Temporal Super-Resolution of Medical Images with Implicit Neural Representations

Temporal image sequences are essential in medical imaging for analyzing motion and dynamic physiological processes. The temporal resolution (i.e., the number of frames in a sequence) plays a major role in the accuracy of downstream tasks. To address this, we propose a temporal super-resolution method based on implicit neural representations (INRs), that models smooth deformations continuously across time. Unlike conventional interpolation techniques that assume linear motion or require extensive training data, our approach optimizes an INR directly for each image sequence. We leverage a continuous time encoding mapped onto a unit circle, allowing for the generation of intermediate frames at any time point and thus creating image sequences with an arbitrary high temporal resolution. To validate our method we test on two 4D CT datasets, with CT scans over a respiratory cycle and a heart cycle. We evaluate the method by reconstructing frames excluded during training and demonstrate that our method outperforms other temporal interpolation methods in several reconstruction quality metrics. Our method provides a flexible, memory-efficient solution for enhancing temporal resolution in medical imaging while maintaining high spatial resolution.