Evaluation And Testing Components In Monai Implementation Of Deep Learning-Based Registration

Image registration is important in many clinical applications. The recent development in deep-learning based methods achieved competitive performance with greater efficiency during inference - crucial for many time-critical applications. Promising results have been reported from scientific evaluation, often with retrospective data in controlled experiments. The next step will inevitably be evaluating these in clinical studies, likely using sizeable external validation datasets in real-world settings. Evaluation is key to demonstrate registration values and patient benefits. In this abstract, we describe two examples, which adopt an efficient approach to evaluate and test many components in the learning-based algorithms by benchmark comparison with their counterparts in open source classical algorithms. We highlight that the adopted approach is simple yet methodical in ensuring the quality of these components in such a new class of algorithms.