Week 2

 Since Dr. Gauthier was in Paris for ISMRM conference this week, mostly of the time I read papers she sent to me or I found interesting and related to my immersion project topic. One of the most interesting paper I read from these paper stack was called "QSMnet". Our lab has pioneered the development and application of Quantitative Susceptibility Mapping (QSM) which solves the dipole inversion problem to get the magnetic susceptibility distribution around the tissue. In our method, we incorporated regularization term to penalize the striking artifacts caused by the ill-posedness of the dipole kernel, and used iterative optimization algorithm like Conjugate Gradient (CG) descent to minimize the objective function. QSM was firstly used to detect the susceptibility distribution around the brain, and has been extended to other applications like cardiac imaging, combination with PET imaging and so on. Recently, deep learning, or Convolutional Neural Network (CNN) specifically speaking, has become extremely popular when dealing with image related computational problems, and image reconstruction problem like QSM is one important case of them. In this QSMnet paper, the authors delicately designed a network structure to mimic the dipole inversion process appeared in QSM problem. They trained the network based on thousands of high quality subjects and applied the state of art deep learning techniques to get optimal results. Surprisingly, they obtained the results which were obviously better than ours. I think I should discuss this paper with the guys in the lab when they are back and possibly come up with other deep learning based method to improve our existing method.

During the most of the time this week, I stayed in the lab working on my thesis project, which focuses on medical image reconstruction problem. I also tried to figure out how to combine the image reconstruction techniques I've utilized to Dr. Gauthier's project, MS lesion related work. My first thought was to use the same network architecture as I'm using for reconstruction now to handle with MS lesion segmentation problem. I believe this should work as long as we have enough data to fit the model. I should talk about this direction with Dr. Gauthier when she is back too.

I really enjoy my PhD thesis topic and also the immersion project, hopefully some useful outcomes could be come out at the end of the immersion.