This repo maintains the project I did for MIT 6.869 Spring 2022, which is medical segmentation of liver and tumor.
- Network structure: U-Net with 16 features on the first layer. Optimizer was Adam. Loss function was crossEntropy
- Data set: liver data set from Medical Segmentation Decathlon. I sliced the 3-D CT images into 3-channel 2-D images for training and validation.
- An experimental idea of fusing the prediction labels along the three directions.
 |
 |
 |
 |
|---|---|---|---|
| CT image | True label | Prediction | Fused prediction |
- segmentation.ipynb contains the main code for training and visualization
- helper_function_seg.ipynb has some helper functions (for data preparation, generating images for the report, display 3D image as gif)
- models/ contains the trained model.
- videos/ contains the test results on a CT image (see details here).