On reconstruction, detection and segmentation with networks

I will talk about three big problems in medical image analysis: MRI
reconstruction, outlier detection and segmentation. The talk will be
divided in two parts. In the first part, I will describe probabilistic
models that use priors learned through neural networks and apply them
for MRI reconstruction and outlier detection. Results will show that
probabilistic models when used with appropriately strong priors can
lead to competitive results in reconstruction and outlier
detection. In the second part, I will describe our efforts towards
learning segmentation models with as little as one labeled volume
through optimized data augmentation. I will pose the training
procedure of a segmentation model as an optimization over segmentation
and augmentation procedures. This joint view will lead to higher
accuracies compared to other augmentation and semi-supervised learning
strategies.