These are my #AI slides for medical deep learning using #radiology and medical imaging examples. Please use them & modify to teach your own group about medical AI.
4. “…mobile computing, the proliferation of inexpensive sensors collecting
terabytes of data, and the rise of machine learning that can use that data,
will fundamentally change the way the global economy is organized.”
Fortune, “CEOs: The Revolution is coming” March 8, 2016
Reference: March 2016 issue of Fortune: “Oracles of Davos.”
6. Deep Learning
Machine Learning
Artificial Intelligence
Automated
Video Game
“Player”
Decision tree
algorithm: Are
we going to
play tennis?
Using a neural
network for
expert systems:
Does the data in
the CT scan
mean your
patient will live?
7. • Paucity of high-quality data.
• Pooling of resources is needed.
• Big data sets necessitate
wholesale anonymization for
integration.
• What is non-identifiable?
• Blockchain your semantically
annotated big data and promote
service interoperability?
Machine Learning can use a Decision Tree
Tennis example from: Learning from Data: Decision Trees. Amos
Storkey, School of Informatics University of Edinburgh
8. • Paucity of high-quality data.
• Pooling of resources is needed.
• Big data sets necessitate
wholesale anonymization for
integration.
• What is non-identifiable?
• Blockchain your semantically
annotated big data and promote
service interoperability?
Tennis example from: Learning from Data: Decision Trees. Amos
Storkey, School of Informatics University of Edinburgh
9. Deep Learning
A set of rules
Intended to execute an operation
With a level of independence from humans
10. Deep Learning
A set of rules
Intended to execute an operation
With a level of independence from humans
11. The Rules
Not a single entry
based on what one
person likes (e.g.
humidity). It is
based on a large set
of data that is used
to train the
computer to make a
set of rules with the
lowest error (cost). Andrej Karpathy & Li Fei-Fei
Dept of Computer Science, Stanford University
12. The Rules
I love tennis.
The first thing I see
here is this logo and
the scoreboard!
Was the computer
trained to see it?
13. Steps to Generate Rules
• Get data
• Clean, Prep, Manipulate data
• Train model
• Test model
• Improve
14. Steps to Generate Rules
• Get data
• Clean, Prep, Manipulate data
• Train model
• Test model
• Improve
15. Steps to Generate Rules
• Get data
• Clean, Prep, Manipulate data
• Train model
• Test model
• Improve
16. Steps to Generate Rules
• Get data
• Clean, Prep, Manipulate data
• Train model
• Test model
• Improve
17. Steps to Generate Rules
• Get data
• Clean, Prep, Manipulate data
• Train model
• Test model
• Improve
playground.tensorflow.org
18. TensorFlow Crash Course and Playground
If you are dedicated to learn, you can do it!
• Label (y): This is the yes/no truth: Is this a lung cancer
that will respond to a new immunotherapy
• Features (xi): The inputs. The scan data (not just a
report), clinical information, lab data
• Examples (x): The patients, each patient has features
(xi). Labeled examples are (x, y) where y is truth
• Model: What is doing the predicting. The model
outputs (x,y’). y’ is the guess of whether or not this is
lung cancer that will respond to new immunotherapy
22. Our task is different, because what we do is different.
globallymealliance.org
Journal of Thoracic and Cardiovascular
Surgery. 2004; 128(5):761-2
23.
24. US FDA Commissioner Scott Gottlieb
“AI holds enormous promise for the future of medicine.” The FDA is
working on a “new regulatory framework” to allow regulators (e.g.
FDA employees) to “promote innovation in this space”… “We expect
to see an increasing number of AI-based submissions in the coming
years, starting with medical imaging devices, and we’re working with
experts in the field.”
25. Deep Learning
A set of rules
Intended to execute an operation
With a level of independence from humans
26. The brain and its neurons
• Logical
– Study protocolling
– Information integration
– Report QA/Dictation Assistance
– Quality Improvement Support
• Image-Based
– Diagnosis
– Segmentation
– Super-Resolution
– Workflow prioritization
– Automated lesion follow-up
– Image-based molecular pathologymy-ms.org Univ of Melbourne
32. Can we automate Tb detection just inside the Arctic circle?
Photos from Iqaluit, Nunavut, 2016
33. Training for human intelligence and
artificial intelligence is completely
different. Is it astounding that the
two methods can have similar
results?
There are several challenges: training
Curry International Tb Center
34. Supervised versus Unsupervised Learning
Applications for medical AI
have used supervised
learning
Deep clinical impact of AI
will come with
unsupervised learning
gocitrusnow.com
35. Supervised versus Unsupervised Learning
Applications for medical AI
have largely used
supervised learning
Deep clinical impact of AI
will come with
unsupervised learning
38. There are big opportunities
1. If there are fewer radiologists, they
will be more informed will make
healthcare more efficient and
economical
2. AI will benchmark a new set of
human “loss functions”
3. Seeing different solutions will enable
us to think creatively, and better
position those with creativity
4. AI can only enhance natural
intelligence, it can’t decrease it
osirix-viewer.com
39. Last example: Contrast in blood vessels from
first-pass CT angiography
Enhancing Natural Intelligence
40. Angiogram
Can We Assess Blood Flow?
1. Contrast Enhancement
2. Computational Fluid Dynamics
Perfusion
Lesion
Hemodynamic
Significance
Conduit
Blood
Flow
Organ
Pathophysiology Accessible by Coronary CTA
41. 300 HU
238 HU
0
100
200
300
400
0 2 4 6 8 10
MeanHU
Distance From Ostium (cm)
@0cm
@9.5cm1 mm intervals (not to scale)
0
100
200
300
400
0 2 4 6 8 10
MeanHU
Distance From Ostium (cm)
1 mm intervals (not to scale)
Mean HU = -8.3×Distance + 316.1
r = 0.845
0
100
200
300
400
0 2 4 6 8 10
MeanHU
Distance From Ostium (cm)
HU at
ostium
HU at 2.5 mm
diameter
300 HU 238 HU
Steigner ML et al. Circ Cardiovasc Imaging. 2010;3:179-186
Contrast Gradients (TAG)
42. Normalized 64 CT Differences in Correlate with TIMI Flow
Chow BJW et al. CT Contrast Opacfication Predicts Coronary Flow. JACC, 2011(57):1280-1288
43. Relationship Between
TAG320 and FFR
“..TAG320 assessment significantly
improves both sensitivity and
specificity to detect clinically relevant
reduction in FFR.”
Some say it is good
46. Let the computer do the talking…
• Logical
– Study protocolling
– Information integration
– Report QA/Dictation Assistance
– Quality Improvement Support
• Image-Based
– Diagnosis
– Segmentation
– Super-Resolution
– Workflow prioritization
– Automated lesion follow-up
– Image-based molecular pathology
47. N=254 (484 vessels) analyzed for all features. Highest predictive value
for contrast density difference
Machine learning for integrating plaque features
to predict ischemia
Dey et al. European Radiology 2018
48. Break-down of Radiology Applications
• Text-Based
– Study protocolling
– Integrating clinical data
– Report QA/Dictation Assistance
• Image-Based
– Prioritize cases and Diagnoses
– Segmentation
– Super-Resolution
– Prioritize cases and Diagnoses
- Automated lesion follow-up
– Image-based molecular pathology
imgkid.com
50. Summary
1. AI healthcare applications will follow the routine use in
our day-to-day lives
2. Deep learning will enhance our natural learning. It can’t
decrease it.
3. Please don’t confuse science fiction with medical care.
The reality for radiology is that computers have helped us
be better doctors at every step of the way. This is just one
of those “big steps”.
Hinweis der Redaktion
This is the talk that I am not giving you! Please login to Linked In and you can download my slides. Feel free to use any slide that I have made in any of your own talks. Every slide from every talk I have given (~500) is open for you to use. Email me at frybicki@toh.ca.
ABB – industrial robots
The last statement is probably a way forward, as we have discussed in this lecture…
Radiomics – using all radiologic signatures available from the images
Radiogenomics – taking all the radiology and integrating with the genomics and seeing how each intersects with the genetic features
The last statement is probably a way forward, as we have discussed in this lecture…
Chow showed that differences correlate with TIMI flow (do not say yet as this comes later: that this is with the aorta normalization that he developed).
In the coming short term, radiologists will be inundated with various applications of AI in clinical practice. The question then becomes – how do we know which applications to trust? How do we assess quality? How do we ensure that an AI application is applicable to our instruments, patients, and datasets?