Representation Learning for imaging
- Leverage representation learning approaches to model a radiologists assessment of prostate MR images from data in which radiologists are confident in their assessment of risk for clinically significant prostate cancer, and use it to disambiguate the equivocal PI-RADS 3 assessments, and avoid unnecessary biopsies in general. Paper
- Leverage underlying tissue-specific information in multi-contrast MR images to learn representations of local regions in an image such that regions belonging to similar tissue types generate similar representations. This constained contrastive learning approach is used to pretrain deep learning models to reduce the amount of labeled data required for downstream segmentation tasks in MRI. Paper
- Leverage shared information in multi-contrast MR images to synthesize Gd-enhanced contrast MR images from non-contrast enhanced MR images that are routinely acquired in the clinic (T1-weighted, T2-weighted, FLAIR). ISMRM Proc
- Synthesize white matter nulled MPRAGE images from conventional T1-weighted MPRAGE images to improve thalamic nuclei segmentation. Paper