Slide QR code

Members

1. Dr. Cheyu Hsu
   National Taiwan University Hospital
   
   
2. Chengen Lee
   National Taiwan University, Institute of Applied Mathematical Sciences
   
   
3. Yuehchou Lee
   National Taiwan University, Mathematics

Introduction

1. Challenge is to predict the one-year prognosis of Brain Metastasis after radiotherapy


2. Brain Metastasis is the kind of brain tumor


3. Brain tumor has the tumor pseudoprogression after radiotherapy

Goal

1. Contruct the workflow to support the radiologist to classify the one-year prognosis after radiotherapy


2. Simulate the multiple radiologists to label the tumor, that is to generate the different label


3. Select the robust Radiomics features

Workflow

Step 1: Image Dataset Preprocessing

1. MRI bias-correction, resampling and normalization


2. Simulation multiple tumor label by using single label (gorund truth)

Step 2: Radiomics Features Extraction

• Radiomics can exract lots of features, such as shape, volume, first-order and high-order features
  
  

  

  
  

  (ref: https://www.radiomics.io/pyradiomics.html)

Step 3: Classification Models Establishment

• eXtreme Gradient Boosting
  

  

  (ref: https://www.educba.com/xgboost-algorithm/)

Step 4: Classification Models Validation

• Use thousands of cases to validate the above steps

Preliminary Result

• All features:

     

• Selected features:

     

Conclusion

1. Reduce the cost of labeling tumors by multiple radiologists



2. Improve AUC and sensitivity of one-year prognosis classification by using selected features

Future Works

1. Validate above all steps by using NHI dataset (over 3000 cases)


2. Construct the segmentation model to support the radiologist to label tumor


3. Combine VAE to improve the result of classification

Thank You

MeDA Lab