210 likes | 296 Views
Treatment Outcome Prediction Model of Visual Field Recovery Using SOM. JOJO 2011.12.22. Outline. Basic knowledge Treatment Outcome Prediction Model Feature selection Self-organizing-maps Conclusion. Outline. Basic knowledge Treatment Outcome Prediction Model Feature selection
E N D
Treatment Outcome Prediction Model of Visual Field Recovery Using SOM JOJO 2011.12.22
Outline • Basic knowledge • Treatment Outcome Prediction Model • Feature selection • Self-organizing-maps • Conclusion
Outline • Basic knowledge • Treatment Outcome Prediction Model • Feature selection • Self-organizing-maps • Conclusion
Basic knowledge 1 Diagnosis of damage to the visual system Reaction time High Resolution Perimetry (HRP) Detection
Basic knowledge 1 Diagnosis of damage to the visual system Diagnostic spots definition:
Basic knowledge 2 Vision Restoration Training(VRT) After damages to visual system, spontaneous recovery happens. When the recovery finished, VRT is used to treat patients. How can we know the results of VRT before it’s applied?
Basic knowledge 3 Treatment Outcome Prediction Step1: build a TOPM with patients’ baseline diagnosis and diagnostic charts Step2: extract features from a patient’s baseline diagnosis chart Step3: predict the treatment outcome with TOPM
Outline • Basic knowledge • Treatment Outcome Prediction Model • Feature selection • Self-organizing-maps • Conclusion
TOMP (FS) Global features
TOMP (FS) Conformitytohemianopiaandquadrantanopia
TOMP (FS) Local features
TOMP (SOM) 1 Theory: Winner takes all
TOMP (SOM) Local feature
TOMP (SOM) 2 Prediction: the winner takes all decided
TOMP (SOM) 3 Results:
TOMP (SOM) 3 Results: (Model evaluation: 10-fold cross validation) P: the number of hot spots N: the number of cold spots TP: correctly classified positive samples FP: incorrectly classified positive samples
TOMP (SOM) 3 Results: (Model evaluation: 10-fold cross validation) ROC:
TOMP (SOM) 3 Model evaluation: 10-fold cross validation 44%±4.7% 6%±1.9% 84.2%±1.4% 45.3%±4.5% 3.2%±0.8% 86.8%±1.1% 68.5%±4.0% 4.7%±1.0% 90.0%±0.8%
Outline • Basic knowledge • Treatment Outcome Prediction Model • Feature selection • Self-organizing-maps • Conclusion
Conclusion Why choose SOM? • Its non-linearity and self-organization methodology allows a comprehensible adaptation to the data distribution. • Simplify the process of data mining and the feature selection phase by conveniently combining both prediction and data exploration.