Presented work is accepted at Korean domestic conference for Medical AI, Korean Society of Artificial Intelligence in Medicine (KOSAIM) 2020.
Special Thanks to Dongmin Choi, the first author and presenter of this work.
(Link to Dongmin Choi Bio: https://www.slideshare.net/DongminChoi6/)
- by Seunghyun Hwang (Yonsei University, Severance Hospital, Center for Clinical Data Science)
Bajaj Allianz Life Insurance Company - Insurer Innovation Award 2024
Diagnosis of Maxillary Sinusitis in Water’s view based on Deep learning model
1. Diagnosis of Maxillary Sinusitis in Water’s view
based on Deep learning model
Dongmin Choi, Seunghyun Hwang, Minkyu Kim,
Sangwook Cho, Byunghan Jang, HwiYoung Kim
Yonsei University, CCIDS
3. Introduction
- Rhinosinusitis is the most prevalent disease in sinonasal
inflammatory disorder.
- CT usually gives the best overall anatomic detail of the
paranasal sinuses. Waters’ views often employ as the first-line
investigation for sinusitis (sensitivity: 0.76 / specificity: 0.79)
Rhinosinusitis
Kim Y, Lee KJ, Sunwoo L, Choi D, Nam CM, Cho J, Kim J, Bae YJ, Yoo RE, Choi BS, Jung C, Kim JH. Deep Learning in Diagnosis of Maxillary Sinusitis Using Conventional Radiography. Invest Radiol. 2019 Jan;54(1):7-15.
Waters’s view
→ If the accuracy of X-ray waters’ view can be
improved by deep learning, patients will not take CT
examination with higher radiation and more cost.
4. Introduction
- Kim Y et al.[1]
proposed majority decision algorithm to
determine a reasonable consensus using three multiple
convolutional neural network (CNN) models: VGG-16,
VGG-19, and ResNet-101. Accuracy of algorithm is high,
but the external test dataset in multiple medical centers did
not be included for reproducibility.
Deep learning approaches
[1] Kim Y, Lee KJ, Sunwoo L, Choi D, Nam CM, Cho J, Kim J, Bae YJ, Yoo RE, Choi BS, Jung C, Kim JH. Deep Learning in Diagnosis of Maxillary Sinusitis Using Conventional Radiography. Invest Radiol. 2019 Jan;54(1):7-15.
[2] Kim HG, Lee KM, Kim EJ, Lee JS. Improvement diagnostic accuracy of sinusitis recognition in paranasal sinus X-ray using multiple deep learning models. Quant Imaging Med Surg.
- Kim HG et al.[2]
compared the diagnostic performance of
convolutional neural network in diagnosing maxillary sinusitis
on Waters’ view radiograph with those of five radiologists.
The AUC score was 0.93 and they argued that CNN can
diagnose as accurately as the expert radiologists.
Class activation maps for
sinusitis (left images) and normal (right images) case
5. Introduction
Both are normal Both have mucosal thickening
Both have air fluid Haziness in right side
https://github.com/KYBiMIL/KHD_2020/blob/master/KHD_2020_PNS.pdf
8. Dataset
Diseases Training set Test set Percentage(%)
Normal 2,500 250 78%
Mucosal thickening 360 40 13%
Air fluid 150 20 6%
Haziness 90 10 3%
Total 3,100 training dataset and 320 test dataset.
The source of the entire dataset is KONYANG University Hospital
This work was supported by 'Breast Cancer Pathology and Sinusitis X-Ray Data Collection Project for Artificial Intelligence
Applications(2020)' funded by the National Information Society Agency(NIA, Korea) & the Konyang University
Hospital(KYUH, Korea).
9. Model & Parameters
Structure of ResNet
K He et al. Deep Residual Learning for Image Recognition. CVPR 2016
ResNet-34 is trained for total 60 epochs with Stochastic Gradient Descent
optimizer, which used weight decay of 0.3 and batch size of 8.
The initial learning rate was 0.0005 and decayed by cosine annealing.
11. Discussion & Conclusion
- We achieved 0.94 of weighted F1 score in test set
- Our proposed model shows the outstanding performance for
diagnosing various types of sinusitis.
- We overcame data imbalance problems through adjustment of
class weights and successfully classified sparse classes.
- Future works
1. Validate on external dataset
2. improve performance for Mucosal thickening