A resolution enhancement plug-in for deformable registration of medical images

Image registration is a fundamental task for medical imaging. Resampling of the intensity values is required during registration and better spatial resolution with finer and sharper structures can improve the resampling performance and hence the registration accuracy. Super-resolution (SR) is an algorithmic technique targeting at spatial resolution enhancement to achieve an image resolution beyond the hardware limitation. In this work, we consider SR as a preprocessing technique and present a CNN-based lightweight resolution enhancement module (REM) which can be easily plugged into the registration networks in a cascaded manner. Different residual schemes and network configurations of REM are investigated to obtain an effective architecture design. Besides, an auxiliary loss is introduced into the cascaded network to empower multi-hierarchical supervision and strengthen the fidelity of the output. In fact, REM is not confined to image registration, it can also be straightforwardly integrated into other vision tasks for enhanced resolution. In the experiments, REM and the cascaded registration network are evaluated on Brain MR images quantitatively and qualitatively at different upscaling factors. It is shown that REM not only improves the registration accuracy, especially when the input images severely suffer from the degraded spatial resolution, but also reconstructs resolution enhanced images which can be exploited for successive diagnosis.