NCCR 2020: Conference Of Very Important Disease (COVID-19) | 24 - 26 August 2020
Young Investigator Awards Presentation
Kim-Ann Git1, Aida binti Abdul Aziz2, Lau Kiew Siong3, Lau Song Lung3, Preetvinder Singh a/l Dheer Singh4, Tan Ying Sern5, Eric Chung6
1-Selayang Hospital
2-Sungai Buloh Hospital
3-Sarawak General Hospital
4-Hospital Raja Permaisuri Bainun
5-Taiping Hospital
6-University of Malaya Medical Centre
https://doi.org/10.5281/zenodo.4004461
Ahmedabad Call Girls Book Now 9630942363 Top Class Ahmedabad Escort Service A...
Multi-Centre Optimization and Validation of an Open Deep Learning Model for COVID-19 Detection on Chest Radiographs
1. Multi-Centre Optimization and Validation of an
Open Deep Learning Model for
Covid-19 Detection on Chest Radiographs
Kim-Ann Git1, Aida binti Abdul Aziz2, Lau Kiew Siong3, Lau Song Lung3,
Preetvinder Singh a/l Dheer Singh4, Tan Ying Sern5, Eric Chung6
1Selayang Hospital 2Sungai Buloh Hospital 3Sarawak General Hospital 4Hospital Raja Permaisuri Bainun
5Taiping Hospital 6University of Malaya Medical Centre
NMRR-20-949-54941
5. Hospital A
Introduction
COVID-Net
? Model Generalizability
Model
Hospital B
Model
Hospital A & B
Model
Hospital C
Model
Model Optimization
Hospital A
Model
Hospital B
Model
Retrain Model
Sensitivity & specificity around 90%
Potential Solutions
6%
94%
External Validation
Used
Not used
6. Objective
To optimize and validate COVID-Net for the prediction of
Covid-19 in CXRs of Malaysian patients
To optimize COVID-Net with
additional training using local
CXRs
To validate and compare the
accuracy of the published
(unoptimized) model with the
optimized model using local CXRs
General Objective
Specific Objectives
7. Methodology
Convenient sampling from each hospital:
Cases: up to 100-150 Covid PCR+ CXRs
Controls: up to 150-200 CXRs from October 2019
Training Set:
318 Cases & 368 Controls
Test Set:
318 Cases & 367 Controls
Data Collection:
Hospital Sungai
Buloh
Sarawak General
Hospital
Hospital Raja
Permaisuri
Bainun
Hospital Taiping
University of
Malaya Medical
Centre
Optimization Validation
Published COVID-
Net Model
(Unoptimized)
Published COVID-
Net Model
(Unoptimized)
Optimized COVID-
Net Model
Optimized COVID-
Net Model
vs
8. Pre-Optimization Validation
Unoptimized (4A)
Prediction
Normal Pneumonia Covid
Ground
Truth
Normal 78 3 235
Pneumonia 1 16 34
Covid 96 11 211
Published (Unoptimized) COVID-Net Model
44.5%
66.4%
Accuracy (All) Sensitivity (Covid-19)
Published (Unoptimized)
9. Optimization
• Training Parameters:
• Image augmentation
• Batch size: 4
• Optimizer: Adam
• Variable learning rate: 0.0001 to 0.00003
• Epochs at each learning rate: 40+
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 10 20 30 40 50 60 70 80 90 100
Epochs
loss accuracy
10. Post-Optimization Validation
Unoptimized (4A)
Prediction
Normal Pneumonia Covid
Ground
Truth
Normal 78 3 235
Pneumonia 1 16 34
Covid 96 11 211
Published (Unoptimized) COVID-Net Model
44.5%
66.4%
77.7%
83.0%
85.3%
92.5%
Accuracy (All) Sensitivity (Covid-19)
Published (Unoptimized)
Partially Optimized
Optimized
90+ Epochs
LR: 0.0001 → 0.00003
Prediction
Normal Pneumonia Covid
Ground
Truth
Normal 269 6 41
Pneumonia 12 21 18
Covid 19 5 294
Optimized COVID-Net Model
20+ Epochs
LR: 0.0001
Prediction
Normal Pneumonia Covid
Ground
Truth
Normal 251 7 58
Pneumonia 13 17 21
Covid 45 9 264
Partially Optimized COVID-Net Model
13. Conclusion
The published (unoptimized) COVID-Net model has mediocre
performance which improved tremendously after optimization.
Model is robust but requires optimization with local CXRs before
consideration for clinical use.
Grad-CAMs do not reflect disease distribution. Further work is
required for model explainability.
14. Limitations
Limited GPU for optimization
→ Cloud Machine Learning Platforms
Small number of centres
→ Ongoing NIH-Crest Collaborative Covid-19 Chest Radiograph (NC4R)
Repository Project
16. Acknowledgements
Dr Aida binti
Abdul Aziz
Dr Lau Song
Lung
Dr Lau Kiew
Siong
Dr Preetvinder
Singh
Dr Eric
Chung
Dr Tan Ying
Sern
The authors would also like to thank the
Director-General, Ministry of Health, for
permission to present this paper.
Dr Alexander
Wong
The authors would
like to thank
Alexander Wong of
the Vision and
Image Processing
Lab, University of
Waterloo, for his
guidance on model
optimization and
explainability.
Team Members Special Acknowledgement
17. References
• Ai, T. Y. (2020). Correlation of chest CT and RT-PCR testing in coronavirus disease 2019 (COVID-19) in China: a report of
1014 cases. Radiology.
• Chan, J. Y. (2020). Improved molecular diagnosis of COVID-19 by the novel, highly sensitive and specific COVID-19-
RdRp/Hel real-time reverse transcription-PCR assay validated in vitro and with clinical specimens. Journal of Clinical
Microbiology.
• Hellewell, J. A. (2020). Feasibility of controlling COVID-19 outbreaks by isolation of cases and contacts. The Lacet Global
Health.
• Johns Hopkins University. (2020, 05 08). Coronavirus Resource Center. Retrieved from Coronavirus Resource Center:
https://coronavirus.jhu.edu/map.html
• Wang, L. a. (2020). COVID-Net: A tailored deep convolutional neural network design for detection of COVID-19 cases
from chest radiography images. arXiv preprint.
18. Multi-Centre Optimization and Validation of an
Open Deep Learning Model for
Covid-19 Detection on Chest Radiographs
Kim-Ann Git1, Aida binti Abdul Aziz2, Lau Kiew Siong3, Lau Song Lung3,
Preetvinder Singh a/l Dheer Singh4, Tan Ying Sern5, Eric Chung6
1Selayang Hospital 2Sungai Buloh Hospital 3Sarawak General Hospital 4Hospital Raja Permaisuri Bainun
5Taiping Hospital 6University of Malaya Medical Centre
NMRR-20-949-54941
Thank You for Your Attention!