Recent advances of AI for medical imaging : Engineering perspectives
1. Deep learning and its application for radiologists
Namkug Kim, PhD namkugkim@gmail.com
Medical Imaging N Robotics Lab. http://mirl.ulsan.ac.kr
Convergence Medicine/Radiology
Biomedical Engineering Center
Asan Medical Center/Univ. of Ulsan College of Medicine
2. Researches with
Hyundai Heavy Industries Co. Ltd.
LG Electronics
Coreline Soft Inc.
Osstem Implant
CGBio
VUNO
Kakaobrain
이해상충
Stockholder
Coreline Soft, Inc.
AnyMedi
Co-Founder
Somansa Inc.
Cybermed Inc.
Clinical Imaging Solution, Inc
AnyMedi, Inc.
Selected Grants as PI
한국연구재단, NRF, South Korea
7T용 4D 자기공명유속영상을 이용한 심뇌혈관 질환의
in-vivo 유동 정량화 SW개발, 2016
4D flow MRI을 이용한 심혈관 질환의 in-vivo 유동 연구,
2015-7
자기공명분광영상 및 MRI의 통합 분석 소프트웨어
개발
산업부, KEIT, South Korea
의료영상 인공지능 과제, 2016-20
3DP 척추 맞춤형 임플란트, 2016-20
3D 프린터 기반 무치악 및 두개악안면결손 환자용 수복
보철물 제작, 재건 시스템 개발, 2015-9
근골격계 복구 수술 로봇 개발, 2012-7
영상중재시술 로봇시스템 개발, 2012-7
보건복지부, KHIDI, South Korea
영상 뇌졸중 예후 예측 및 치료방침 결정 시스템 개발,
2012-8
관동맥 관류 CT 의 자동 진단 프로그램을 활용한
허혈성 질환의 진단과 치료, 2013-6
산학협력
Hyundai Heavy Industry, Osstem Implant, S&G Biotech,
Coreline soft, Midas IT, AnyMedi, Hitachi Medical, Japan,
3. 5
MBC 다큐 스페셜 ‘미래인간 AI’, 2016.12.05
영상유도 중재 수술로봇 ; 현대중공업 폐영상 분석 SW ; 코어라인 소프트
3D 프린터 응용 ; 서울아산병원
Movie Clips
4. Major Breakthroughs in Feedforward NN
K. Fukushima Yann Lecun G. Hinton, S. Ruslan
Neocognitron
Neocognitron (1979)
• By Kunihiko Fukushima
• First proposed CNN
Convolutional Neural Networks (1989)
• Yann Lecun et.al
• Back propagation for CNN
• Theoretically learn any function
LeNet-5 architecture
Alex krizhevsky ,
Hinton
LeNet-5 (1998)
• Convolutional networks
Improved by Yann Lecun
et.al
• Classify handwritten digits
D. Rumelhart, G. Hinton, R.
Wiliams
1960 1970 1980 1990 2000 2010 2012
Perceptron
XOR
Problem
Golden
Age
1957 1969 1986
F. Rosenblatt
M. Minsky, S. Papert
• Adjustable weights
• Weights are not
learned
• XOR problem is not linearly
separble
• Solution to nonlinearity
separable problems
• Big computation, local
optima and overfiting
CNN Breakthrough (2012)
• By Alex Krizhevsky et al.
• Winner of ILSVRC2012 by
large marginDark Age (AI
winter)
Back propagation
(1981)
• Train multiple layers
Multi-layer
Perceptron (1986)
5. Why Deep Learning ?
From MS speech group
Speech Recognition
6.66%
5.98%
5.10% 4.94% 4.80%
0.00%
1.00%
2.00%
3.00%
4.00%
5.00%
6.00%
7.00%
Google 2014
Baidu 2015
Human Level
MS 2015
Google 2015
Object (Image) Recognition (ImageNet)
Face Verification Gene Network Structure Inference
97.35% 97.45% 97.53%
98.52%
99.15%
99.47%
99.63%
96.00%
96.50%
97.00%
97.50%
98.00%
98.50%
99.00%
99.50%
100.00%
←Using DL
6. Major Components in Deep Learning
Breakthroughs
Algorithms
Parallel
computing
Big Data
Unsupervised pre-training
Supervised training for deeper models
NVIDIA® CUDA® 코어 5760
메모리 클럭 7.0 Gbp
표준 메모리 설정 12288 MB
7. Where do we use Deep Learning ?
Auto drive (pedestrian/traffic sign recognition)
• Netflix movie recommendation
• Language translation
• Breast cancer detection
• Skype translator
• Gesture/pose detection*
*Neverova, Natalia, et al. "Hand Pose Estimation through Weakly-Supervised Learning of a Rich
Intermediate Representation." arXiv preprint arXiv:1511.06728 (2015)
14. 생성 모델 (풍경)
Eyescream project in Facebook AI Research
(using Laplacian pyramid Generative adversarial
network; LAPGAN)
http://soumith.ch/eyescream/
15. 생성 모델 (얼굴, 거실)
Deep Convolutional Generative Adversarial Networks (DCGAN)
Rotations are linear in latent space
Bedroom generation
Arithmetic on faces
19. Bio Plausible Neural Network
Mimic human visual recognition system
Each unit connected to a small subset of other units
Based on what it sees, it decides what it wants to say
Units must learn to cooperate to accomplish the task
38
22. From Shallow to Deep Learning
Shallow learning
SVM
Linear & Kernel Regression
Hidden Markov Models (HMM)
Gaussian Mixture Models (GMM)
Single hidden layer MLP
Artificial Neural Net (ANN)
...
Limited modeling capability of concepts
Cannot make use of unlabeled data
24. Neural Networks
• Machine Learning
• Knowledge from high dimensional data
• Classification
• Input: features of data
• supervised vs unsupervised
• labeled data
• Neurons
25. Neural Networks (NN); 반복적인 에러교정
Forward propagation
Sum inputs, produce activation, feed-forward
Input 𝑋
Hidden
𝑊1
𝑊2 Neuron
y
𝑥1
𝑥2
𝑥3
𝑥 𝑛−1
𝑥 𝑛
…
Inputs
Output
𝑧 = 𝑏 +
𝑖
𝑥𝑖 𝑤𝑖
𝑦 = 𝐻(𝑧)
Output 𝑌 =f(x)
Scaling function Activation function Activation function
Information Propagation
Error BackPropagation
Output Comparison
Weights
29. Need for Multiple Units and Multiple Layers
Multiple boundaries
are needed (e.g.
XOR problem) ->
Multiple Units
More complex
regions are needed
(e.g. Polygons) ->
Multiple Layers
48
31. Best Practice
Normalization
Prevent very high weights, Oscillation
Overfitting/Generalisation
Validation Set, Early Stopping
Mini-Batch Learning
update weights with multiple input vectors combined
32. Challenges in Training Feedforward NN
Training multiple layers
Vanishing gradient problem
Gradients are diluted as layers go deep
Only use labeled data
most data is unlabeled
Stuck in local minima
Over-fitting
Too many hyper parameters
33. Problems with Backpropagation
Limitations
Get stuck in local optima
start weights from random positions
Error attenuation, long fruitless training
Slow convergence to optimum
large training set needed
Only use labeled data
most data is unlabeled
Backpropagation (BP) barely changes lower-layer
parameters (vanishing gradient)
Therefore, deep networks cannot be fully (effectively) trained
with backpropagation
52
34. Breakthrough with Backpropagation
Breakthrough
Recent – Long patient training with GPUs and special hardware
Deep belief networks (unsupervised pre-training)
Convolutional neural networks (reducing redundant parameters)
Rectified linear unit (constant gradient propagation)
53
35. Rectified Linear Units
More efficient gradient propagation, derivative is 0 or constant,
just fold into learning rate
More efficient computation: Only comparison, addition and mul
tiplication.
Leaky ReLU f(x) = x if > 0 else ax where 0 ≤ a <= 1, so that derivat
e is not 0 and can do some learning for that case.
Lots of other variations
Sparse activation: For example, in a randomly initialized network
s, only about 50% of hidden units are activated (having a non-z
ero output)
CS 678 – Deep Learning 54
36. Convolutional Neural Networks (CNN)
A type of feed-forward neural network
Inspired by biological process
Weight sharing (convolution) +
Subsampling (pooling)
Reducing the number of parameters (Reduce over-
fitting)
Translation invariance
Input
28 × 28
Feature maps
4@24 × 24
Feature maps
4@8 × 8
Feature maps
8@4 × 4
Feature maps
8@2 × 2
Feature maps
8 ⋅ 2 ⋅ 2 × 1
Output
10 × 1
Convolution
layer
Max-pooling
layer
Convolution
layer
Max-pooling
layer
Reshape Linear layer
[LeCun, 1998]
38. Convolutional Neural Networks (CNN)
Convolution과 Pooling (Subsampling)을 반복하여 상위 Feature
를 구성
Convolution은 Local영역에서의 특정 Feature를 얻는 과정
Pooling은 Dimension을 줄이면서도, Translation-invariant한
Feature를 얻는 과정
39. Convolutional Neural Networks (CNN)
Neural network with sparse connections
Learning algorithm:
Backpropagation on convolution layers and fully-connected layers
41. Visualization of FilterBank
VGG16 architecture
ImageNet
most filters
identical,
rotated by some non-random
factor (typically 90 degrees).
– rotation-invariant.
The rotation observation
holds true in block4_conv1.
Textures similar to that found
in the objects in
block5_conv2
61
43. 26.2%
16.4%
13.5%
12.9%
11.8%
7.3%
6.7%
4.9% 4.8%
3.6%
Breakthroughs in CNN
2012 2013 2014 2015
Alexnet (2012)
• 1st place in 2012
• 5 conv layers + 3
fully connected layers
• Dropout & ReLU
ImageNet Large Scale Visual Recognition Challenge
Results
SIFT + FVs (2012)
• 2nd place in 2012
• SIFT + fisher vectors
• No CNNs
ReLU (Rectified Linear Units)
Alexnet
architecture
Data augmentation (flip,
random crop)
Dropout
44. 26.2%
16.4%
13.5%
12.9%
11.8%
7.3%
6.7%
4.9% 4.8%
3.6%
Recent Breakthroughs in CNN
2012 2013 2014 2015
ZF Net (2013)
• 3rd place in 2013
• By Matthiew
Zeiler & Rob
Fergus
• Variant of
Alexnet
Alexnet (2012)
• 1st place in 2012
• 5 conv layers + 3
fully connected
layers
• Dropout & ReLU
Clarifai (2013)
• 1st place in 2013
• Deep Learning startup
• Founded by Matthiew
Zeiler
• Variant of Alexnet
VGG Networks (2014)
• 2nd place in 2014
• By Oxford computer vision
group
• 19 layers deep
GoogLeNet (2014)
• 1st place in 2014
• 24 layers of convolution
• Memory efficient network
ImageNet Large Scale Visual Recognition Challenge
Results
SIFT + FVs (2012)
• 2nd place in 2012
• SIFT + fisher vectors
• No CNNs
Overfeat (2013)
• 2nd place in 2013
• By NYU
• Variant of Alexnet
GoogLeNet
architecture
45. 26.2%
16.4%
13.5%
12.9%
11.8%
7.3%
6.7%
4.9% 4.8%
3.6%
Recent Breakthroughs in CNN
2012 2013 2014 2015
ZF Net (2013)
• 3rd place in 2013
• By Matthiew
Zeiler & Rob
Fergus
• Variant of
Alexnet
Alexnet (2012)
• 1st place in 2012
• 5 conv layers + 3
fully connected
layers
• Dropout & ReLU
Clarifai (2013)
• 1st place in 2013
• Deep Learning startup
• Founded by Matthiew
Zeiler
• Variant of Alexnet
VGG Networks (2014)
• 2nd place in 2014
• By Oxford computer vision
group
• 19 layers deep
GoogLeNet (2014)
• 1st place in 2014
• 24 layers of convolution
• Memory efficient network
Batch normalization (2015)
• By Google
• Simple but powerful
normalization algorithm
Parametric ReLU (2015)
• By Facebook
ImageNet Large Scale Visual Recognition Challenge
Results
SIFT + FVs (2012)
• 2nd place in 2012
• SIFT + fisher vectors
• No CNNs
Overfeat (2013)
• 2nd place in 2013
• By NYU
• Variant of Alexnet
Parametric ReLU
Residual Learning (add skip connections)
Deep Residual Network
(2016)
• Winner of ILSVRC 2015
• By MSRA
• More than 100 layers deep
with skip connections
52. Chest X-ray
75
Data cleansing SW
with AI
Gold Standard
(Generate 1000 data per class)
Manual Drawing SW
Nodule Interstitial Opacity
Consolidation Pleural Effusion
53. Anonymization
77
Data 수집
- 정상 1,821,455
- 비정상 287,626
DICOM header
Meta inf.
Chest (AP,Lateral)
Study description
filtering
Anonymizer
1. Batch conversion
2. 익명화
1. DICOM patient info 0008 tag
2. PatientID, SEX, AGE, NAME, Birth date, etc
3. Research ID generation
3. Efficient Error Handling
MIRL@AMC ananoymizer
Anonymizer@MIRL*
54. Cleansing
Data Description
- # 9,589 from 500,000
- Resize into 100x100
- Standardization
CNN
Training Data
(3,000/3,000 )
Validation Data
(1,000/1,000)
Test Data
(1000/589)
Normal Class
Abnormal Class
Model Precision Sensitivity Specificity
CNN 99.3% 99.8% 98.4%
CNNBN 100% 99.9% 100%
H. C. Shin,“Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset
Characteristics and Transfer Learning, “IEEE Transactions on Medical Imaging, 2016, pp. 1285-1298.
55. Chest X-ray CAM
Weakly Supervised Learning + Class Activation Map
- Resnet50 ILSVRC Pre-trained Model
58. Chest PA YOLO
• Fine-Tuned You Only Look Once(YOLO) model
• YOLO model-26 layers including 24 convolution layers and 2 fully
connected layers
• Only final layer training
• Only 6 classes of abnormal lesion, our last layer requires C = 6
60. • Nodule result ROI
1. Two nodules are detected in in Chest X-ray image
2. Nodule are only detected in Chest X-ray image
• Consolidation result ROI
1. Consolidation and pleural effusion are simultaneously detected in Chest X-ray image
2. Consolidation is only detected in Chest X-ray image
① ②
① ②
Examples
61. • Interstitial Opacity and Cardiomegaly result ROI
1. Two interstitial opacity are detected in Chest X-ray image.
2. Cardiomegaly is detected in Chest X-ray image .
3. Cardiomegaly and two pleural effusion are detected in Chest X-ray image.
②①
③
Examples
62. • Pleural Effusion result ROI
1. Cardiomegaly and Pleural Effusion are detected in Chest X-ray image
• Pneumothorax result ROI
1. Pneumothorax are detected in Chest X-ray image
2. Pneumothorax and pleural effusion are simultaneously detected in Chest X-ray image
①
②①
Examples
63. AI in Radiology
CAD for COPD
Compare various classifiers for OLD classification
Lee Y., et al, CMPB, 2009. 93(2): p. 206-15.
Adding shape features for accuracy enhancement of emphysema
quantification., J Digit Imaging, 22:136-148, 2009,
SPIE 2007 Honorable Mention Poster Award / JDI Best paper 2009
Compare texture-based quantification vs density-based quantification for
emphysema quantification, Investigative Radiology 43:395-402 , 2008
CAD for DILD
Study feasibility for DILD, Korean J Radiol 10:455-463, 2009
Context sensitive SVM for whole lung quantification, IFMIA 2011, IWPFI
2011, JDI 2011
CrossVendor study for ILD (GE vs Siemens), Kim N , et al, RSNA 2010, Med
Phys 2013
Texture based segmentation for iodine quantification in DECT of ILD, RSNA
2013, ER 2015
3D extension of texture analysis in DILD, RSNA 2014
CADD for DILD with regional parenchyma classification, WIP in Med Phys
Deep learning for DILD, WIP in JDI
DILD lung segmentation using deep learning, WIP in Med Phys
Noise reduction of HRCT using auto-encoder, WIP in Med Phys
Ensemble method for ILD classifier, JDI accepted
CAD for AVUS (Ultrasounds), Thyroid Nodule (Med Phys), Lymph node meta (Acta
Radiol), Pulmonary Embolism (Eur Radiol), …
67. Results
2D (82.7%) • 3 X 2D (81.7%) • 3D (84.7%)
• 비교를 위해 augmentation 적용 안됨.
따라서 3D의 경우 variation이 큼. (많은 parameter)
68. CNN Airway Segmentation
80 COPD Patients’ Inspiration CT
69 CT volumes are included in training
11 CT volumes are NOT included in training
GS : Manual segmentation
107
Axial 3 slices, Sagittal 3 slices, Coronal 3 slices
32 x 32 x 3 x 3
Weights are shared
2-class
classification
CT
Volume
Probabili
ty
Volume
CNN
For each voxels inside lungs
Segment
ed
Airway
Hard thresholding (0.51) and
Select the connected component
73. 113
The Previous Research
for Quantification Definition of Similar Lung Images
Extraction of Distribution Features Extraction of Distribution Features
* https://en.wikipedia.org/wiki/Large_margin_nearest_neighbor
* Y.J.Chang, et al,. “A support vector machine classifier reduces interscanner variation in the HRCT classification of regional
disease pattern in diffuse lung disease: Comparison to a Baysian classifier”, Medical Physics 40 (5), 051912 (2013)
74. 114
Evaluation: Statistics
( * p-value < 0.01, ** p-value
< 0.001 )
Evaluation: Recall Accuracy
Fitting Test Data
Acknowledgement
This work was supported by the Industrial
Strategic technology development program
(10072064, Development of Novel Artificial
Intelligence Technologies To Assist Imaging
Diagnosis of Pulmonary, Hepatic, and Cardiac
Diseases and Their Integration into
Commercial Clinical PACS Platforms)
funded by the Ministry of Trade Industry and
Energy (MI, Korea)
76. 116
AP 통신 : 인공지능(로봇)이 기사 작성, 기술 공개
초당 2,000개 기사 작성 가능
기존의 300개 기업 실적 -> 3000개 기업 실적 커버
77. 117
Right now, about 80% of Americans who need a lawyer can't afford one
"With ROSS, lawyers can scale their abilities and start to service this very
large untapped market of Americans in need,"
80. 정밀의료(Precision Medicine) Initiative
“And that’s why we’re here today. Because something called precision
medicine … gives us one of the greatest opportunities for new medical
breakthroughs that we have ever seen.”
President Barack Obama
January 30, 2015
정밀의료 치료기기 핵심 : 빅데이터 + 인공지능
빅데이터
▶ 인공지능
▶ 환자 맞춤형 진단/치료/공공
120/27
81. 의료 빅데이터 + 인공지능
의료 빅데이터를 이용하는
정밀의료 실현
사물인터넷 유전자검사
의료영상
환자 모니터링
하루 37만건의 의무기록 (연
간 약 1TB)
연간 약 2백만 영상데이터
(30TB)
3.8×109 염기쌍을 가진 DNA
정보
대형 종합 병원
82. 의료비 절감의 필요성
의료효율성
국내현황
Efficacy (효용성)
Cost (비용)
Equation = Efficacy / Cost
• 정밀치료를 통한 효용성 극대화
• 효율적 의료시스템 구축을 통한 비
용 감소
• 고령화 사회 등으로 GDP대비 의료비 비율 빠르게 증가하고 있으며 상승폭 또한 커지고 있
음
• 국민 소득 및 소비수준을 고려하여 의료비는 변화에 반비례하여 감소되어야 할 필요성 존재
83. 인공지능 관련 국내외 시장 현황
사회 및 산업 각 분야시장의 핵으
로 부상중인 인공지능
2020년까지 연평균 53.65% 성장
률을 기록할 것으로 예상1)
국내의 인공지능 시장 규모 예상
2020년 2조2천억원
2025년 11조원
2030년 27.5조원2)
1) 2016 market and market 2) KT경제경영연구소
123
84. Global Trends Drive Momentum in health care industry
Data Explosion 150+ exabytes Amount of healthcare data today1
Over 230K Active clinical trials2
80% Healthcare data that comes from unstructured data sources3
Dynamic
Delivery
Environment
50% Expected alternative payments form the Centers of Medicare and Medicaid by
2018 4
75%+ Percentage of patients expected to use digital health services in the future5
90K Expected shortage of physicians by 2020 6
Value vs Volume 4.7 trillion Estimated global economic impact of chronic disease by 2030 7
3 trillion Estimated US healthcare spending 8
100’s Approx. amount of decisions a person living with Type 1 Diabetes makes a day9
Efficient and
effective R&D
1 in 10 Clinical trials in cancer that are shut down due to lack of participation10
2.6B Average costs to develop a new pharma drug11
<10% Amount of drug currently in development that make it market12
124
1: NCBI: big data analytics in healthcare: promise and potential 2: ClinicalTrials.gov, 3: NIH 4 CMS 5: McKinsey Healthcare's Digital Future July 2014,
6: AAMC Report The complexities from 2014 ro 2025 7: WEF global economic burden non-communication diseases 8: Health affaires. Team analysis
9: OpenAps.org 10: Bio-clinical development success rates 20, Health Economic volume 47, May 2016
1. Life expectancy data, WHO, 2012
2. 2015 Global life sciences outlook: adapting in an era of transformation, Deloitte DTTL, 2014
3. Informa Pic Market Line Extracted Oct 2014
85. Opportunity
126
8 trillion : Industry
Size
2 trillion : waste in industry
Better experience
Imaging :
Unnecessary tests
Lower cost
Oncology:
Variability of Care
Better outcomes
Life sciences:
Failed clinical trials
Government:
Fraud, Waste and Abuse
Value Based Care:
Cost of chronic disease
360 billion : total IT and
healthcare market
opportunity
*IBM Watson
86. 국제적 동향
인공지능 의료적용 현황 (해외)
1) marketandmarkets 2016.02
세계 시장 규모는 5.05조원 (2020년 예측치), 향후 5년 성장률 53.65%으로 新성장 사업분야1)
의료 문서, 영상 빅 데이터를 활용한
진단 및 처방이 가능한 기술을 개발
헬스케어 분야부터 정밀의료 산
업까지 폭넓은 기술 로드맵 보유
폐암진단을 위한 딥러닝
기반 분석 시스템
131
89. 인공지능 의료적용 현황 (국내)
국내 동향
• 삼성과 LG를 비롯한 IT 기업의 높은 인공지능 분야 기술력
• 세계적 수준의 임상의료 및 임상시험 기술
• 최근 뷰노코리아, 루닛 등 인공지능기반 의료분야 스타트업들이 조기성과를 보이는 등 의료분야
인공지능 산업 기반 확대
미만성폐질환에서 CT로부터 질병을 판별. 질병
판독. 질병의 진행상황, 치료법 등 전문의 판단
을 도움을 주는 시스템 개발
소아 골연령 판정 보조 시스템 개발
흉부단순촬영 등 영상 데이터를 딥러닝 기술로
학습, 결핵, 유방암등 검출하는 기술을 개발
시스템생물학에 인공지능을 결합시
켜 기존 약물 개발 과정을 개선
135
90. 인공지능 의료적용 분야
인공지능 분야
시각지능
언어지능
판단지능
자동분류
요약/창작
공간지능
임상시험
케이스선정
신약개발프로세스
진료보조
비서서비스
음성인식 의무기록
데이터기반 정밀의료
유전체분석
약혼합사용 및 합병증 예측
진단검사추천
판독보조
정상유무판정
유사증례검색
판독문 생성
병리분야 판독 보조
물류, 수술실, 병실 운영
로봇수술
의료 인공지능
인공지능 의료적용 분야
136
91. 인공지능 의료적용 분야 (임상 중심 분류)
정상 유무 판정
초기진단 빅 데이터와 차세대
인공지능 기법을 통해 정밀판독
이 없이도 초기 감별진단
유사증례검색
DB의 수많은 증례로부터 유사
한 증례를 검색, 시각화하여 진
단에 도움
예비 판독문 생성
인공지능 기반 의료영상 분석기
법과 자연어 처리기술을 융합,
영상전문의의 판독을 보조할 수
있는 수준의 판독문 자동생성
병리분야 판독 보조
병리영상 빅데이터에 인공지능
기법의 적용, 진단, 발병기전 분
석, 예후 예측 등에 활용
임상시험
인공지능 기술을 기반으로 단일
병원 및 병원 클러스터간 물류, 수
술실, 병실의 효율적 운영
물류, 수술실, 병실 운영
로봇수술
인공지능 기술을 통해 의료 로봇
의 수술을 계획, 위험도 예측, 침
습부위 최소화
진료보조
데이터기반 정밀의료
판독보조
신약
약물 개발과정에 인공지능 기법을 적용, 질병 치
료에 더욱 효과적인 약물의 조합과 용도 변경 탐
색, 약물 후보군 및 임상환자 군의 최적화
케이스 선정
인공지능 기반 검색기법을
통해 적합한 질병 및 환자
를 탐색, 임상 시험의 준비
기간을 단축, 객관성 향상
비서서비스
IoT기술, 음성 인식기술 및 인공지능 기술을 융합,
효율적인 예약, 진단 및 진료 프로세스, 업무 정보
업데이트 및 맞춤형 큐레이션
음성인식 의무기록
진단 및 판독 내용의 기록을 자동화,
의료분야 전문용어를 판독 및 구조
화할 수 있는 수준의 음성 인식 및
문서 생성 기술
진단 검사 추천
정밀 진단 및 치료를 위해 인공지능
및 빅데이터를 활용, 정확도를 높이
고 위험도를 낮추기 위한 추가 진단
검사 프로세스를 추천
유전체
맞춤 의료를 위해 유전체, 멀티모달 의료 영
상 및 임상병리 빅데이터를 바탕으로 연관
성을 분석, 모델링하여 예후 예측, 진단 및
치료
약 혼합 사용 및 합병증 예측
치료 및 약물 사용 시 사례기반 위험성 또는 합병
증의 위험을 알려 주어 의사의 최종 결정을 보조
137
92. 인공지능 의료적용의 어려움과 대책
139
의료환경의
표준화의
어려움으로
인공지능 훈련을
위한 라벨링 어려움
개인정보보호에
따른 데이터의
비개방성
의료기기 인허가,
신의료기술인정 및
수가 장벽
인공지능 전문가와
의료전문가의 단절
의료빅데이터의
표준화와 연계 연구
필요
인공지능기술개발을
위한 공개 데이터
개발과 이를 이용한
그랜드 챌린지
방식의 기술 진흥
필요
연구기획 초기부터
관련 종합 대책
정책적 지원 방안
마련
인공지능 교육,
연구개발 산업화를
망라하는 생태계
구축 필요
어려움
대책
93. Strategy
Proof of principal
Google
Retinopathy, JAMA 2016
– 130k images, 54 ophthalmologists
– Reference standard
– Severity, image quality, L/R, FoV
Ophthalmology, Pathology, Radiology,
Dermatology
Cloud/Vision/Speech API
AI make AI
Amazon, MS
Naver
J Proj. Clova, 오감 AI
Line chat bot since 2014
141
94. Technical Issues on Medical Imaging
Data collection
More (clean) data
Accurate annotation
Legal issues
Models selection
Deeper network
Off the shelf models
Result Interpretations
Neural network visualization
Human-friendly interpretation
142
95. Machine Operable, Human Readable
Visual attention
Category – feature mapping
Sparsity and diversity
150
96. Machine Operable, Human Readable
Evidence hotpot for lesion visualization
“SpineNet: Automatically Pinpointing Classification
Evidence in Spinal MRIs”
151
98. MD-friendly Interpretation
Breast cancer risk prediction through BI-RADS categorization of
mammography
Analysis and visualization for breast density prediction
Mapping and visualization of patient by predicted breast cancer risk
score
153