Non-invasive assessment of the terminal ileum’s mucosal healing plays key role in managing Crohn’s disease (CD) patients. We develop machine-learning models to predict terminal-ileum’s mucosal healing from big-data databases of:
1) semi-quantitative clinical interpretation of Magnetic Resonance Imaging (MRI) data of CD patients, and
2) MRI images of CD patients. Our approach provides more accurate assessment of the terminal ileum’s mucosal healing compared to classical linear methods.
Mechanisms to determine deep-neural-networks confidence in their prediction by estimating their predictions’ uncertainty play a critical role in adopting deep-learning techniques for safety-critical clinical applications. We introduce a principled way to non-parametrically characterize the true posterior distribution of the neural-network predictions through stochastic gradient Langevin dynamics (SGLD). We demonstrated very high correlation between our measures of uncertainty and out-of-distribution data in MRI registration. Further, our approach improved registration accuracy and robustness.
In-vivo quantification of tissue biophysical properties plays a key role in personalized medicine. Motivated by classical model-fitting approaches, we introduce a new class of deep-neural-network architectures and training processes, to enable accurate and reliable quantification of tissue biophysical properties from quantitative MRI data. We demonstrated the added-value of our approach for Intra-Voxel Incoherent motion analysis of Diffusion-Weighted MRI data with clinical applications in oncology and gastroenterology.
