LITS_Hybrid_Comp_Net (Paper Link - https://arxiv.org/pdf/1909.04797.pdf)

Score of Test Set from LiTS website -

Built With/Things Needed to implement experiments

• Python - Python-2
• Keras - Deep Learning Framework used
• Numpy - Numpy
• Sklearn - Scipy/Sklearn/Scikit-learn
• CUDA - CUDA-8
• CUDNN - CUDNN-5 You have to register to get access to CUDNN
• LiTS - LiTS-dataset website
• [12 gb TitanX]- Used for all networks mentioned here
• [OpenCV] - To visualize 2D slices. Can be installed using Pip
• Requires X forwarding and MatplotLib to open the interactive Histogram Distribution clicker
• Nibabel - To open nii images. Can be installed using Pip

Preprocessing

Run Preprocess_create_numpy_array_from_data.py to obtain the 3d numpy arrays of the subjects Controls are mentioned in the file itself in the initial comments but to summarize please read the following

• Set filepath to load data
• Set filepath to save the 2D slices as .png to visualize them
• Set filepath to save the numpy arrays
• The for loop of line 58 can help you control the files you need changes on
• The values of the numpy array are normalized between 0-255 from initial values and not 0-1 and has to be made 0-1 before being passed to the network. The value is kept between 0-255 to easily visualize as png but for the network please use 0-1

When the code is run it will pop up an Intensity Distribution and the aim is to select the rightmost peak of the graph as shown below -

The following images show the preprocessed files resulting from this method -

You will notice that the rightmost peak intensity varies across the subject and this is expected.

If you have difficulty to obtain the right most peak, you may redo that specific file using the for loop control in line 58.

In the event this is too troublesome a simple fix is to normalize the subject by setting all pixels below -200 to 0 and all pixels over 250 to 0 and segmenting the liver region seperately first. Following this the Distribution graph returns only one peak and it is easy to select it.

The function def find_min_and_max(img) in line 25 is the fuction which actually does the left and right range selections and can be reused as desired.

2D Network to segment liver region and also to predict large tumors

This network is based on the idea of the Complementary Networks presented in the paper here - https://arxiv.org/abs/1804.00521. The exact architecture details are mentioned in the paper - https://arxiv.org/pdf/1909.04797.pdf in section 2.3. The Tumor detection Networks has to be compiled and trained twice.

• The 2D CompNet for the liver segmentation is trained for 40 epochs using the Adam optimizer with a learning rate of 5e-5
• Train the networks using the Adam optimizer with a learning rate of 5e-5, and having an early stopping scheme with the tolerance being set to 5; then we train the networks with a learning rate of 1e-6 using an early stopping with a tolerance of 10 trials. Both steps have 150 maximum number of epochs for training.

3D Network to segment small tumors

This network is a 3d version working on small 3D patches and exact details of the architecture is mentioned in section 2.3 in https://arxiv.org/pdf/1909.04797.pdf. The Tumor detection Networks has to be compiled and trained twice.

• Train the network using the Adam optimizer with a learning rate of 5e-5, and having an early stopping scheme with the tolerance being set to 5; then we train the networks with a learning rate of 1e-6 using an early stopping with a tolerance of 10 trials. Both steps have 150 maximum number of epochs for training.

email me - rd31879@uga.edu for any questions !! Am happy to discuss