Bone & Bone Lesion Segmentation in SPECT/CT Using U-Net

This is an implementation of the bone and bone lesion segmentation, multi-modal U-Net in Nuclear Medicine Imaging, the network was trained on the simulated SPECT/Attenuation-Map images and tested on the patient PET/CT images.

Main script: boneSegUnet/run/train_UNet.py

U-Net:

Network Architecture

Trainable Parameters: 15,317,115.

Example Segmentation Results

Validation Set (SPECT/CT simulations)

Dice similarity coefficient: 0.958 for lesion, 0.963 for bone

Testing Set (PET/CT)

Resulting an undesired result with the unseen patient PET/CT images.

Example Feature Maps of the Testing Set (PET/CT)

Redundant feature maps were observed, meaning that we could reduce the number of filters/parameters.

CT encoding stage, 4th layer:

SPECT encoding stage, 4th layer:

About Myself