Modern machine learning tools have demonstrated that learning from data can be greatly beneficial for unsupervised problems. We are developing algorithms that leverage machine learning as well as traditional physics-based models to provide better estimation of underlying structures in MRI. These methods will not only improve our abilities to estimate such latent variables more accurately and with fewer measurements but will provide a robust framework enabling scientists and clinicians to extract information and biomarkers from these latent variables.
A common issue of quantitative MR methods is the poor efficiency in terms of the relation between resolution, signal to noise ratio (SNR), and acquisition time. In research settings, we can often afford to spend longer scan times to enhance SNR or spatial resolution. However, high SNR and long acquisition times are impractical in the clinic and uncomfortable for the participant, further increasing the occurrence of nuisance factors such as motion artifacts. As our research is moving towards multi-contrast integration, reduction of scan time for each contrast modality is becoming more important. It is in this setting that developing methods that require minimal number of measurements can be greatly beneficial. These inverse problems (I.e. reconstruction) are traditionally solved by formulating an optimization problem that can be expensive to compute in practice. Instead, we deploy data-enabled approaches to alleviate this huge computational burden by searching for parametric functions that approximate these solutions more efficiently. These alternatives are given by particular instantiations of deep convolutional networks. While this learning or fitting process can itself be computationally demanding, the load is shifted from deployment (where efficiency is crucial) to training (which can be done off-line and in advance).
Our goal therefore is to design, implement and validate deep learning based approaches to approximate the solution to inverse problems in MRI while reducing the number of necessary measurements. We will leverage implicit data-adaptive prior knowledge that will provide significantly faster reconstructions and of high SNR and reconstruction quality, even while acquiring a reduced number of measurements.
(1) Lai KW., Aggarwal M., van Zijl P., Li X., Sulam J. (2020) Learned Proximal Networks for Quantitative Susceptibility Mapping. In: Martel A.L. et al. (eds) Medical Image Computing and Computer Assisted Intervention – MICCAI 2020. MICCAI 2020. Lecture Notes in Computer Science, vol 12262. Springer, Cham.