778731218 Abortion Clinic /Pills In Harare ,
263778731218 Abortion Clinic /Pills In Harare ,
sisternakatoto
 
Charaka Samhita Sutra sthana Chapter 15 Upakalpaniyaadhyaya
Charaka Samhita Sutra sthana Chapter 15 UpakalpaniyaadhyayaCharaka Samhita Sutra sthana Chapter 15 Upakalpaniyaadhyaya
Charaka Samhita Sutra sthana Chapter 15 Upakalpaniyaadhyaya
Dr KHALID B.M
 
NVBDCP.pptx Nation vector borne disease control program
NVBDCP.pptx Nation vector borne disease control programNVBDCP.pptx Nation vector borne disease control program
NVBDCP.pptx Nation vector borne disease control program
Sapna Thakur
 
New Directions in Targeted Therapeutic Approaches for Older Adults With Mantl...
New Directions in Targeted Therapeutic Approaches for Older Adults With Mantl...New Directions in Targeted Therapeutic Approaches for Older Adults With Mantl...
New Directions in Targeted Therapeutic Approaches for Older Adults With Mantl...
i3 Health
 
Physiology of Special Chemical Sensation of Taste
Physiology of Special Chemical Sensation of TastePhysiology of Special Chemical Sensation of Taste
Physiology of Special Chemical Sensation of Taste
MedicoseAcademics
 
HOT NEW PRODUCT! BIG SALES FAST SHIPPING NOW FROM CHINA!! EU KU DB BK substit...
HOT NEW PRODUCT! BIG SALES FAST SHIPPING NOW FROM CHINA!! EU KU DB BK substit...HOT NEW PRODUCT! BIG SALES FAST SHIPPING NOW FROM CHINA!! EU KU DB BK substit...
HOT NEW PRODUCT! BIG SALES FAST SHIPPING NOW FROM CHINA!! EU KU DB BK substit...
GL Anaacs
 
Superficial & Deep Fascia of the NECK.pptx
Superficial & Deep Fascia of the NECK.pptxSuperficial & Deep Fascia of the NECK.pptx
Superficial & Deep Fascia of the NECK.pptx
Dr. Rabia Inam Gandapore
 

Recently uploaded (20)

heat stroke and heat exhaustion in children
heat stroke and heat exhaustion in childrenheat stroke and heat exhaustion in children
heat stroke and heat exhaustion in children
 
micro teaching on communication m.sc nursing.pdf
micro teaching on communication m.sc nursing.pdfmicro teaching on communication m.sc nursing.pdf
micro teaching on communication m.sc nursing.pdf
 
Flu Vaccine Alert in Bangalore Karnataka
Flu Vaccine Alert in Bangalore KarnatakaFlu Vaccine Alert in Bangalore Karnataka
Flu Vaccine Alert in Bangalore Karnataka
 
Tom Selleck Health: A Comprehensive Look at the Iconic Actor’s Wellness Journey
Tom Selleck Health: A Comprehensive Look at the Iconic Actor’s Wellness JourneyTom Selleck Health: A Comprehensive Look at the Iconic Actor’s Wellness Journey
Tom Selleck Health: A Comprehensive Look at the Iconic Actor’s Wellness Journey
 
Non-respiratory Functions of the Lungs.pdf
Non-respiratory Functions of the Lungs.pdfNon-respiratory Functions of the Lungs.pdf
Non-respiratory Functions of the Lungs.pdf
 
Are There Any Natural Remedies To Treat Syphilis.pdf
Are There Any Natural Remedies To Treat Syphilis.pdfAre There Any Natural Remedies To Treat Syphilis.pdf
Are There Any Natural Remedies To Treat Syphilis.pdf
 
New Drug Discovery and Development .....
New Drug Discovery and Development .....New Drug Discovery and Development .....
New Drug Discovery and Development .....
 
Triangles of Neck and Clinical Correlation by Dr. RIG.pptx
Triangles of Neck and Clinical Correlation by Dr. RIG.pptxTriangles of Neck and Clinical Correlation by Dr. RIG.pptx
Triangles of Neck and Clinical Correlation by Dr. RIG.pptx
 
Evaluation of antidepressant activity of clitoris ternatea in animals
Evaluation of antidepressant activity of clitoris ternatea in animalsEvaluation of antidepressant activity of clitoris ternatea in animals
Evaluation of antidepressant activity of clitoris ternatea in animals
 
For Better Surat #ℂall #Girl Service ❤85270-49040❤ Surat #ℂall #Girls
For Better Surat #ℂall #Girl Service ❤85270-49040❤ Surat #ℂall #GirlsFor Better Surat #ℂall #Girl Service ❤85270-49040❤ Surat #ℂall #Girls
For Better Surat #ℂall #Girl Service ❤85270-49040❤ Surat #ℂall #Girls
 
Maxilla, Mandible & Hyoid Bone & Clinical Correlations by Dr. RIG.pptx
Maxilla, Mandible & Hyoid Bone & Clinical Correlations by Dr. RIG.pptxMaxilla, Mandible & Hyoid Bone & Clinical Correlations by Dr. RIG.pptx
Maxilla, Mandible & Hyoid Bone & Clinical Correlations by Dr. RIG.pptx
 
How to Give Better Lectures: Some Tips for Doctors
How to Give Better Lectures: Some Tips for DoctorsHow to Give Better Lectures: Some Tips for Doctors
How to Give Better Lectures: Some Tips for Doctors
 
Phone Us ❤85270-49040❤ #ℂall #gIRLS In Surat By Surat @ℂall @Girls Hotel With...
Phone Us ❤85270-49040❤ #ℂall #gIRLS In Surat By Surat @ℂall @Girls Hotel With...Phone Us ❤85270-49040❤ #ℂall #gIRLS In Surat By Surat @ℂall @Girls Hotel With...
Phone Us ❤85270-49040❤ #ℂall #gIRLS In Surat By Surat @ℂall @Girls Hotel With...
 
263778731218 Abortion Clinic /Pills In Harare ,
263778731218 Abortion Clinic /Pills In Harare ,263778731218 Abortion Clinic /Pills In Harare ,
263778731218 Abortion Clinic /Pills In Harare ,
 
Charaka Samhita Sutra sthana Chapter 15 Upakalpaniyaadhyaya
Charaka Samhita Sutra sthana Chapter 15 UpakalpaniyaadhyayaCharaka Samhita Sutra sthana Chapter 15 Upakalpaniyaadhyaya
Charaka Samhita Sutra sthana Chapter 15 Upakalpaniyaadhyaya
 
NVBDCP.pptx Nation vector borne disease control program
NVBDCP.pptx Nation vector borne disease control programNVBDCP.pptx Nation vector borne disease control program
NVBDCP.pptx Nation vector borne disease control program
 
New Directions in Targeted Therapeutic Approaches for Older Adults With Mantl...
New Directions in Targeted Therapeutic Approaches for Older Adults With Mantl...New Directions in Targeted Therapeutic Approaches for Older Adults With Mantl...
New Directions in Targeted Therapeutic Approaches for Older Adults With Mantl...
 
Physiology of Special Chemical Sensation of Taste
Physiology of Special Chemical Sensation of TastePhysiology of Special Chemical Sensation of Taste
Physiology of Special Chemical Sensation of Taste
 
HOT NEW PRODUCT! BIG SALES FAST SHIPPING NOW FROM CHINA!! EU KU DB BK substit...
HOT NEW PRODUCT! BIG SALES FAST SHIPPING NOW FROM CHINA!! EU KU DB BK substit...HOT NEW PRODUCT! BIG SALES FAST SHIPPING NOW FROM CHINA!! EU KU DB BK substit...
HOT NEW PRODUCT! BIG SALES FAST SHIPPING NOW FROM CHINA!! EU KU DB BK substit...
 
Superficial & Deep Fascia of the NECK.pptx
Superficial & Deep Fascia of the NECK.pptxSuperficial & Deep Fascia of the NECK.pptx
Superficial & Deep Fascia of the NECK.pptx
 

Basic deep learning & Deep learning application to medicine

  • 1. Deep learning application to medicine Department of Nuclear Medicine, Seoul National University Hospital Hongyoon Choi
  • 2. CONTENTS • Basic knowledge : Linear regression to deep learning • Overview of deep learning • Real application to medical data
  • 3. B a s i c k n o w l e d ge : L i n e a r re g re s s i o n to d e e p l e a r n i n g
  • 4. L i n e a r re g re s s i o n to d e e p l e a r n i n g Age prediction No. of wrinkles Age Regression: Predict continuous value Supervised Learning: “Right answers” given
  • 5. L i n e a r re g re s s i o n to d e e p l e a r n i n g Lung tumor (Benign vs malignancy) Tumor Size Classification: Predict discrete value Supervised Learning: “Right answers” given 0: Benign 1: Malignant
  • 6. L i n e a r re g re s s i o n to d e e p l e a r n i n g Supervised Learning Unsupervised Learning Regression Classification Clustering Generative model Semi-Supervised Learning Reinforcement Learning
  • 7. L i n e a r re g re s s i o n to d e e p l e a r n i n g Training Set Learning Algorithm hHeight Estimated Weight Training set Height(X) Weight(Y) 170 155 180 175 190 160 . . . 71 49 80 63 91 52 . . . h(X) = a0 + a1X
  • 8. L i n e a r re g re s s i o n to d e e p l e a r n i n g h(X) = a0 + a1X a  Parameters How to choose “a”? Basic idea: Input a0 and a1 to minimize “Some Target” Some Target = Cost Function J(a0, a1) = Mean of (h(X) – Y)2
  • 9. L i n e a r re g re s s i o n to d e e p l e a r n i n g 0 1 2 3 0 1 2 3 Simply, a0 = 0 and only consider a1 0 2 4 6 -0.5 0 0.5 1 1.5 2 2.5 X Y a1 J(a1)
  • 10. L i n e a r re g re s s i o n to d e e p l e a r n i n g h(X) = a0 + a1X J(a0, a1) = Mean of (h(X) – Y)2 Hypothesis Cost Function Target Minimize Cost Parameters a0 and a1
  • 11. L i n e a r re g re s s i o n to d e e p l e a r n i n g a0 a1 J
  • 12. L i n e a r re g re s s i o n to d e e p l e a r n i n g Height Weight a0 a1 Optimal points of a0 and a1
  • 13. L i n e a r re g re s s i o n to d e e p l e a r n i n g How to optimize? Gradient Descent • Basic algorithm for training deep learning model • Briefly, • Start any points of a0 and a1 • Iteratively changing a0 and a1 until reach minimum J(a0,a1)
  • 14. L i n e a r re g re s s i o n to d e e p l e a r n i n g a0 a1 J
  • 15. L i n e a r re g re s s i o n to d e e p l e a r n i n g aj  aj - 𝛂 𝝏 𝝏𝒂𝒋 𝑱(𝒂 𝟎, 𝒂 𝟏) (j= 0 and 1) Repeat until convergence Gradient Descent Algorithm 𝛂: Learning Rate a1 0 2 4 6 -0.5 0 0.5 1 1.5 2 2.5 J(a1) Positive slope
  • 16. L i n e a r re g re s s i o n to d e e p l e a r n i n g a1 0 2 4 6 -0.5 0 0.5 1 1.5 2 2.5 J(a1) a1 0 2 4 6 -0.5 0 0.5 1 1.5 2 2.5 J(a1) Small 𝛂 large 𝛂
  • 17. L i n e a r re g re s s i o n to d e e p l e a r n i n g • Multiple Variables h(X) = a0 + a1X1 + a2X2+a3X3 … Weight Prediction using Height, Waist circumference, Head circumference, … J(ai) = Mean of (h(X) – Y)2 aj  aj - 𝛂 𝝏 𝝏𝒂𝒋 𝑱(𝒂)Repeat
  • 18. L i n e a r re g re s s i o n to d e e p l e a r n i n g 0 2 4 6 -0.5 0 0.5 1 1.5 2 2.5 Single Parameter Two Parameter >3 Parameters : n-dimensional parabolic shape
  • 19. L i n e a r re g re s s i o n to d e e p l e a r n i n g Lung tumor (Benign vs malignancy) Tumor Size 0: Benign 1: Malignant h(x) = aX h(X) = 1 if X > 3cm h(X) = 0 if X < 3cm Classification
  • 20. L i n e a r re g re s s i o n to d e e p l e a r n i n g Lung tumor (Benign vs malignancy) Tumor Size 0: Benign 1: Malignant Classification Borderline  value of 0~1 “Logistic Function” =“Sigmoid Function” What we want… 1) 0< h(X) <1 2) For borderline, h(X) ~ 0.5 ℎ 𝑋 = 1 1 + 𝑒−𝑍 Z = a0 + a1X (X: tumor size)
  • 21. L i n e a r re g re s s i o n to d e e p l e a r n i n g Lung tumor (Benign vs malignancy) Tumor Size 0: Benign 1: Malignant Classification Decision Boundary ℎ 𝑋 = 1 1 + 𝑒−𝑍 Z = a0 + a1X (X: tumor size)0.5 3 cm H(X) > 0.5  Malignant H(X) < 0.5  Benign Interpretation of h(X) ~ Probability of malignant
  • 22. L i n e a r re g re s s i o n to d e e p l e a r n i n g Classification x1 x2 1 2 3 1 2 3 Classification with two variables x1 x2 1 2 3 1 2 3 Decision boundary = Threshold x1 x2 1 2 3 1 2 3Y=1 Y=0 Sigmoid
  • 23. L i n e a r re g re s s i o n to d e e p l e a r n i n g Classification Classification with multiple variables x1 x2 1 2 3 1 2 3 ℎ 𝑋 = 1 1 + 𝑒−𝑍 Z = a0 + a1X1+a2X2 Linear regression h(X) = a0+a1x1+a2x2+… Logistic classification h(X) = sig(Z) , Z = a0+a1x1+a2x2+…
  • 24. L i n e a r re g re s s i o n to d e e p l e a r n i n g Classification Tumor Size 0: Benign 1: Malignant Tumor Size 0: Benign 1: Malignant ℎ 𝑋 = 1 1 + 𝑒−𝑍 Z = a0 + a1X Changing a0 Changing a1 How to optimize aj?  Make convex cost function J(a0, a1) = Mean of (h(X) – Y)2 For linear regression, 0 2 4 6 -0.5 0 0.5 1 1.5 2 2.5
  • 25. L i n e a r re g re s s i o n to d e e p l e a r n i n g Classification Cost function for logistic classification J(a0, a1) = -log(h(X)) if Y = 1 -log(1-h(X)) if Y = 0 h(X) J 10 (h(X) is 0~1) Y=1 h(X) J 10 (h(X) is 0~1) Y=0
  • 26. L i n e a r re g re s s i o n to d e e p l e a r n i n g Classification Cost function for logistic classification J(a0, a1) = -[Y log (h(X)) + (1-Y) log (1-h(X))] Cost function for logistic = “Binary crossentropy” aj  aj - 𝛂 𝝏 𝝏𝒂𝒋 𝑱(𝒂)Repeat Optimization algorithm : Same with linear regression
  • 27. L i n e a r re g re s s i o n to d e e p l e a r n i n g • Logistic regression as a Perceptron Lesion size Circularity Hounsfield unit x1 x2 x3 Z w1 w2 w3 b0 Activation function Sigmoid(Z) Output 1: Malignancy 0: Benign Find optimized W for minimized error  Gradient descent
  • 28. L i n e a r re g re s s i o n to d e e p l e a r n i n g • Perceptron vs neuron
  • 29. L i n e a r re g re s s i o n to d e e p l e a r n i n g x1 x2 x1 x2 1 2 3 1 2 3 Linear classification Nonlinear classification • Limitation of single-layer perceptron
  • 30. L i n e a r re g re s s i o n to d e e p l e a r n i n g Layer 3Layer 1 Layer 2 H(X)input Output layer a1 (2) = g(w11x1+w12x2+w13x3) a2 (2) = g(w21x1+w22x2+w23x3) a3 (2) = g(w31x1+w32x2+w33x3) where g: activation function (sigmoid) a1 (3) = H(X) = g(w21a1 (2)+w22a2 (2)+w23a3 (2)) Single-layer perceptron to neural network
  • 31. L i n e a r re g re s s i o n to d e e p l e a r n i n g Non-linear classification example: XOR/XNOR x1 and x2 are binary (0 or 1). x1 x2 Y=1 Y=0 (XNOR Problem)
  • 32. L i n e a r re g re s s i o n to d e e p l e a r n i n g Non-linear classification example: XOR AND X2 X1 +1 -30 20 20 H(X) X2 X1 +1 -10 20 20 H(X) OR sigmoid(-10)~0 sigmoid(10)~1
  • 33. L i n e a r re g re s s i o n to d e e p l e a r n i n g Non-linear classification example: XOR AND X2 X1 +1 a1 -30 20 20 (NOT x1) AND (NOT x2) a2 10 -20 -20 +1 H(X) -10 20 20 OR X1 X2 0 0 0 1 1 0 1 1 a1 a2 0 1 0 0 0 0 1 0 1 0 0 1 H(X)
  • 34. L i n e a r re g re s s i o n to d e e p l e a r n i n g Output Lesion size Circularity Lesion size Circularity Malignancy Benign Non-linear classification
  • 35. L i n e a r re g re s s i o n to d e e p l e a r n i n g H O Output How to calculate gradient descent ? 𝜕𝐽 𝜕𝑤2 = 𝜕𝐽 𝜕𝑂 𝜕𝑂 𝜕𝑍 𝑂 𝜕𝑍 𝑂 𝜕𝑤2 w2 𝑍 𝑜 = ෍ 𝑊2 𝐻 J = Cost (O, Y) (O = h(X)) J= -[Y log (h(X)) + (1-Y) log (1-h(X))] w1 𝜕𝐽 𝜕𝑤1 = 𝜕𝐽 𝜕𝑍ℎ 𝜕𝑍ℎ 𝜕𝑤1 H=sig(Zh) O=sig(ZO) Zh=W1X ZO=W2H “Back Propagation” 𝜕𝐽 𝜕𝑍ℎ = 𝜕𝐽 𝜕𝑍 𝑂 𝜕𝑍 𝑂 𝜕𝐻 𝜕𝐻 𝜕𝑍ℎ
  • 36. L i n e a r re g re s s i o n to d e e p l e a r n i n g The era of artificial brain!
  • 37. L i n e a r re g re s s i o n to d e e p l e a r n i n g You see this: But the camera sees this: • Limitation of conventional neural network
  • 38. L i n e a r re g re s s i o n to d e e p l e a r n i n g • Limitation of conventional neural network pixel 1 pixel 2 Cars “Non”-Cars 50 x 50 pixel images→ 2500 pixels (7500 if RGB) pixel 1 pixel 2 = A point at 7500 dimensional axis  7500 multivariable logistic regression
  • 39. L i n e a r re g re s s i o n to d e e p l e a r n i n g Curse of dimensionality
  • 40. L i n e a r re g re s s i o n to d e e p l e a r n i n g Curse of dimensionality
  • 41. L i n e a r re g re s s i o n to d e e p l e a r n i n g More dimensions More sparse in data space  Easy to overfit
  • 42. L i n e a r re g re s s i o n to d e e p l e a r n i n g More dimensions  More weights (Parameters to learn) 1990~2000 Better manual features instead of raw pixel value + Kernel-based learning
  • 43. L i n e a r re g re s s i o n to d e e p l e a r n i n g • Hard to learn deep layer Problem of Vanishing Gradient
  • 44. L i n e a r re g re s s i o n to d e e p l e a r n i n g • Hard to learn deep layer Slope~0 Slope~0 Output Error Backpropagation 𝜕𝐽 𝜕𝑤2 = 𝜕𝐽 𝜕𝑂 𝜕𝑂 𝜕𝑍 𝑂 𝜕𝑍 𝑂 𝜕𝑤2 O=sig(ZO) ~0
  • 45. L i n e a r re g re s s i o n to d e e p l e a r n i n g Overcome Limitations & To Deep Learning… • Automatic feature extraction from raw data • Learning deep-layered neural network Algorithm Hardware Big Data
  • 46. Overview of Deep Learning
  • 47. Deep learning • Nonlinear problem & multilayer perceptron 0 1 1 3 5 7 8 8 0 0 0 1 3 3 5 3 0 0 1 2 4 7 7 1 0 0 2 3 8 5 7 6 2 5 8 8 8 4 9 5 0 0 8 8 6 4 2 3 128x128 =16,384 Output - Require good manual features - Raw data  Too big. - More layers?  Difficulty in learning
  • 48. Deep learning • MLP to Deep learning - Require good manual features - Raw data  Too big. - More layers?  Difficulty in learning • Automatic feature extraction from raw data • New activation function & Stochastic gradient descent • Methods for reducing overfitting
  • 49. Deep learning • Train deep layer Problem of Vanishing Gradient Slope~0 Slope~0
  • 50. Deep learning • Train deep layer Nonlinearity function Sigmoid  ReLU , tanh, ELU, Leaky ReLU Sigmoid ReLU Slope~0 Slope~0 Output Error Backpropagation Slope = 1
  • 51. Deep learning • Train big data aj  aj - 𝛂 𝝏 𝝏𝒂𝒋 𝑱(𝒂)RepeatGradient Descent 50x50x3 ~7500 pixel data per image 100,000 images of cars and non-cars Front : - Estimate 100,000 h(X) Back = update ‘a’ - Cost calculated by 100,000 cost(Y,h(X)) a1 0 2 4 6 -0.5 0 0.5 1 1.5 2 2.5 J(a1) Small 𝛂 Per iteration
  • 52. Deep learning • Train big data Training data 1 weight update Training data 1 weight update per mini-batch Multiple weight updates Gradient Descent Stochastic Gradient Descent (=batch gradient) (=minibatch stochastic gradient)
  • 53. Deep learning • Train big data Faster and efficient to reach global minima
  • 54. Deep learning • Problem of overfitting Apple Not apple Because it’s yellow Not apple Because it’s not round
  • 55. Deep learning • Problem of overfitting Dropout
  • 56. Deep learning • Automatic feature extraction efficiently – Convolutional Neural Network for Image – Recurrent Neural Network for Sequential Data
  • 57. Deep learning • Convolutional Neural Network • Local feature extraction • Translational invariance • Sparsity compared with fully-connected layer
  • 58. Deep learning • Convolutional Neural Network • Local feature extraction • Translational invariance • Sparsity compared with fully-connected layer Fully-connected layer Convolutional layer
  • 59. Deep learning • Convolutional Neural Network Convolutional filter 0 1 1 3 5 7 8 8 0 0 0 1 3 3 5 3 0 0 1 2 4 7 7 1 0 0 2 3 8 5 7 6 2 5 8 8 8 4 9 5 0 0 8 8 6 4 2 3 Image Filter 1 0 1 0 1 0 1 0 1 2 6 12 22 27 28 2 5 15 16 30 24 11 17 24 29 33 24 15 19 32 28 27 27 Output Number of feature maps = number of convolutional filters Instead number of nodes
  • 60. Deep learning identify line / some texture identify head lights and wheels identify Car!
  • 62. Deep learning • Convolutional Neural Network Initial Data : 256 x 256 Ear Ear Eye Eye Nose Tail Foot Foot After convolutions and poolings =Abstracted features
  • 63. Deep learning • Convolutional Neural Network Abstracted Features Feature vectors Multivariate Logistic Dimension 256x256x3 4x4x1024 4096 1000
  • 64. Deep learning • Convolutional Neural Network ImageNet Challenge Results 28.2% 2010 25.8% 2011 16.4% 2012 Shallow model AlexNet 11.7% 2013 6.7% 2014 3.57% 2015 GoogleNet ResNet 8-layers 22-layers 152-layers
  • 65. Deep learning • Convolutional Neural Network ImageNet Challenge Results AlexNet GoogleNet ResNet
  • 66. Deep learning Pedestrian Car Motorcycle Truck • Cf> Multiple output (instead of binary classification) 4 Output nodes, instead of 1 node Y = [ 1, 0 , 0, 0] for pedestrian Y = [ 0, 1, 0, 0 ] for car Y = [ 0, 0, 1, 0 ] for motorcycle Y = [ 0, 0, 0, 1 ] for truck Activation function: Softmax, instead of sigmoid
  • 67. Deep learning • Recurrent Neural Network
  • 68. Deep learning • Recurrent Neural Network
  • 69. Deep learning • Recurrent Neural Network Vision Deep CNN Language Generating RNN “A group of people shopping at an outdoor market. There are many vegetables at the fruit stands.” Neural Machine Translation Google Translate Text, Music Generation https://www.youtube.com/watch?v=A2gyidoFsoI Combined with CNN : Image caption generation
  • 70. Deep learning • Current Concept of Deep learning Deep layered neural network Output + Data type-specific layers Convolution Recurrent Modification for training + ReLU activation SGD training Dropout Batch normalization Variable Cost Function …
  • 71. Deep learning Nomenclature Supervised Learning Unsupervised Learning Reinforcement Learning • Regression • Classification • Clustering • Generative model • Algorithms react to environment Machine Learning
  • 72. Deep learning Unsupervised learning , particularly generative model Transfer learning One-shot learning Bayesian modeling Mobile-friendly model Manifold and non- Euclidean data Current trends of deep learning High accuracy to various purposes/situations
  • 73. Deep learning Flexible and scalable deep learning • Transfer learning Car ImageNet-based model as a feature extractor Train only last layer
  • 74. Deep learning Flexible and scalable deep learning • Generative model z~N(0,1) G: generator Fake image Real image D: Discriminator 1: real 0: fake Can be composed with convolutional, FC layers, Batch normalization, regularization, etc. Various cost functions / combined cost functions MSE, CE, Adversarial, KLD, etc.
  • 75. Deep learning Flexible and scalable deep learning Ref. DeepMind, NIPS 2017 https://tykimos.github.io/
  • 76. Deep learning in Medicine
  • 77. Deep learning in Medicine ROC curve - better than dermatologists Esteva, Andre, et al. Nature 2017
  • 78. Deep learning in Medicine Diabetic Retinopathy  Better or equivalent to ophthalmologists Normal DM Gulshan, Varun, et al. JAMA 2016 ChestXnet  Equivalent/Superior to radiologists (?) Pranav Rajpurkar, … Andrew Ng, 2017. Arxiv DL for medical imaging: Supervised learning using CNN
  • 79. Deep learning in Medicine FDA approve a device for diagnosing diabetic retinopathy (2018.4) AI-aided system (CT angiography for large vessel occlusion) Year of AI invasion to clinic
  • 80. Deep learning in Medicine DL for medical imaging: Supervised learning using CNN AD & NC MCI-converter & non-converter FDG and amyloid PET to predict future cognitive decline Choi H and Jin KH Arxiv 2017
  • 81. Deep learning in Medicine https://adfdgpet.appspot.com Online Demo Input file Web application Output: likelihood for AD & predicted cognitive score Output: Cognitive dysfunction-related map p(Alzheimer|X)
  • 82. Deep learning in Medicine https://insight.lunit.io/ Web application
  • 83. Deep learning in Medicine Web application https://modelderm.com Han SS, et al. J Invest Derm 2018
  • 84. Deep learning in Medicine Laborious Work Replaced by DL: Segmentation Choi, H., & Jin, K. H. J Neurosci Methods 2016 de Brebisson, et al. CVPR 2015.
  • 85. Deep learning in Medicine Laborious Work Replaced by DL: Detection Liu Y, et al. Arxiv 2017
  • 86. Deep learning in Medicine Enhance Image Acquisition & Quality Dahl et al. Arxiv 2017 Normal dose abdomen CT Low dose abdomen CT Low dose abdomen CT+CNN Chen H et al. Biomed Opt Exp 2017
  • 87. Deep learning in Medicine Image Generation Cat Cat Common deep learning model Generative model z = f(x) where x: data, z: discriminative features f: classifier model x = g(z) where x: data, z: latent g: generation function
  • 88. Deep learning in Medicine Generative Adversarial Network z~N(0,1) G: generator Karras T, et al. Arxiv 2017 G: generator Isola P, et al. arxiv, 2016.
  • 89. Generative Adversarial Network Structural MR generation from PET Florbetapir PET Generator: U-net Skip connection Generated MR PETandgeneratedMRPETandrealMR Discriminator Real or Fake Generative Adversarial Networks for MR generation z G(z) z & G(z) z & x Choi H and Lee DS, J Nucl Med 2017. Deep learning in Medicine RealMRI Generated MRI 18F-Florbetapir PET
  • 90. Deep learning in Medicine Conditional Generation Antipov G, Arxiv 2017
  • 91. Encoder Latent features Generator + Age Latent features + Age VAE model for brain PET generation  Brain metabolism aging movie Choi H,… Lee DS. Biorxiv 2017 Deep learning in Medicine Conditional Generation
  • 92. Choi H,… Lee DS. Biorxiv 2017 Estimating normal population distribution Deep learning in Medicine
  • 93. Deep learning in Medicine Omics data Quang D et al. NAR 2016 Predicting Function from DNA sequences
  • 94. Deep learning in Medicine Omics data Low risk group High risk group Deep learning-based risk score Choi H and Na KJ, Biomed Res Int. 2017
  • 95. Deep learning in Medicine Disruptive Innovation: Raw medical & healthcare data Diet + Previous Glucose Level Future Glucose Level & Scheduling Insulin Sugar.iq from Medtronic
  • 96. Deep learning in Medicine Disruptive Innovation: Raw medical & healthcare data
  • 97. Deep learning in Medicine Disruptive Innovation: Raw medical & healthcare data HTC DeepQ Tricoder Predicting PVC from daily EKG Diagnosis of otitis media
  • 98. Deep learning in Medicine Disruptive Innovation: Raw medical & healthcare data Deep learning facilitates left-shifting