SLIDE 1 Deep Learning, Lung Cancer Screening & the Data Science Bowl 2017
Bram van Ginneken, Arnaud Arindra Adiyoso Setio, Colin Jacobs
Diagnostic Image Analysis Group, Radboud University Medical Center, Nijmegen, The Netherlands Fraunhofer MEVIS, Bremen, Germany
SLIDE 2 Background and disclosures
- 1996-2001: PhD Detection of tuberculosis in chest radiographs:
CAD4TB now in use in >20 countries
- 2004: Research on automated analysis of chest CT scans for lung cancer
screening, with many partners in Europe, USA, Canada, Korea
- 2010: Chair of Diagnostic Image Analysis Group, 40 researchers at Radboud
University Medical Center, Nijmegen, The Netherlands; working for Fraunhofer MEVIS in Bremen, Germany
- 2010: Developed software for automated reading of chest CT screening scans,
available via Mevis Medical Solutions (Veolity) and InVivo (DynaCAD Lung)
- 2014: Founder of Thirona
- 2016: Involved in Data Science Bowl
SLIDE 3 Contents
- Why should you do lung cancer screening with CT
- Human reading of screening CT scans
- Computer reading of screening CT scans
- 2017 Data Science Bowl results
SLIDE 4 Lung cancer is the biggest cancer killer
Lung cancer can grow unnoticed for years, because the lungs are very big organs When the patient goes to the doctor with complaints and lung cancer is diagnosed, the median size of the cancer is 4 cm, and usually it has already metastasized
SLIDE 5
SLIDE 6
Early detection is our only hope
SLIDE 7
SLIDE 8
SLIDE 9 Screening with CT
- Hold your breath, slide through scanner in <10 seconds
- High resolution (<1mm) 3D image of the lungs
- Very low dose possible
- No contrast material needed, very user friendly
- No high-end scanner needed: cheap (~200 k$)
SLIDE 10
Baseline
SLIDE 11
After 1 year 6mm nodule
SLIDE 12
After 2 years 2cm nodule Squamous cell lung cancer Still early stage!
SLIDE 13
NLST led to screening implementation
SLIDE 14
Screening centers in the United States
SLIDE 15
Human reading of screening CT scans
SLIDE 16
SLIDE 17
Nodule malignancy calculator
SLIDE 18 Radiologists often don’t agree on nodules
- Show table and images from LIDC study...
SLIDE 19
SLIDE 20
SLIDE 21
Some radiologists see many more nodules...
SLIDE 22
Radiologist often don’t agree on nodule type
SLIDE 23 What type of nodule is it?
Complete agreement non-solid nodule Complete agreement solid nodule
SLIDE 24 What type of nodule is it?
Complete disagreement part-solid nodule Complete disagreement part-solid nodule
SLIDE 25
Lung-RADS 4X
SLIDE 26
The Lung-RADS 4X category in practice
SLIDE 27
The Lung-RADS 4X category in practice
SLIDE 28
Computer reading of screening CT scans
SLIDE 29 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.01 0.10 1.00 10.00 100.00 Sensitivity Average number of false alarms per scan
All together Dayton Pisa Utrecht Torino Bremen Florianópolis Bari Hamburg
2009: Finding nodules with computers
SLIDE 30
2017: Finding nodules with deep learning
SLIDE 31
www.grand-challenge.org
SLIDE 32
LUNA16: Nodule detection competition
SLIDE 33
LUNA16 leaderboard
SLIDE 34 Nature Scientific Reports 2017
Deep learning for nodule type classification
SLIDE 35
SLIDE 36
SLIDE 37
SLIDE 38 Case from the Danish Lung Cancer Screening Trial
May 2005: 640 mm3 / 10.7mm;
Jan 2010: 1042 mm3 / 12.5mm;
Automatic segmentation
SLIDE 39
Data Science Bowl 2017 Results
SLIDE 40 Data Science Bowl 2017
- Teams were provided with a single low dose CT scan from a lung cancer screening participant
- Task: predict if this subject was diagnosed with lung cancer within one year of the CT scan
- Approach of most teams:
- Look for suspicious nodules
- Estimate probability of malignancy for the most suspicious nodule(s)
SLIDE 41
SLIDE 42 ROC analysis Data Science Bowl test set
Lung-RADS category Non-cancer cases Cancer cases
1 66 2 2 138 4 3 28 10 4A 63 68 4B 54 67
4B 4B,4A 4B,4A,3,2 4B,4A,3
SLIDE 43 ROC analysis Data Science Bowl test set
Lung-RADS category Non-cancer cases Cancer cases
1 66 2 2 138 4 3 28 10 4A 63 68 4B 54 67
4B 4B,4A 4B,4A,3,2 4B,4A,3
SLIDE 44 ROC analysis Data Science Bowl test set
4B 4B,4A 4B,4A,3,2 4B,4A,3
SLIDE 45 ROC analysis Data Science Bowl test set
4B 4B,4A 4B,4A,3,2 4B,4A,3
SLIDE 46 Conclusions and future work
- DSB 2017 provides very promising results
- In lung cancer screening the vast majority of all scans are follow-up scans: estimating probability of lung
cancer given a single CT scan is therefore not the most important and most relevant problem
- Future competitions:
- Estimate probability of lung cancer given multiple CT screening rounds
- Extend the problem, taking into account clinical information
- Add liquid biopsy results (blood tests), genetic tests, etc
- Predict 5-year probability of lung cancer: personalized screening intervals
