Raai 2019 clinical unmet needs and its solutions of deep learning in medicine3
1. Clinical Unmet Needs and their
Solutions of Deep Learning in
Medicine
Namkug Kim, PhD
Medical Imaging & Intelligent Reality Lab.
Convergence Medicine/Radiology,
University of Ulsan College of Medicine
Asan Medical Center
South Korea
2. Conflict of Interests
Establishments of Somansa Inc, CyberMed Inc, Clinical Imaging Solution Inc, and Anymedi
Inc.
Researches with LG Electronics, Coreline Soft Inc.,, Osstem Implant, CGBio, VUNO,
Kakaobrain, AnyMedi Inc.
4. AMC : a leading hospital in Korea
▪ Total area of 524,000 m2
▪ No.1 in number of patients and
treatments
▪ No.1 in performing surgeries on
top 6 cancer patients & 30
major diseases
New BuildingEast BuildingWest Building
Education &
Research Center
Convergence
Innovation Building
Dormitory
Total Beds
2,715
licensed beds
Outpatients Inpatients
13,380
(Daily average)
2,586
(Daily average)
Medical Staffs
7,924
(MDs: 1,720
RNs: 3,627
Others: 2,577)
Surgical Operations
91,000 Annual
(250 / day)
JapanChina
South
Korea
North
Korea
Russia
China
5. A world best hospital in organ transplantation & cancer surgery
Transplantation
in Korea
Liver Heart Kidney Pancreas
34.9% 47.2% 14.6% 64.8%
98
85
88
Liver
98 97
95
Heart
95
86 86
Kidney
95 95
85
Pancreas
AMC (Asan Medical Center)
UNOS (United Network For Organ Sharing)
KONOS (Korea Network For Organ Sharing)
Survival rates
(1 year)
6. Investing heavily in clinical and inter-disciplinary research
Achieve a world-class medical reputation to
aggressively invest in R&D and
clinical trial
Clinical Research
Investigator (MD)
685
Lab Technician
839
Basic Research
Investigator (PhD)
119
Administration
39
(Units : Number of personnel)
Selected research-driven hospital by Korean
government
R&D funding US$ 1.2 billion by
the government (10 hospital, 3 yrs)
Over 1,600 Investigators for Research
AI
Established a new research
center for artificial intelligence in
medical imaging
3D
Printing
Leading 3D printing technology
in medical field
Genome
Genomic studies on rare genetic
diseases, cancer, chronic
diseases, etc.
Medical
Robot
Robotic training center for
training & researching on robotic
surgical skills
⋯
Development of clinical applications by open
network and translating advanced
technologies
1st ship building company
in the world
7. Beyond Human-level Performance
• Now, Machines Beat Human in Tasks Once Considered Impossible
5:0
vs Fan Hui
(Oct. 2015)
4:1
vs Sedol Lee
(Mar. 2016)
Modified by KH Jung, PhD
9. Five tribes of AI researches
From Pedro Domingos, Professor, University of Washington at MLconf ATL
10. Bio Plausible Neural Network
Mimic human visual
recognition system
Neocognitron, proposed by
Hubel & Wiesel in 1959
Visual primary cortex by
cascading from S-Cell to C-Cell
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
Nobel prize winner @ 1981
10From Gallant and van Esses, Simon Thorpe
11. CNN : Major Breakthroughs in Feedforward NN
K. Fukushima Yann Lecun G. Hinton, S. Ruslan
Neocognitron (1979)
• By Kunihiko Fukushima
• First proposed CNN
Convolutional Neural Networks (1989)
• Yann Lecun et.al
• Back propagation for CNN
• Theoretically learn any function
Neocognitron
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. RosenblattM. 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)
1950
Neocognitron (1959)
• Hubel & Wiesel
• by cascading from S-
Cell to C-Cell
12. 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]
16. Computational map
16
Dense
Few
Millions
#ofVariables
Completeness of Data (Sampling)Sparse
More
• Compute
• Data
• Storage
• Bandwidth
Computationally
Intractable SpaceDeep Learning
Neural Nets
Statistical Analysis
Algorithms, Closed Form Solutions
Expert
Systems
Insufficient
data for analysis
[Un]supervised Learning
Models
Intuition
Log scale
Log scale
17. Machine Learning vs Deep Learning
— Scale Matters
— Millions to Billions of parameters
— Data Matters
— Regularize using more data
— Productivity Matters
— It’s simple, so we can make tools
Data & Compute
Accuracy Deep Learning
Many previous
methods
Deep learning is most
useful for large problems
Modified by Nvidia DLI
20. AI + Healthcare Market Size
Healthcare AI Market Size:
10B USD@ 2015 -> 67B USD(7조원)@2021, CAGR : 42%/y*
*Prost and Sullivan, ** WHO, ***Variant Market Research, ****KHDI
Global healthcare expenditure**
Global IT healthcare market**
AI healthcare market @ SouthKorea****Ads 20B USD <<< Healthcare 9,500B USD (50x)
21. AI Medical Device
21
• Verily@Google: Normal vs Abn, Anti-aging, Life
Prolongation
• IBM: Truven, 40B USD
• Apple: GSK, EMR +iPhone Healthcare Platform
• Facebook: Incurable dx, Human cell atlas, 5000M USD
• Zebra Medical Vision
• AI medical imaging Dx : 1st place of investment
• Medical imaging reading cloud service/1 USD
22. AI Medical Device Cleared in FDA@2018
22
https://twitter.com/erictopol/status/1028642832171458563
23. 1st AI Doctor; IDx-DR
the first device to provides a
screeningdecision without
the need for a clinician to
also interpret the image or
results
IDx-DR
>=mildretinopathy;
refer to an eye care professional
< mildretinopathy
rescreen in 12 months
23
24. Requirements
DICOM/ITK/HL7/FHIR library
Physics of Imagingmodalities : X-
ray, CT, MRI, US
Domain knowledge of EMR, EHR,
PHR, HIS
Computer Science : Preprocessing,
post-processing
DL algorithms
Public / Private Datasets
Medical experts
24
25. Challenges
Deep Learning Needs Why
Data Scientists New computing model
Latest Algorithms Rapidly evolving
Fast Training Impossible -> Practical
Deployment Platforms Must be available everywhere
26. Opportunities for AI inmedicine
Ball, John, Erin Balogh, and Bryan T. Miller, eds. Improving diagnosis in health care. National Academies Press, 2015.
27. Opportunities for AI inmedicine
Ball, John, Erin Balogh, and Bryan T. Miller, eds. Improving diagnosis in health care. National Academies Press, 2015.
28. E.g., AI Application in Medical Imaging
Almost all aspects
Image transformation
Lesion segmentation
Lesion classification
Lesion detection
Findingsimilar cases
Assistance of interpretation
29. TASKS
Image SegmentationObject Detection
Image Classification +
Localization
Image Classification
(inspired by a slide found in cs231n lecture from Stanford University)
Nvidia DLI Education Materials
30. Clinical Unmet Needs andSolutions
1. Augmentation, curriculum learning, one / multi-shot learning to solve diseases’ imbalance, rare or a
small number of dataset
2. Efficient anonymization, curation, and smart labeling for cheap labeling
3. Domain adaptation or Image normalization to overcome differences of multi-center trials
4. Interpretability, explainablity and visualization to mitigate black box property
5. Uncertainty of medical data, and artificial intelligence prediction
6. Reproducibility study of deep learning using repeatedly scanned images
7. Generative model
8. Novelty (Anomaly) detection under unsupervised learning for human decision in later
9. Big data PACS and Content based image retrieval
10. Deep radiomics and deep survival
11. Clinical decision support system
12. Develop physics-induced machine learning with well-known physics and medical laws
13. Active and Robust learning to overcome adversarial attack and input changes
14. Applications on 3D Printing, AR/VR, Robotics, Optic devices, Etc
31
31. 1. Perlin noise augmentation
Infinity augmentation strategy
Perlin noise by complexity theory
32Bae HJ, Kim N, Sci Report 2018
32. Perlin noise
Randomnoise is too harsh to be natural
The Perlin noise : by simply addingup noisy
functions at a range of different scales.
33
Apply smooth filter to the interpolated noise
121
242
121
34. GAN augmentation to imbalanceddataset
35Hojjat Salehinejad, et al., ICASSP 2018
35. GAN synthetic images
X-ray, Pathology
classification task
DCGAN, PGGAN
2-3% accuracy drop
36https://arxiv.org/ftp/arxiv/papers/1904/1904.08688.pdf
Examining the Capability of GANs to Replace Real Biomedical Images in Classification Models Training
41. Comparison Results
42
❑ Curriculum learning vs. Baseline (w/o curriculum)
• Both models converged well.
• Curriculum learning showed better result, more stable and shorter training time.
Baseline
Epoch 1257
Loss 0.163
Acc
AMC: 96.1
SNUBH: 92.7
Curriculum
learning
Epoch 199
Loss 0.140
Acc
AMC: 97.2
SNUBH: 94.2
Loss on tuning set
Epoch
Loss
42. 1. Curriculum learning : Endoscopy of Colonic Polyp
Polyp detection
in endoscopy
(video)
Close-up shot on
polyp of interests
Pathological
Diagnosis
(multi-modal)
• https://arxiv.org/abs/1512.03385 “Deep Residual Learning for Image Recognition”
• https://arxiv.org/abs/1512.04150 “Learning Deep Features for Discriminative Localization”
AMC Digestive Internal Medicine Byeon JS
Procedure of Detection and Classification of colon polyps
CAM result of classification
Colonoscopy Automation
Histopathology
Hyperplastic
polyp
: HP 65
166
Sessile serrated
adenoma
: SSA 101
Benign
adenoma
: BA 521
655Mucosal or
superficial
submucosal
cancer
: MSMC 134
Deep
submucosal
cancer
: DSMC 101 101
44. Further
46
Integration of CT virtual colonoscopy Depth generation from CT virtual colonoscopy
endoscopyCT virtual colonoscopy
45. 2. Smart Labeling : Concept
Manual label Smart label
Smart label
Manual label
200 hours
Per 100 cases
500 hours
Per 100 cases
Manual label
SW label
Label
Correction
Total Dataset
Manual Labeling
Results
Sub-dataset
Manual Labeling
Training model
DL labeling Label correction Results
Active learning
Additional dataset
46. 2. Smart Labeling : Ex
Manual label Smart label
50-60% faster
Abdominal multi-organ segmentation (kidney)
• Parenchyma, renal cell carcinoma, artery, vein,
ureter
• 4-5 hours per case by expert manual segmentation
Abdominal multi-organ segmentation (kidney)
• DL model segment images at first
• Expert corrects segmentation result of DL models
• 1-2 hours per case
47. 2. Smart Labeling; Comparison of AI Generated Dataset
Kim YG, Kim N, et al, Sci Report 2019; Go HJ, Pathology, AMC, Sci Report (2018)
Peri-tubular capillary (PTC) counting
48. 2. Smart Labeling : Surgical Imaging
AI assisted labeling with semantic
segmentation from medical images
Pancreas
Stomach, kidney, liver, etc
Cervical spine
(Med Phys Revision)
MICCAI 2018 Segmentation Decathlon 2nd place
Semantic segmentation of AI
saving time
• No human
interaction
• 10 msec/slice
more accurate
• More
reproducible
• More robust
saving cost
• Fast
segmentation
• 2~3 times faster
after correction
Tip!
Maxillary sinus, mandible,
mandibular canal (RSNA 2018) Glioblastoma (GBM)
Lung lobe (JDI 2019)
Breast (RSNA 2017)
Airway (MIA 2019)
49. 2. Smart Labeling; Medical Segmentation Decathlon
51
10 organs; 52 labels
MSD challenge, MICCAI 2018
Cascaded U-Net
2nd Place
52. Human Segmentation
Human Segmentation
AI_ 1st Test
AI_ 1st Test
AI_ 2st Test
AI_ 2st Test
AI_ 3rd Test
AI_ 3rd Test
rebuilding ground truth increasing data
0.88
2. Smart Labeling; Active Learning
DICE 0.91 0.95 10 sec
TH Kim, KH Lee, Kim N, Sci Report Submitted; Kyoung YS, Nephrology, AMC; Submitted to Sci Report
53. X-ray Line Tracing
55
The results of Cardiomegaly border line segmentation
Yang DH, , Radiology AMC; Park JW, Kualldam Dental Hospital
61. Denoising network
Yang, Q., Yan, P., Kalra, M. K., & Wang, G. (2017). CT image denoising with perceptive deep neural networks. arXiv preprint arXiv:1702.07019.
65. Super Resolution : Compressive Sensing
CS 6.84 -> CS5.56
PredictOriginal Difference
Training Result with ~3000 training images
AMC Radiology Jung SC
71. 4. Interpretability vs Accuracy
73
PredictionAccuracy
Explainability
Learning Techniques (today)
Neural Nets
Statistical
Models
Ensemble
Methods
Decision
Trees
Deep
Learning
SVMs
AOGs
Bayesian
Belief Nets
Markov
Models
HBNs
MLNs
SRL
CRFs
Random
Forests
Graphical
Models
Explainability
(notional)
DARPA; Distribution Statement "A" (Approved for Public Release, Distribution Unlimited)
72. DARPA XAI (eXplanable AI)
74
New
Approach
Create a suite of
machine learning
techniques that
produce more
explainable models,
while maintaining a
high level of learning
performance
PredictionAccuracy
Explainability
Learning Techniques (today) Explainability
(notional)
Neural Nets
Statistical
Models
Ensemble
Methods
Decision
Trees
Deep
Learning
SVMs
AOGs
Bayesian
Belief Nets
Markov
Models
HBNs
MLNs
Model Induction
Techniques to infer an explainable model
from any model as a black box
Deep Explanation
Modified deep learning techniques to
learn explainable features
SRL
Interpretable Models
Techniques to learn more structured,
interpretable, causal models
CRFs
Random
Forests
Graphical
Models
DARPA; Distribution Statement "A" (Approved for Public Release, Distribution Unlimited)
73. 4. Visualization : Machine Operable, Human Readable
Visual attention
Class Activation Map (CAM)
Category – feature mapping
Sparsity and diversity
75
77. Image Pattern Classification on DILD HRCT
(a) Consolidation (b) Emphysema (c) Normal
(d) Ground-glass opacity (e) Honeycombing (f) Reticular opacity
Image
Feature
Extraction
Machine
Learning
Training dataset Testing set
Accuracy (%)
SVM Classifiers
GE GE 92.02 ± 0.56
Siemens Siemens 89.92 ± 0.36
GE+Siemens GE+Siemens 89.73 ± 0.43
N : 100 ROIs in each class of GE, Siemens CT
Kim GB, Kim N, et al, Journal of Digital Imaging, 2017
78. CNN networks
80
Input Image Conv. Layer #1 Conv. Layer #2 Conv. Layer#3 Conv. Layer#4 FC #1 FC #2
Local Response
Normalization
Dropout
100
6
20
20
19
19
64
64
64 64
9
9
4
4
4
4
4
4
4
4
3
3 3
3
3
3
Max
Pooling
Max
Pooling
Local Response
Normalization
• Two basic data replication
✓ random cropping and flipping: 20 x 20 pixels of ROI patches (top left, top right, center, bottom left, bottom right) & horizontally flipped
✓ mean centering and random noising: random noise of Gaussian distribution with mean zero and standard deviation
Kim GB, Kim N, et al, Journal of Digital Imaging, 2017
79. 4. Image Pattern Classification on DILD HRCT
(a) Consolidation (b) Emphysema (c) Normal
(d) Ground-glass opacity (e) Honeycombing (f) Reticular opacity
DEEP
Learning
Kim GB, Kim N, et al, Journal of Digital Imaging, 2017
80. 4. t-SNE Map ; SVM vs CNN
82
O = GE; X = SiemensSVM CNN
Kim GB, Kim N, et al, Journal of Digital Imaging, 2017
83. 5. Uncertainty
Uncertainty of training data
In clinical situation, it is common
Deep Bayesian Modeling
Uncertainty of classification/prediction of Machine
Learning
85
85. 6. Reproducibility
Test-retest is one of most important issues of biomarker
Multiple scanning within similar date
Evaluate reliability of AI/Deep learning
Chest PA (Nodule)
Nodule Size on Chest PA with YOLO
– 56% : Variability of Size Marker
– Chest PA 50 pairs
DILD CT
87
86. 6. Chest PA : 5 Disease Patterns
▪ Curriculum learning for multi-label classification
Results
Lee SM, Seo JB, Radiology, AMC
87. 6. Chest PA : Extra-validation
▪ Extra validation using multi-center datasets
Results
AMC
Densenet 201
True
Lesion Normal
Pred.
Lesion 690 44
Normal 30 1170
AMC
Resnet 152
True
Lesion Normal
Pred.
Lesion 668 61
Normal 52 1153
Sensitivity: 95.8%, Specificity: 96.4% Accuracy: 96.2% Sensitivity: 92.8%, Specificity: 95.0% Accuracy: 94.2%
SNUBH
Densenet 201
True
Lesion Normal
Pred.
Lesion 99 14
Normal 1 86
SNUBH
Resnet 152
True
Lesion Normal
Pred.
Lesion 100 38
Normal 0 62
Sensitivity: 99%, Specificity: 86% Accuracy: 92.5% Sensitivity: 100%, Specificity: 62% Accuracy: 81.0%
Lee SM, Seo JB, Radiology, AMC
88. 6. Chest PA Reproducibility : Nodule Detection
▪ Reproducibility analysis on CNN-based detection
Comparison of various networks
✓YOLO v2
✓VUNO-net
Nodule
(N = 121)
1st
P N
2nd
P 94 3
N 15 9
✓Faster RCNN
✓Mask RCNN
Nodule
(N = 121)
1st
P N
2nd
P 90 8
N 11 12
Nodule
(N = 121)
1st
P N
2nd
P 90 3
N 13 15
Nodule
(N = 121)
1st
P N
2nd
P 97 4
N 14 6
Lee SM, Seo JB, Radiology, AMC
Nodule
(N = 121)
1st
P N
2nd
P 116 1
N 2 2
✓Reader 2
Nodule
(N = 121)
1st
P N
2nd
P 111 3
N 4 3
✓Reader 1
90. PGGAN; Chest PA Xray
92
Progressive growing GAN (PGGAN)
https://arxiv.org/abs/1710.10196
Tero Karras, et al
Progressive Growing of GANs for Improved Quality, Stability, and Variation,
97. Conditional generators
GAN Architecture2
[3] Jun-Yan Zhu et al., Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks, ICCV, 2017
[4] Xi Chen et al., InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets, NIPS, 2016
InfoGAN4
Style transfer Networks CycleGAN3
Latent Space of GAN
[2] “An intuitive introduction to Generative Adversarial Networks (GANs)“, https://medium.freecodecamp.org/an-intuitive-introduction-to-generative-adversarial-networks-gans-7a2264a81394
Disentangling Latent Space
98. [5] TL-GAN: transparent latent-space GAN, https://github.com/SummitKwan/transparent_latent_gan
PGGAN6
CNN Classifier7
[6] Tero Karras et al., Progressive Growing of GANs for Improved Quality, Stability, and Variation, ICLR, 2018
[7] Beomhee Park et al., A Curriculum Learning Strategy to Enhance the Accuracy of Classification of Various Lesions in Chest-PA X-ray Screening for Pulmonary Abnormalities, Scientific Reports, submitted
Proposed Model Architecture5
99. CNN Classifier7 based on supervised learning
[7] Beomhee et al., A Curriculum Learning Strategy to Enhance the Accuracy of Classification of Various Lesions in Chest-PA X-ray Screening for Pulmonary Abnormalities
Dataset Chest PA X-ray images from
6069 healthy subjects and 3417 patients at AMC,
1035 healthy subjects and 4404 patients at SNUBH
Task To detect 5 pulmonary abnormalities in Chest PA X-ray(CXR)
:Nodules(ND), consolidation (CS), interstitial opacity (IO), pleural effusion (PE), and pneumothorax (PT)
AMC SNUBH
Sensitivity 85.4% 97.9%
Specificity 99.8% 100.0%
AUC* 0.947 0.983
Feature Extractor Classifier
*AUC: Area Under the Curve
Feature Extraction with Classifier
100. Vector Orthogonalization
t-SNE projection of latent space representations on MNIST dataset8 HouseHolder QR Factorization9
Linear Regression on Latent Vectors
[8] “THE MNIST DATABASE of handwritten digits”, http://yann.lecun.com/exdb/mnist/
[9] “QR decomposition using Householder transformations”, https://www.keithlantz.net/2012/05/qr-decomposition-using-householder-transformations/
Finding Feature Axes
104. 7. Generative Adversarial Network (GAN)
Deep Convolutional Generative Adversarial Networks (DCGAN)
Rotations are linear in latent space
Bedroom generation
Arithmetic on faces
106. 8. Novelty (Untrained catergory)
In clinical situation
Novelty is everywhere, especially supervised learning
Rare diseases, but well known to medical doctors
Hard to training
How to determine novel (untrained) category
Unsupervised learning
Semi-unsupervised learning
Normal vs abnormal
Abnormality Detection
108
120. Big-Data PACS for Quantitative COPD Reading
❖ Auto DICOM retrieve from legacy PACS.
❖ Auto processing supported by Deep-learning.
❖ Stored all the quantification values for every COPD
patient
❖ Dashboard showing the overall statistics.
❖ Query by quantification.
❖ Similar case can be easily queried by quantification
values when reading a patient’s chest CT images.
RSNA 2018
125. Content Based Image Retrieval in DILD CT
DILD Classification using SVM
Normal
Honeycombing
Ground -glass opacity
Consolidation
Emphysema
Reticular opacity
𝑁. 𝐿 . 𝑣𝑜𝑥𝑒𝑙𝑠
𝑙𝑢𝑛𝑔 𝑣𝑜𝑥𝑒𝑙𝑠
𝐻. 𝐶. 𝑣𝑜𝑥𝑒𝑙𝑠
𝑙𝑢𝑛𝑔 𝑣𝑜𝑥𝑒𝑙𝑠
𝑅. 𝑂. 𝑣𝑜𝑥𝑒𝑙𝑠
𝑙𝑢𝑛𝑔 𝑣𝑜𝑥𝑒𝑙𝑠
: 4 x 4 x 4 x 6
= 384 features
Red : IPF Blue : COP Green : NSIP
76.7% and 81.7% in Top3, and Top5
Same patient with no change CTs
3.88 ± 0.96 5-scale visual scoring
by two radiologists
RSNA 2019
126. Content Based Image Retrieval in COPD CT
COPD feature extraction for CBIR
Top 1; similarity score 5/5 Top 2; similarity score 4/4
Top 3; similarity score 4/4 Top 4; similarity score 4/4 Top 5; similarity score 2/3
Query case
CBIR retrieval for similar chest CT with COPD
60.0% and 68.0% in Top3, and Top5
Same patient with no change CTs
3.55±0.98 5-scale visual scoring
by two radiologists
RSNA 2019
130. CT Image Texture Features of NSCLC
134
Unsupervised
Hierarchical
Clustering
131. GBM from PCNSL (primary central nervous system lymphoma)
136
Park JE, Kim HS, et al, Sci Report 2019
132. GBM from PCNSL (primary central nervous system lymphoma)
137
Park JE, Kim HS, et al, Sci Report 2019
133. COPD Deep Radiomics : Survival Prediction
Method
KOLD cohort
Random survival forest
Extraction of deep features
Design a CNN model for classifying 5-
year survival
Each slice was trained separately, and
used for extracting representative
deep features of it
138
Ishwaran, H., et al., “Random survival forests,” arXiv:0811.1645.
Conv2D
Batchnorm
ReLu
Maxpooling2D
Conv2D
Batchnorm
ReLu
Maxpooling2D
Conv2D
Batchnorm
ReLu
Maxpooling2D
Conv2D
Batchnorm
ReLu
Maxpooling2D
Conv2D
Batchnorm
ReLu
Maxpooling2D
FCN(1,024)
5-year
survival
CT slice
(a) A patient survived less than 5-year.
(b) A patient survived more than 5-year.
Data C-index
Deep radiomics features (A1+A2+A3+A4+A5+A6) 0.8412
Demographic information 0.7688
PFT 0.7498
Handcrafted features 0.5994
Accuracy of survival prediction
CAM of survival prediction
CNN architecture for Deep radiomics
134. Extra-Validation
Malaysian
111→102 COPD patients
Death: 18 / censored: 0 / surviving: 84
11 CT slices based on preselected anatomic landmarks
3 images in the coronal plane
– Vertebral body, center of tracheal carina and superior vena cava
2 images in the sagittal plane
– Center of left and right lung
6 images in the axial plane
– Upper 2 slices and lower 2 slices at intervals of 2 cm on the center of tracheal carina
139
Mortality data (based on 'copddeath)
The median follow-up time as of 31 June 2017 was 1000 days (2 years and 9 months) (range, 60 to 1400 days). Eighteen
out of 112 (16%) died during this time. Death causes were as follows: COPD (n=8), pneumonia (n=4), congestive cardiac
failure/ coronary artery disease (n=2), lung carcinoma (n=1), stroke (n=1), motor traffic accident (n=1) and intestinal
obstruction (n=1). The final two causes are considered as non-COPD associated causes. As such, 16 patients died of COPD
associated death (14.2%).
AMC Radiology, Seo JB
137. 11. CDSS on Liver Cancer
Kim KM, GI, AMC
Clinical data
(20 variables)
RFA / Operation Or not
Operation
Classifier 1
Prediction B
TACE Or not
TACL+RT
Prediction C
Classifier 2 Classifier 3
sorafenib
Others
Or not
Or not
RFA
Prediction A
LT
Classifier 4
Prediction G
Prediction D
Supportive
care
Classifier 5
Classifier 6Prediction E
Prediction F
147
138. CDSS on Liver Cancer
Treatment recommender system for liver
cancer
Patient’s baseline data: 20 variables
1,000 patients
Hierarchical decision modeling
Survival prediction
Cox hazard regression, random forest survival
148
Survival Curve on Tx of Liver Cancer
Kim KM, GI, AMC
141. 12. CNN for Steady Flow Prediction
151
𝑦=f’(x)
142. 14. Surgical Robot with AI
Will robot steal surgeons’ job?
NO.
Will robot CHANGE surgeons’ job?
It may...
Will robot and SUPER COMPUTER steal surgeons’ job?
…
143. 14. Simple Mastoidectomy Surgery Video
Simple Mastoidectomy Surgery Video
Understanding
Surgical Step Understanding
9 Steps
153
Example
AMC ENT Jung JW
Classification algorithm of
surgical stage by learning
video data
중이염 동영상 내
수술도구(위),
Henle recognized
(below)
149. 14. Hospital Workflow Improvement
Intelligent Hand
Hygiene Support
@ Lucile Packard Children's
Hospital at Stanford
Intelligent ICU Clinical
Pathway Support
@ Stanford Hospital and Clinics
@ Intermountain Healthcare
Intelligent Senior Wellbeing
Support
@ Palo Alto Medical Foundation
@ On Lok Senior Health Services
@ Intermountain Healthcare
** Stanford AI-assisted Care
Courtesy of Ha HS, VAIIM consulting, modified
150. 14. Beyond computer vision
160
Computer vision to discern clinical behaviors
Bedside Computer Vision, Yeung S, Fei-Fei Li et al, NEJM 2018 April
H Rosette *
151. Hand Hygiene Monitoring
Clean
Clips Images Clips Images
Training 428 8,325 444 11,988
Validation 46 871 43 1,047
Testing 10 340 8 516
Total 484 9,536 495 13,551 Test accuracy: 0.82
156. The limitations and solutions of 3DP in medicine
166
Limitations Solutions
Slow printing speed New 3DP for 100~1000 times faster
Variations of clinical imaging protocols ;
especially thick slice thickness
Domain adaptation;
Super resolution with DL
Limitation of materials Develop new materials and mechanism
Slow segmentation for specific anatomy Training DL for semantic
segmentation
New manpower Train new expertise with help of AI
Physicians’ additional efforts Communication tools with AI
Already doing well Find new killer apps with 3DP
RSNA 2018
157. Semantic Segmentation
167
Medical Images
Super Resolution Module
Auto Design Module
GAN
진단용
저해상도 영상
3D모델링용
고해상도 영상
train/test set
Generator Discriminator Y/N
Similar?
Auto Segmentation Module
3D모델링용
고해상도 영상
Segmented
mask
FCN Network
Labeling
Pre-processor
train/test set
End-to-End learning
Segmented
mask
Surgical planner
Ref. tracker Border tracker Surgical
planning
Y/N
meet
Design
spec..?
DeepLearning
RSNA 2018
158. AI Solution for 3DP
168
Data
acquisition
DL Modeling 3DP
Confirm by
an expert
Confirm by
M.D.
Application
Yes
NoNo
Yes
Segmentation with human-AI interaction
Communication
with AI
Low-resolution image
for diagnosis
High-resolution image for
3DP with super-resolution
Semantic segmentation method
Super-resolution method
▪ Super-resolution convert low-resolution
images into high resolution images.
▪ The result images of super-resolution could
be better 3D models.
Gold Standard 1st training
Segmentation enhancement with human-AI interaction
2nd training 3rd training
AI based 3DP workflow
Manual
segmentation result
Result of 2D U-net
(about 30 sec / case)
Result of 3D U-net
(about 1 sec / case)
RSNA 2018
162. Artificial Intelligence (+Big Data) Will Redesign
Healthcare
1. Artificial Intelligence (x) ->Augmented Intelligence (o)
2. Just Starting
3. Precision medicine / Mining medical records / Designing
treatment plans
4. Getting the most out of in-person and online consultations /
Health assistance and medication management / Open AI
helping people make healthier choices and decisions
5. Assisting repetitive jobs
6. Drug discovery / Clinical trial Case Matching
7. Analyzing, redesigning a healthcare system
http://medicalfuturist.com/ Aug, 2016
163. Job Openings @ MI2RL, AMC
Post-doc research fellow, PhD Students, Researchers
173
AMC, UoU
Seoul South Korea
Contact to namkugkim@gmail.com
164. Collaborators
Radiology
Joon Beom Seo, SangMin LeeA,B, Dong Hyun, Yang, Hyung Jin Won, Ho Sung Kim,
Seung Chai Jung, Ji Eun Park, So Jung Lee,Jeong Hyun Lee, Gilsun Hong
Neurology
Dong-Wha Kang, Chongsik Lee, Jaehong Lee, Sangbeom Jun, Misun Kwon, Beomjun Kim
Cardiology
Jaekwan Song, Jongmin Song, Young-Hak Kim
Emergency Medicine
Dong-Woo Seo
Pathology
Hyunjeong Go, Gyuheon Choi, Gyungyub Gong, Dong Eun Song
Surgery
Beom Seok Ko, JongHun Jeong, Songchuk Kim, Tae-Yon Sung
Internal Medicine
Jeongsik Byeon, Kang Mo Kim
Anesthesiology
Sung-Hoon Kim, Eun Ho Lee