FCN-CTSCAN

• Using fcn to segment CT images

Requirement:

• python 2.7
• tensorflow
• numpy
• matplotlib
• h5py (if you want to use my preprocessed data)

================================================

Abdominal CT image segmentation using fully convolutional networks

(Original method proposed in Jonathan Long, 2015)

• Blue: air
• Light blue: bone
• Orange: liver
• Teal/Green (depending on your screen): kidney
• Red: everything else

Dataset: http://www.ircad.fr/research/computer/

Input construction:

• Instead of using a single CT scan image, I combined 3 adjacent images into one stack. The input of the network has the shape of [batch_size, 480, 480, 3].
• Two zero images will be padded before and after the top/bottom CT scan image.

Preprocessing:

• Contrast limited histogram equilization was performed on each CT image (CLAHE)).
• Computing mean and std (shape is [1,1,1,3]) on train data. Subtract each stack of 3 iamge to the mean and divide them with the std.

Network architecture:

• The classification network is very similar to VGG. However, I reduced the number of params and layers since the number of train/test samples and classes are not as large as PASCAL VOC or imagenet.
• The rest of the fcn architecture is generally the same. Bilinear upsampling was used instead of tranposed convolution to avoid overfitting.

Training procedure:

• Before traing the actual fcnn, I had to pretrain the classification network first (as you can see in fcn_pretrain.py). Training samples for this part was generated randomly from the preprocessed data.
• The training method and hyper params can be seen in fcn_ctscan.py.

Note: my current PC is very limited in computing power, therefore for each patient I only sample half of the images. This resulted in a somewhat watered down precision for each class. I will redo this experiment later on a high-end computer, my expectation is that the overall accuracy will be higher.

Some preliminary results: Confusion matrix (except for precision and recall, the unit is pixel):

Other metric:

• Pixel accuracy: 0.976
• Mean accuracy: 0.979
• Mean IU (intersection over union): 0.777
• Frequency weighted IU: 0.957