1. Applications of AI in Radiology
Boris Geyzer

2. About me Started in medical imaging commercially in 2002
Wholly owned subsidiary of Canon Medical Systems focused on DR technology
Product management in digital imaging
Channel management
Shifted focus to Interventional Radiology – Omega Medical Imaging
ERCP imaging
Cardiac Cath and Electrophysiology
Currently Sales Director for MID – exclusive partners for Samsung throughout USA
Live in Florida (on the weekends) with wife, April, and son, Elijah

3. What is AI?
Artificial Intelligence (AI) The theory and development of computer systems able to perform tasks that normally require human intelligence.

4. History of AI Alan Turing
“Godfather” of theoretical computer science and artificial intelligence.
Cryptanalyst who worked during WWII to break German “Enigma” encryption
Died June 7, 1954 of cyanide poisoning at age of 41

5. Source: Paul Marsden, Digital Intelligence Today

6. AI Roadblocks

7. Moore’s Law
Observation made by Gordon Moore (co-founder of Intel) in 1965, that number of transistors per square inch on integrated circuits double every 2 years.
Transistors
Computations

8. Moore’s Law

9. AI Roadblocks

10. Cost of Computing

11. Cost of memory
How much would your current daily PACS storage usage cost in the 80’s?
Modality
Avg File Size
# of Patients
Storage Needed
Cost DR
25MB 50
1.25GB
$424,750 CT
200MB 20
4.00GB
$1,359,200 MR UL
50MB 30
1.50GB
$509,700
DBT
1GB
30.00GB
$10,194,000
Total Storage
Total Cost
Daily
40GB
$13,846,850
Annual
14.6TB
$5,054,100,250

12. Cost of memory
How much would your current daily PACS storage usage cost in the 80’s?
Modality
Avg File Size
# of Patients
Storage Needed
Cost DR
25MB 50
1.25GB
$424,750 CT
200MB 20
4.00GB
$1,359,200 MR UL
50MB 30
1.50GB
$509,700
DBT
1GB
30.00GB
$10,194,000
Total Storage
Total Cost
Daily
40GB
$13,846,850
Annual
14.6TB
$5,054,100,250

13. Time keeps on slippin’…into the future
Cost of computing decreased
Cost of storage decreased
More access to PCs throughout community
Internet more available
How it applied to medical imaging
New applications abound
PACS
HIS/RIS
Teleradiology

14. Early AI in the news

15. Deep Blue vs. Kasparov
First tournament (1996) – Kasparov won 5 out of 6 against Deep Blue (early version)
REMATCH tournament (1997): Kasparov wins 1, Deep Blue wins 2, and tie 3X; Deep Blue considered winner
Difference between 1996 and 1997 – IBM doubled processing power of system with new hardware; increased logic in ‘chess chip’.
Deep Blue was able to search between 100 and 200 million positions per second; searching between 6 and 8 pairs of moves to a maximum of 20 pairs.

16. Moving to Big Data

17. Moving to Big Data Definition:
extremely large data sets that may be analyzed computationally to reveal patterns, trends, and associations.
Murray Campbell, engineer on Deep Blue, on how AI today is different now compared to 20 years ago:
“People also started gathering—just as part of their business—a lot more data that provided fodder for the machine-learning algorithms of the day. Eventually we started realizing that combining all these things could produce some remarkable results.”

18. AI needs Structured Data
Source: Monica Rogati, Data Science Advisor, hackernoon.com, August 1, 2017

19. Structured Data
Source: EMC

20. Structured Data - Examples

21. Structured Data - Examples

22. Algorithms + Last component... COMPUTING POWER INEXPENSIVE STORAGE
STRUCTURED DATA
COMPUTING POWER +
INEXPENSIVE STORAGE
Algorithms

23. Algorithms Definition:
a process or set of rules to be followed in calculations or other problem- solving operations

24. Algorithms Machine Learning Deep Learning
The practice of using algorithms to parse data, learn from it, and then make a determination or predication about something in the world.
Machine is “trained” using large amounts of data and algorithms to ‘learn’ the task.
“Computer Vision” most applicable – OCR, shape and edge detection
Deep Learning
Artificial Neural Networks algorithms to continually train the computer itself similar to how human brains function.
Continually improving

25. Deep Learning AI

28. AI in Radiology

29. Your next radiologist? Probably not!
Diffusion weighted images are normal, indicating absence of any acute ischemic lesion….
Probably not!

30. Why the uproar... Professor Geoffrey Hinton
Godfather of neural networks
Stated:
We should stop training radiologists
Within next 5 years (maybe 10), Deep Learning image analysis will be better than radiologist abilities
“Like the coyote already over the edge of the cliff who hasn’t yet looked down”.

31. The reality
To replace Radiologists, Artificial General Intelligence (AGI) is needed
AGI
Human level intelligence achieved by AI
Be able to:
Reason
Plan
Solve problems
Think abstractly
Comprehend complex ideas
Learn quickly
Learn from experience
Study conducted at annual AGI conference of when people believe AGI would be achieved:
By 2030: 42%
By 2050: 25%
By 2100: 20%
After 2100: 10%
Never: 2%

32. The reality Thousands of published research articles since 2017
Each are individual for a specific focus
Implementation is key
Published articles are not commercial applications
No ability to conglomerate and associate all techniques into one field

33. Grid Scatter Correction
Examples in Radiology
Detection
Workflow
Grid Scatter Correction
Mammography CAD

34. Why this matters? AIding Detection
*A recent US nationwide research on malpractice suits [2] showed that the most common cause of medical malpractice suits against radiologists was error in diagnosis (mainly failure to diagnosis instead of delay).
Commercially available AI tools in aiding detection:
Mammography CAD
Lung CAD
Bone Suppression Imaging
Radiographic Line/Tube Image Enhancement
Ischemic Stroke detection
More coming!
Why this matters?
Source: J. S. Whang, S. R. Baker, R. Patel, L. Luk, and A. Castro III, “The causes of medical malpractice suits against radiologists in the United States,” Radiology, vol. 266, no. 2, pp. 548–554, 2013.

35. AIding Detection
High Physician Concern About Malpractice Risk Predicts More Aggressive Diagnostic Testing In Office-Based Practice
Source:
Estimating Cost Savings from Early Cancer Diagnosis
Source:
Improving patient outcomes:

36. AIding Detection Bone Suppression:
“The results also showed that radiologists were more confident in making diagnoses regarding the presence or absence of an abnormality after rib-suppressed companion views were presented”
Source: Proceedings Volume 9037, Medical Imaging 2014: Image Perception, Observer Performance, and Technology Assessment; 90370D (2014); doi: /

37. AIding Detection Lung CAD:
“CAD has the potential to equalize performance among readers by reducing individual detection errors of lung nodules on chest CT.”
Source:

38. AIding Workflow Line Enhancement Software:
“Tube and line interpretation in portable chest radiographs was assessed using a new visualization method. When using the new method, radiologists' interpretation time was reduced by 30% vs. standard modality processing and window and level (23 vs. 33 s).”
Source: Foos DH, Yankelevitz DF, Wang X, Berlin D, Zappetti D, Cham M, Sanders A, Novak Parker K, Henschke CI. Improved visualization of tubes and lines in portable intensive care unit radiographs: a study comparing a new approach to the standard approach. Clinical Imaging; Volume 35, Issue 5, September–October 2011, Pages 346–352.

39. AI systems working together
Oncology
Pulmonology
Radiology
Cardiology

40. AI systems working together
AI systems, integrated together can potentially:
Help radiologist identify the patient nodule earlier with advanced imaging technologies (at the modality) and with advanced detection algorithms (at the reading station)
Help oncologist give more precise outcome statistic based on big data available from NIH and other large datasets
Using genetic testing and AI to see which drug deliveries may be most effective based on patient demographic
Help physicians practice medicine more confidently

41. AI shaping the future – for the patient
Personalized Treatment Plan and Outcomes
“Big Data” Pathology and Statistical Correlation
Improved Detection and categorization

42. The future for Radiologists?
Possible Pneumothorax

43. Thank You!