About the Comparison and Evaluation of Real-Valued Multimodal Medical Point Clouds
Christian Janorschke, Jingyang Xie, Xinyu Lu, Floris Ernst · Current Directions in Biomedical Engineering · 2025
Abstract Two- or three- dimensional point clouds are common representations of data in the medical context, especially in medical imaging and segmentation. When assigning realvalued coordinates and transformations - e.g., for registration with an additional imaging modality - metrics like the Dice similarity coefficient, which rely on set operations on discrete point sets, have to be approximated. We applied different calculation methods for multiple metrics on a synthetic dataset of basic shapes, combined with modifications like transformations or occlusions. The evaluation resulted in a wide range of values for all metrics, depending on the modification, calculation method, and parametrization. This variability suggests both a potential clinical risk and the opportunity for selectively reporting more favourable results in publications. Therefore, and considering the influence of point spacing and distribution, we recommend reporting multiple metrics from diverse categories in the context of real-valued point clouds.