Liver Tumors Segmentation-CNNs

This project segments tumors in the Liver using 2 cascaded CNNs. We use 3D-IRCADb 01 as our dataset. View the Thesis for more details about the model, dataset, methodology and results.

There are 3 notebooks in this project

• 3d-ircadb-01_util.ipynb Responsible for renaming files, merging masks and augmenting the dataset

• liver_CNN.ipynb The first CNN, it uses a CNN to segment the Liver and extract the ROI

• tumor_CNN_final.ipynb The second CNN, it segments the tumors using a CNN after masking other organs using the extracted ROI from the first CNN

There is the models directories which contains the trained models

• liver_model_final_resunet.h5 which is the first CNN trained for 20 epochs
• tumor_weights_final_50epochs.h5 which is the second CNN trained for 50 epochs
• tumor_weights_final_100epochs.h5 which is the second CNN trained for 100 epochs

Dependencies

I used Anaconda for package management but pip will work all the same

• Tensorflow-GPU
• Keras (Tensorflow) already comes with Tensorflow
• Numpy
• OpenCV
• Matplotlib
• Pydicom
• Jupyter Notebook
• Scipy
• Scikit-learn
• PIL
• Seaborn
• Imageio

Dataset Structure

The structure of the dataset should be as follows ( if you don't want to change the code :grinning: )

train
|_ patients               # Contains all the CT-slices from all patients together with no directories inside
|_ masks                  # Contains the masks for the Liver and Tumors for the patients
   |_ merged_livertumors  # Contains the masks of tumors after merging each slice's tumor masks together
   |_ 1.1_liver           # Contains the mask for the liver of each slice for patient 1
   |_ 1.2_liver           # Contains the mask for the liver of each slice for patient 2
   |_ 1.3_liver           # Contains the mask for the liver of each slice for patient 3
      .
      .
      .
   |_ 1.20_liver          # Contains the mask for the liver of each slice for patient 20