際際滷

際際滷Share a Scribd company logo
2018. 10. 17.
襭 覿 覦 豌襴
- Introduction to Medical Image
Analysis and Processing -
ModuLab DLC-Medical1
(1) Dataset
v 一危 (豢豌: 襭蠍郁)
則 IRB (Institutional Review Board, )
 豬渚 蟆曙  郁規 : 襭螻  , 襭蠍郁 讌, 渚 讌, 蟲一 
則 螳語覲 覲危碁 (碁 覦豢 X)
v Open Datasets (e.g. Grand Challenges in Biomedical Image Analysis)
襯碁襯危 螳 燕  覯覈 覲願
1947 1964 1979 1987 2004
渚 
蟯襴蠍一 
覈 るΜ 覦
 蟯 覯襯
(2) Research
則 レ(Prospective Study) vs レ 郁規(Retrospective Study)
則 螻牛 蠏 vs  蠏
 Impact Factor
Why Important?
Medical images account for
at least 90% of all medical data!
The largest data source in the health-care industry
Object
Detection
Semantic
Segmentation
Visual
Tracking
Action
Recognition
Captioning
Question
Answering
Image
Classification
(1) Dataset Cleaning (80%)
 Raw dataset  Trainable dataset
e.g. ROI (Segment), Outlier 蟇 
(2) Class Imbalance Check
 Statistical Data Visualization
(3-1) Dataset Splitting
則 Train : Validation : Test (; View of Machine Learning Researcher)
= Train : Test : Extra Validation (; View of Clinician)
則 Class Imbalance 螻!
譬1 (n=600) 譬2 (n=240)
ex) Train : Validation : Test = 4 : 1 : 1
譬3 (n=240)
400 160 160 100 40 40
Train Validation Test
100 40 40
(3-2) Dataset Splitting
則 レ(Prospective Study)
ex) (12~18) Train : Validation : (19) Test = 4 : 1 : 1
則 レ 郁規(Retrospective Study)
ex) (1) Train : Validation : (2) Test = 4 : 1 : 1
2-1) Test 一危一 襦 覿襴 覈語 . (1螳 set)
2-2) Test 一危一 覦蠑語願覃 覈語 . (6螳 set)
1-1) Train : Validation 一危一 襦 覿襴 . (1螳 set)
1-2) Train : Validation 一危一 覦蠑語願覃 . (5螳 set)
 1螳 豕 覈
 覲旧螳 覈碁れ 蠏
ModuLab DLC-Medical1

More Related Content

ModuLab DLC-Medical1

  • 1. 2018. 10. 17. 襭 覿 覦 豌襴 - Introduction to Medical Image Analysis and Processing -
  • 3. (1) Dataset v 一危 (豢豌: 襭蠍郁) 則 IRB (Institutional Review Board, ) 豬渚 蟆曙 郁規 : 襭螻 , 襭蠍郁 讌, 渚 讌, 蟲一 則 螳語覲 覲危碁 (碁 覦豢 X) v Open Datasets (e.g. Grand Challenges in Biomedical Image Analysis) 襯碁襯危 螳 燕 覯覈 覲願 1947 1964 1979 1987 2004 渚 蟯襴蠍一 覈 るΜ 覦 蟯 覯襯
  • 4. (2) Research 則 レ(Prospective Study) vs レ 郁規(Retrospective Study) 則 螻牛 蠏 vs 蠏 Impact Factor
  • 5. Why Important? Medical images account for at least 90% of all medical data! The largest data source in the health-care industry
  • 7. (1) Dataset Cleaning (80%) Raw dataset Trainable dataset e.g. ROI (Segment), Outlier 蟇 (2) Class Imbalance Check Statistical Data Visualization
  • 8. (3-1) Dataset Splitting 則 Train : Validation : Test (; View of Machine Learning Researcher) = Train : Test : Extra Validation (; View of Clinician) 則 Class Imbalance 螻! 譬1 (n=600) 譬2 (n=240) ex) Train : Validation : Test = 4 : 1 : 1 譬3 (n=240) 400 160 160 100 40 40 Train Validation Test 100 40 40
  • 9. (3-2) Dataset Splitting 則 レ(Prospective Study) ex) (12~18) Train : Validation : (19) Test = 4 : 1 : 1 則 レ 郁規(Retrospective Study) ex) (1) Train : Validation : (2) Test = 4 : 1 : 1 2-1) Test 一危一 襦 覿襴 覈語 . (1螳 set) 2-2) Test 一危一 覦蠑語願覃 覈語 . (6螳 set) 1-1) Train : Validation 一危一 襦 覿襴 . (1螳 set) 1-2) Train : Validation 一危一 覦蠑語願覃 . (5螳 set) 1螳 豕 覈 覲旧螳 覈碁れ 蠏