Presenter: Konstantin Pogorelov, Simula Research Laboratory, University of Oslo, Norway
Paper: http://ceur-ws.org/Vol-1984/Mediaeval_2017_paper_3.pdf
Video: https://youtu.be/V2vFNXKSFrM
Authors: Michael Riegler, Konstantin Pogorelov, Pål Halvorsen, Carsten Griwodz, Thomas de Lange, Kristin Ranheim Randel, Sigrun Losada Eskeland, Duc-Tien Dang-Nguyen, Mathias Lux, Concetto Spampinato
Abstract: The Multimedia for Medicine Medico Task, running for the first time as part of MediaEval 2017, focuses on detecting abnormalities, diseases and anatomical landmarks in images captured by medical devices in the gastrointestinal tract. The task characteristics are described, including the use case and its challenges, the dataset with ground truth, the required participant runs and the evaluation metrics.
MediaEval 2017 - Medical Multimedia Task: Multimedia for Medicine: The Medico Task at MediaEval 2017 (Overview)
1. Multimedia for Medicine:
The Medico Task
Michael Riegler, Konstantin Pogorelov, Pål Halvorsen, Carsten Griwodz,
Thomas de Lange, Kristin Ranheim Randel, Sigrun Losada Eskeland,
Duc-Tien Dang-Nguyen, Mathias Lux, Concetto Spampinato
Email: michael@simula.no, konstantin@simula.no
2. Medico
Bringing together IT and medicine, focusing on detecting
abnormalities, diseases and anatomical landmarks in images captured
by medical devices in the gastrointestinal tract
Goals of the task: Help to improve the health care system using
multimedia methods to reach the next level of multimedia-assisted
computer diagnosis, detection and interpretation of abnormalities
Attract more researchers to medical use cases
3. GI Tract Challenges
Many types of diseases can affect the human digestive system
Screening of the gastrointestinal (GI) tract using different
types of traditional endoscopy
– is costly (colonoscopy: US - $1100/patient, $10 billion dollars)
– consumes valuable medical personnel time (1-2 hours)
– does not scale to large populations
– is intrusive to the patient
Current developments in technology may potentially enable
automatic algorithmic screening and assisted examinations
a true interdisciplinary activity with high chances of societal impact
4. Colorectal Cancer Mortality 2012
Most common cancer for men in Norway
Second most common cancer for women in Norway
5. Live Automatic Detection
System to assist doctors during live
traditional endoscopy procedures
Second pair of eyes
Support for inexperienced doctors
Automatic tagging of lesions
Automated reports generation
Better procedure documentation
[1] van Rijn, J. C., Reitsma, J. B., Stoker, J., Bossuyt, P. M., van Deventer, S. J., and Dekker, E.
Polyp miss rate determined by tandem colonoscopy: a systematic review.
The American journal of gastroenterology 101, 2 (2006)
6. Wireless Video Capsule (Capsular Endoscopy)
better scale
less intrusive
possible to combine
examinations!?
less expensive?
(detection might lead to an endoscopy)
expensive
does not scale
intrusive
7. Our Goals
A complete system for:
– Live Traditional Endoscopy
– Capsular Endoscopy
Automatic detection of different abnormalities in the
digestive system
– HD sources
– Real-time and faster
– High recall and precision
– Automated reports generation
9. Common Challenges
Blurry images due to camera motion
Objects too close to camera
Under or over scene lighting
Flares
Artificial objects and natural "contaminants“
Low resolution of Capsular Endoscopes
No proper support for medical reports generation
11. Subtasks
Detection
Detection of different diseases and landmarks
Use as few images in the training dataset as possible
Efficient detection
Solve the classification problem as fast as possible
Efficient detection (fare evaluation, the same hardware)
Evaluate detection algorithms on the same hardware
Report generation (Experimental)
Automatically create a text-report for a medical doctor for three
video cases
12. Dataset
Hard to find annotated data allowed to use (even) for research
Two data sets have been published at the MMSys 2017 open
dataset track: Kvasir and Nerthus (bowel preparation quality)
Kvasir: the open source dataset consist of 8,000 annotated GI tract
images in 8 different classes: 500 in dev. + 500 in test set per class
13. Participants and submissions
12 teams registered and requested the data
5 teams successfully submitted the results
Austria, China, India, Italy, Norway, Pakistan, US…
46 submissions in total
Detection – Successful
Efficient detection – Partially Successful
Efficient detection – No-Show
Report generation – No-Go
14. The Approaches
Machine Learning
– Manifold learning
– Support vector machine
– Random forest and random tree
– Logistic model tree
– Regression-based
– Unsupervised clustering
Convolutional Neural Networks
– Direct classification
– Transfer learning
Features Extraction
– Deep features
– Global and Local features
15. Metrics
Multi-class generalization of Matthews correlation coefficient (R_K
statistic)
– Correlation coefficient between the observed and predicted
classifications
– Perfect for unbalanced multi-class case
– Maximum value +1 for perfect prediction
– Minimum value between -1 and 0 depending on the true
distribution
• Frames per second (FPS)
16. The Results – Best Performing
Team Training
set size
Multi-class
MCC
F1 FPS
SCL-UMD 3200 0.827 0.848 1.3
FAST-NU-DS 4000 0.736 0.767 2.3
ITEC-AAU 3600 0.724 0.755 1.4
HKBU 1000 0.663 0.703 2.2
SIMULA 4000 0.802 0.826 46.0
Random - -0.001 0.124 -
ZeroR - 0 0.125 -
18. The Results – Smallest Training Set
Team Training
set size
Multi-class
MCC
F1
ITEC-AAU 400 0.607 0.649
FAST-NU-DS 732 0.649 0.689
HKBU 800 0.648 0.692
SCL-UMD 3200 0.827 0.848
SIMULA 4000 0.802 0.826
Random - -0.001 0.124
ZeroR - 0 0.125
19. Conclusions
1. Good classification performance achieved
2. Small amount of training data would not stop us!
3. Real-time is still challenging
4. Combined approaches are required for medical image
analysis
5. Medical task is interesting for the community
20. Future of the task
1. Localization/segmentation of findings and lesions
2. Exploiting domain expert knowledge – more data!
3. Integration of various data, multi-modality – new sensors,
doctors’ records, audio recordings, patient context
information
4. Automated reporting
21. Thank You! Questions?
All the data is released publicly and available at:
http://datasets.simula.no/kvasir/