Deep Learning based Segmentation for Multi MR Imaging Protocols using Transfer Learning for PET Attenuation Correction

Magnetic resonance (MR) image segmentation is a robust technique used for PET attenuation correction. However, the segmentation of the brain into different tissue classes is a challenging task because of the similarity of pixel intensity values. The objective of this work is to propose a deep learning network to segment T1-weighted MR images of a dataset consists of 50 patients. Additionally, transfer learning is applied to segment another MR image protocol which is T2-weighted. The pretrained network with T1-weighted images is finetuned then tested with a dataset of 14 patients only. The Dice coefficients of air, soft tissue, and bone classes for T1-weighted MR images are 0.98, 0.92, and 0.79 respectively. The results of transfer learning show the feasibility of finetuning a deep network trained with T1-weighted images to segment T2-weighted images.