3D Medical Segmentation GAN

Arda Mavi	Deniz Yuret
Ayranci Anadolu High School	Koç University

3D Liver Segmentation with GAN

This project created under Koç University Summer Research Program

Architecture:

Input Shape: n x 128 x 128 x 128
Output Shape: n x 128 x 128 x 128

In this project we purpose to segmentation medical scans without unsuccessful loss functions in segmentation area like Mean Squared Error (not useful for segmentation) or Dice Coefficient (using for area comparison but not useful for gradient descent optimization function) and for benefit the best use of GAN algorithmic logic.

! Project and Documentations are under construction!

New dataset expected...

For your dataset support: Arda Mavi e-Mail

Currently Used Dataset: DEU Liver Segmentation Dataset

Contents:

For Users

• Segmentation Scans

For Developers

• Processing Dataset
• Model Training

! Important Notes

To-Do List

For Users

Segmentation Scans:

python3 predict.py <Scan_files_path>

For Developers

Processing Dataset:

python3 get_dataset.py

Model Training:

python3 train.py

Important Notes

• Used Python 3.6.0 with Anaconda
• Install necessary modules with sudo pip3 install -r requirements.txt command.
• Load CUDNN(used cudnn/7.0.5/cuda-9.0 for this project) module before training.

To-Do List:

• Normalization of DICOM Liver datas.
• Optimize memory uses in test process.