UNI-MB - logo
UMNIK - logo
 
ALL libraries (COBIB.SI union bibliographic/catalogue database)
  • Centerline dice metric implementation is not streamlined [Elektronski vir]
    Podobnik, Gašper, 1996- ; Vrtovec, Tomaž
    Structures prominently expressed in a single dimension are commonly referred to as network-like or tubular. One of the main characteristics of such structures is their topology, making the ... preservation of their topological connectedness crucial in image analysis tasks, such as automatic image segmentation. This is challenging because tubular structures, such as vessels, neurons and muscle fibers, are often very thin, and a segmentation error of just one pixel or voxel can considerably impact the topology of the resulting segmentation. These challenges have spurred research into the definition of adequate validation metrics and loss functions that measure or penalize topological deviations. One such recently proposed metric is the centerline Dice (clDice), which has been widely adopted by the medical image analysis community. Although clDice has a sound theoretical foundation, its practical implementation poses challenges. Our empirical analysis reveals that the skeletonization algorithm, a critical component of clDice computation, that was adopted by open-source implementations of clDice is sensitive to the order of image coordinate axes and does not account for arbitrary voxel size, subsequently impacting the metric value. Our experiments on the three-dimensional brainstem, carotid arteries and spinal cord segmentation masks revealed that these deviations in clDice can measure of up to 25.6%pt. We provide an intuitive explanation for these observations, propose guidelines to mitigate this issue, and outline potential future directions to enhance the robustness and precision of the clDice metric implementation.
    Source: Medical imaging 2025 [Elektronski vir] : Image Processing : 17-20 February 2025 San Diego, California, United States (7 str. = Progress in biomedical optics and imaging (Online). - ISSN 2410-9045. - 7 str.)
    Type of material - conference contribution ; adult, serious
    Publish date - 2025
    Language - english
    COBISS.SI-ID - 233985539