Utilization of deep learning-based registration method in Parkinson’s disease diagnostic tool

Motivation: Image registration is a fundamental technique widely used in medical image analysis. However, the ANTs-based registration method has a limitation in terms of long processing time, which may make it difficult to apply in practice. Goal(s): To address this limitation, we changed the registration method used in Heuron IPD, a commercialized Parkinson's disease diagnostic tool, from ANTs-based method to a deep learning-based method. Approach: Specifically, we used TransMorph, an unsupervised medical image registration model and evaluated its performance on prospectively collected data. Results: The execution time of registration process was reduced while maintaining diagnostic performance of Heuron IPD. Impact: By changing the ANTs-based SyN registration method in Heuron IPD with TransMorph, a deep learning-based model, we reduced the registration processing time by approximately 32 times while maintaining the diagnostic performance of Heuron IPD.