Deep Learning Model for Automated Detection and Classification of Central Canal, Lateral Recess, and Neural Foraminal Stenosis at Lumbar Spine MRI
; RSNA Radiology article, 11 May 2021, citation 78
1. Motivation
; RSNA 2024의 주제와 유사하기 때문에 추후 RSNA 2024 competiton에 참여한다면 도움이 될 것 같아 선택하게 됨
2. Intro
ⅰ) limitations of previous work
- only binary classification ( absent / present )
- only focus on L4-L5 level
- weakly supervised label ( extracted from radiologist report ; nlp )
U-Net Segmentation; W/o ROI crop
sagital & axial scan combine(concat)
+) how to deal with imbalance data ;weighted categorial cross entropy
ⅱ) Goal of this paper(article)
-
no DL model has been developed
to assess stenosis at all three regions of interest (ROIs)
along the lumbar spine -
ROI Detection ( left, center, right )
; roi 영역 외 noise 또는 상관없는 패턴이 학습에 영향 줄까봐 crop 진행 -
DL model was developed to
automatically detect and classify central canal, lateral recess,
and neural foraminal stenosis in the lumbar spine using axial
and sagittal MRI sequences
3. Method
ⅰ) data preperation
- Sagittal T1 & Axial T2
- internal - 3 years, Singapore ( train - 2 radiologist / test - 3 radi.. )
- external (inference) - 1 year, Saudi ( 1 radi.. )
ⅱ) DL archietecture
- Tried 9 ver DL ; no change in medel archietecture but different train/val data set -> use avg dl model(voting?) for comparison W/ radiologist
- 2 stage ; ROI detection(Faster R-CNN W/ ResNet10 + CNN classificaiton W/ 6 layers) ; github 참고 + 해당 코드 부분
4. Results & discussion
ⅰ) statistic analyst
- ROI detection : IOU + Recall
- Classification : Kappa + Sensitivity + Specifities
Gwet K ; AC1 말하는건지 cohen's k 말하는 건지 모르겠음. AC1 같긴 함. reference label과의 합치도 구하는 것 같음
-> kappa ; radiologist와 비슷하거나 좀 더 낮음. dichotomous에서 간극 적음
sensitivity / specifiteis ; radiologist보다 dl이 더 높은 경우도 존재
external ; normal classi.. - substantial Kappa, dicho.. - almost perfect agree
ⅱ) visualization - attribution
- heatmap overlay (Cam?) ; attribution check + visualization
ⅲ) limitation
; single expertise reference standard set without consensus
1. select common LSS classificaiton reference standard. But there's still controversy -> Use controversal reference standard
2. test set reference labeled by single experts -> no consensus labeling make it biased towards expert
3. manually labeled by radiologist(labor-intense) ; highly supervised -> limited number of training data
5. Future Work(?) / Conclusion
- weighted categorial cross entropy loss
- Roi
- SpineAI 코드 참고
공부하다 정리한거
-
kappa(카파) 계수 - k values
: 관찰자 간 의견 일치 정도
Reference - https://velog.io/@yeonheedong/%ED%86%B5%EA%B3%84-Cohens-Kappa-%EA%B3%84%EC%88%98 -
Intra-reader variability: measurement differences from the same reader, which usually stem from the difficulty of the imaging case or reader interpretation drift (i.e., deviation from study-specific image interpretation criteria on which readers have been trained)
Inter-reader variability: measurement differences between two or more readers, usually due to readers’ skill.
?
- sagittal & axial 어떻게 같이 썼다는거?? feature map concat?
ResNeXt
https://velog.io/@krec7748/ResNeXt
https://velog.io/@pre_f_86/ResNeXt-%EB%85%BC%EB%AC%B8-%EB%A6%AC%EB%B7%B0#6-%EC%BD%94%EB%93%9C%EA%B5%AC%ED%98%84
Gwet's AC1 vs Cohen's kappa
heatmap overlay
https://www.tensorflow.org/tutorials/interpretability/integrated_gradients?hl=ko
