Supervisors: Dr. Christine Tanner, Neerav Karani, Prof. Ender Konukoglu
Organ motion prediction assists with image-guided radiation therapy by increasing targeting accuracy. Motion models can be created offline from 4D magnetic resonance images, which are reconstructed from 2D navigator slices and data slices. Acquiring less navigator slices by predicting some via interpolation can reduce acquisition time overall. An CNN-based interpolation method has been proposed for interpolating sequences of 2D navigator slices. Pursuing this work, we incorporate spatial transformation components into the neural network for explicitly learning suitable transformations. We start with an approach, which addresses frame interpolation as local convolution over input images with spatially-adaptive convolution kernels. Inspired by this idea, we propose a CNN to learn displacement fields between input images. The performance is evaluated on the dataset consisting of MR navigator slices. It enables efficient extraction of motion patterns and achieves better interpolation performance.