Data-Driven AI
for Entertainment and Healthcare
Demetri Terzopoulos
UCLA Distinguished Professor & Chancellor’s Professor of Computer Science
Co-Founder & Chief Scientist, VoxelCloud, Inc.
Visual Computing
• Computer graphics (synthesis)
• Computer vision (analysis)
Talk Overview
1. AI/ML in computer graphics
- Human Modeling and Animation
2. AI/ML in computer vision
- Medical Image Analysis
3. The future & questions
 Images & Videos
 Mathematical models
Computer
Models
Images /
Videos
Computer
Vision
Computer
Graphics
“Final Fantasy: The Spirits Within”
(Square Pictures, Inc., 2001)
Virtual Humans in Movies and Games
These characters are neither autonomous nor intelligent
“Metal Gear Solid” game
Motion Capture Technology
3D tracking of body-attached IR reflectors
Physics
The Artificial Life Approach
Comprehensive computational models of humans and animals
• Modeling the body and mind
Biomechanics / Locomotion
Perception
Behavior
Learning
Cognition
The Artificial Life Modeling Pyramid
Artificial
Intelligence
Realistic Biomechanical Modeling of the Human Body
• Almost all the articular bones and skeletal muscles
– 75 bones (165 DOFs), 846 muscles
• Volumetric finite element soft tissue model
– 354K tetrahedral elements
3D Musculature Model
Biomechanical Actuator Model
Upper Body Actuators
Biomechanical Simulation
Biomechanical Simulation
How Can we Control Complex, State-of-the-Art
Biomechanical Human Models Like These?
The natural way is through neuromuscular control
• The advanced virtual human models can learn to control themselves
like real humans do!
– This is accomplished using massive quantities of training data
– The training data are synthesized by the human models themselves
Anatomical Structure of the Neck
Skeletal System
• 7 cervical vertebrae and a skull
coupled by 3-DOF joints
• Ligaments/disks
passive joint springs
• Equations of motion
0)qb(q,qM(q)  
moment
arm matrix
active
muscle
force
neural
inputmass
a),q(q,P(q)f)qb(q,qM(q) c
 
gravity, Coriolis,
passive elastic
forces
Biomechanical Neck Model
Total of 72 anatomically-based muscle actuators
in 3 layers
48 deep muscles
(16 longus colli, 16 erector, 16 rotator)
6 muscles at each joint increase controllability
12 intermediate muscles
(scalerius: 4 anterior, 4 posterior, 4 capitis)
12 superficial muscles
(2 sternomastoid, 2 cleidooccipital, 8 trapezius)
The big challenge is co-actuation and control
What Would Leonardo da Vinci Think of This?
Neuromuscular Control of the Musculoskeletal Model
muscles
muscle
contraction
forces
skeletal
system
environment
gravity,
applied
force
bio-
mechanical
face
head pose
voluntary
controller
feedfwd
signal
setpoint
signal
proprioceptive feedback
(pose, velocity of head)
reflex
controller
muscle
activation
levels
muscle feedback
(strain/strain rate)
ts
tf
tu
tt s
Trained Deep Neural Networks
• Set random target pose
Training the Neural Networks
• Using inverse kinematics, compute desired muscle lengths
• Using inverse dynamics, compute muscle activations to
achieve desired muscle lengths (under gravity)
Training the Neural Networks
Training the Neural Networks
Target Pose
Activations
• Repeat with about 20K random target poses
Learned Neck-Head-Face-Eye Behavior
Realistic Animation of Swimming
Biomechanical human model immersed in fluid
• Advanced multi-physics simulation
Learning to Swim From Examples
Close-up View of the Biomechanical
Swimmer Model
Medical Image Analysis
Deformable models: A powerful, model-based MIA
approach
• Segmentation
• Registration
• Shape reconstruction and modeling
• Motion estimation and analysis
Tongue Tracking in Ultrasound
[Kambhamettu et al]
Using an Active Contour Model
Interactive Serial Segmentation of
Biomedical Images
EM neuronal tissue sections
Retinal Angiogram Segmentation
Level-Set Active Contour Model
Gland Segmentation
𝑝1
𝑝2
𝑝3
𝑝4
𝑝5
𝑝 𝑁
Data-Driven Learning of
Statistical Deformable Models
Training data: Numerous expert-segmented images
2N
p
.
x1
yN
0
...
.
.
.......
.
.... ....
...
....
𝑝1
𝑝2
𝑝3
𝑝4
𝑝5𝑝 𝑁
Active Shape Model (ASM)
[Cootes & Taylor, 1992]
Labeled model Trained model Fitting the model
in training image
PC 1
PC 2
Intelligent, Model-Based Medical Image Analysis:
Deformable Organisms
Corpus Callosum Organism
fornix
2
3N-2
1
genu
spleniumrostrum
body
NN-1
upper/right
lower/left
fornix
2
3N-2
1
genu
spleniumrostrum
body
NN-1
upper/right
lower/left
Deformable Organism’s AI Planner
Multiple Deformable Organisms
Driving Forces in Medical Imaging
Exploding data volumes
• Procedure volume growth
• Imaging technology advances
• Manpower shortage
Evidence-based diagnosis
• Early-stage disease screening
• Longitudinal tracking
• Experience shortage
500
25,000
2012 2020
Medical Data (Petabytes)*
14,400+
Possible diagnoses (WHO)
12,000,000+
Misdiagnoses in the US per year**
**Singh et al 2012
Data-Driven,
Machine Learning Approaches
Deep Learning is “revolutionizing”
computer vision and other fields
• It is playing an increasingly important role
in Medical Image Analysis
AI and Deep Learning in Medicine
Automated, accelerated, and accurate insight from
massive medical data
• Lower cost
• Higher efficiency
• Fewer misdiagnoses
Case Study: Lung Cancer
Deadliest cancer worldwide
• One in five cancer deaths is from lung cancer
Early detection is critical
18,000,000+
New lung cancer cases per year
15,000,000+
Deaths from lung cancer per year
49% 45%
30% 31%
14%
5% 1%
0%
25%
50%
75%
IA IB IIA IIB IIIA IIIB IV
5-yr survival by stage
>85%
Diagnosed at late stage
>50%
Die within one year of diagnosis
Deep Convolutional Neural Networks
Applied to Lung Nodule Detection/Analysis
Imaging-Based Medical Analytics and Diagnostics
AI in Medicine
NewMargin
Ventures
Sequoia
CapitalTencent Holdings
Lung Cancer Screening Platform
Imaging
Medical data
Pathology
Web based interface
Secure data storage
Cloud AI engine
Reports and insights
Local archive
Malignancy Assessment with AI
Performance trajectory
• 04/2016: >90% consistency with expert panel
• 06/2016: >85% accuracy vs. ground truth information
• 12/2017: pushing the limits of CT-based diagnosis
Additional Use Cases
Coronary heart disease
Diabetic retinopathy
Conclusion
Data-driven AI has enormous potential in the
entertainment and healthcare industries
• Much initial success of data-driven machine learning approaches to
control advanced biomechanical models of humans and other animals
• The most promising avenue of future innovation in medical imaging is
data-driven machine learning methods working in combination with
powerful model-based image analysis methods
• Much more research must be done to realize the full potential in real-
world industrial applications
Thank You !
Terzopoulos.com

Data-Driven AI for Entertainment and Healthcare

  • 1.
    Data-Driven AI for Entertainmentand Healthcare Demetri Terzopoulos UCLA Distinguished Professor & Chancellor’s Professor of Computer Science Co-Founder & Chief Scientist, VoxelCloud, Inc.
  • 2.
    Visual Computing • Computergraphics (synthesis) • Computer vision (analysis) Talk Overview 1. AI/ML in computer graphics - Human Modeling and Animation 2. AI/ML in computer vision - Medical Image Analysis 3. The future & questions  Images & Videos  Mathematical models Computer Models Images / Videos Computer Vision Computer Graphics
  • 3.
    “Final Fantasy: TheSpirits Within” (Square Pictures, Inc., 2001) Virtual Humans in Movies and Games These characters are neither autonomous nor intelligent “Metal Gear Solid” game
  • 4.
    Motion Capture Technology 3Dtracking of body-attached IR reflectors
  • 5.
    Physics The Artificial LifeApproach Comprehensive computational models of humans and animals • Modeling the body and mind Biomechanics / Locomotion Perception Behavior Learning Cognition The Artificial Life Modeling Pyramid Artificial Intelligence
  • 6.
    Realistic Biomechanical Modelingof the Human Body • Almost all the articular bones and skeletal muscles – 75 bones (165 DOFs), 846 muscles • Volumetric finite element soft tissue model – 354K tetrahedral elements
  • 7.
  • 8.
  • 9.
  • 10.
  • 11.
  • 12.
    How Can weControl Complex, State-of-the-Art Biomechanical Human Models Like These? The natural way is through neuromuscular control • The advanced virtual human models can learn to control themselves like real humans do! – This is accomplished using massive quantities of training data – The training data are synthesized by the human models themselves
  • 13.
  • 14.
    Skeletal System • 7cervical vertebrae and a skull coupled by 3-DOF joints • Ligaments/disks passive joint springs • Equations of motion 0)qb(q,qM(q)   moment arm matrix active muscle force neural inputmass a),q(q,P(q)f)qb(q,qM(q) c   gravity, Coriolis, passive elastic forces
  • 15.
    Biomechanical Neck Model Totalof 72 anatomically-based muscle actuators in 3 layers 48 deep muscles (16 longus colli, 16 erector, 16 rotator) 6 muscles at each joint increase controllability 12 intermediate muscles (scalerius: 4 anterior, 4 posterior, 4 capitis) 12 superficial muscles (2 sternomastoid, 2 cleidooccipital, 8 trapezius) The big challenge is co-actuation and control
  • 16.
    What Would Leonardoda Vinci Think of This?
  • 17.
    Neuromuscular Control ofthe Musculoskeletal Model muscles muscle contraction forces skeletal system environment gravity, applied force bio- mechanical face head pose voluntary controller feedfwd signal setpoint signal proprioceptive feedback (pose, velocity of head) reflex controller muscle activation levels muscle feedback (strain/strain rate) ts tf tu tt s Trained Deep Neural Networks
  • 18.
    • Set randomtarget pose Training the Neural Networks
  • 19.
    • Using inversekinematics, compute desired muscle lengths • Using inverse dynamics, compute muscle activations to achieve desired muscle lengths (under gravity) Training the Neural Networks
  • 20.
    Training the NeuralNetworks Target Pose Activations • Repeat with about 20K random target poses
  • 21.
  • 22.
    Realistic Animation ofSwimming Biomechanical human model immersed in fluid • Advanced multi-physics simulation
  • 23.
    Learning to SwimFrom Examples
  • 24.
    Close-up View ofthe Biomechanical Swimmer Model
  • 25.
    Medical Image Analysis Deformablemodels: A powerful, model-based MIA approach • Segmentation • Registration • Shape reconstruction and modeling • Motion estimation and analysis
  • 26.
    Tongue Tracking inUltrasound [Kambhamettu et al] Using an Active Contour Model
  • 27.
    Interactive Serial Segmentationof Biomedical Images EM neuronal tissue sections
  • 28.
  • 29.
    Level-Set Active ContourModel Gland Segmentation
  • 30.
    𝑝1 𝑝2 𝑝3 𝑝4 𝑝5 𝑝 𝑁 Data-Driven Learningof Statistical Deformable Models Training data: Numerous expert-segmented images 2N p . x1 yN 0 ... . . ....... . .... .... ... .... 𝑝1 𝑝2 𝑝3 𝑝4 𝑝5𝑝 𝑁
  • 31.
    Active Shape Model(ASM) [Cootes & Taylor, 1992] Labeled model Trained model Fitting the model in training image PC 1 PC 2
  • 32.
    Intelligent, Model-Based MedicalImage Analysis: Deformable Organisms Corpus Callosum Organism fornix 2 3N-2 1 genu spleniumrostrum body NN-1 upper/right lower/left fornix 2 3N-2 1 genu spleniumrostrum body NN-1 upper/right lower/left
  • 33.
  • 34.
  • 35.
    Driving Forces inMedical Imaging Exploding data volumes • Procedure volume growth • Imaging technology advances • Manpower shortage Evidence-based diagnosis • Early-stage disease screening • Longitudinal tracking • Experience shortage 500 25,000 2012 2020 Medical Data (Petabytes)* 14,400+ Possible diagnoses (WHO) 12,000,000+ Misdiagnoses in the US per year** **Singh et al 2012
  • 36.
    Data-Driven, Machine Learning Approaches DeepLearning is “revolutionizing” computer vision and other fields • It is playing an increasingly important role in Medical Image Analysis
  • 37.
    AI and DeepLearning in Medicine Automated, accelerated, and accurate insight from massive medical data • Lower cost • Higher efficiency • Fewer misdiagnoses
  • 38.
    Case Study: LungCancer Deadliest cancer worldwide • One in five cancer deaths is from lung cancer Early detection is critical 18,000,000+ New lung cancer cases per year 15,000,000+ Deaths from lung cancer per year 49% 45% 30% 31% 14% 5% 1% 0% 25% 50% 75% IA IB IIA IIB IIIA IIIB IV 5-yr survival by stage >85% Diagnosed at late stage >50% Die within one year of diagnosis
  • 39.
    Deep Convolutional NeuralNetworks Applied to Lung Nodule Detection/Analysis
  • 40.
    Imaging-Based Medical Analyticsand Diagnostics AI in Medicine NewMargin Ventures Sequoia CapitalTencent Holdings
  • 41.
    Lung Cancer ScreeningPlatform Imaging Medical data Pathology Web based interface Secure data storage Cloud AI engine Reports and insights Local archive
  • 42.
    Malignancy Assessment withAI Performance trajectory • 04/2016: >90% consistency with expert panel • 06/2016: >85% accuracy vs. ground truth information • 12/2017: pushing the limits of CT-based diagnosis
  • 43.
    Additional Use Cases Coronaryheart disease Diabetic retinopathy
  • 44.
    Conclusion Data-driven AI hasenormous potential in the entertainment and healthcare industries • Much initial success of data-driven machine learning approaches to control advanced biomechanical models of humans and other animals • The most promising avenue of future innovation in medical imaging is data-driven machine learning methods working in combination with powerful model-based image analysis methods • Much more research must be done to realize the full potential in real- world industrial applications
  • 45.