The Mathematics of Radiation Oncology Contour Datasets and the Implications for Automated Segmentation Algorithms

In radiation oncology, the accuracy of tumor contouring plays a crucial role in treatment planning, directly influencing patient outcomes. Metrics like Dice similarity coefficients and Hausdorff distances are often employed to assess the similarity between different annotations for a contour. Interobserver variability significantly impacts these metrics and, in turn, influences clinical decisions such as the determination of prescription target volume margin size. This talk will explore the mathematical foundations of these metrics and their role in quantifying contour discrepancies, delving into how variability among annotators affects the theoretical upper limits of Dice similarity during inference of automated segmentation models. We will introduce the basics of 3D automated segmentation models and the underlying mathematics of convolutional neural networks.