Computer Vision News 30 Over time, more advanced deep-learning-based image registration methods have been developed. “LapIRN utilizes the advantages of a multi-resolution strategy while maintaining the non-linearity of the feature maps throughout the coarse-to-fine optimization scheme,” said Tony Mok about his winning LapIRN method. “The notorious hyperparameter tuning problem is addressed by introducing a conditional image registration learning paradigm [see Figure]. The framework of LapIRN is flexible and can be easily transferred to various medical image registration applications with minimal effort.” Yet, Learn2Reg still attracts submissions of conventional methods yielding impressive results. Interestingly, when executed on GPUs, these conventional methods match the speed of their deep-learning counterparts. The two deformation fields achieve similar accuracy metrics, yet Deformation 2 is smoother, making it more plausible. This shows the importance of evaluating the deformation field's plausibility as well. Image Registration Challenge
RkJQdWJsaXNoZXIy NTc3NzU=