Addressing the Bias of the Dice Coefficient

While self-configuring U-Net architectures excel at a vast majority of supervised medical image segmentation tasks, they strongly rely on the chosen loss function. We demonstrate that a commonly employed Dice or cross entropy loss leads to a bias of the trained network, that is critical for the clinical application of airway segmentation from CT scans. The effort to produce the most accurate segmentations is skewed towards larger anatomical structures, leaving smaller peripheral airways with poorer quality. To address this bias, we explore several different choices of amending the label definition, including morphological dilation, and find that separating the binary airway segmentations into at least two distinct structures yields substantial improvements of approximately 4% in peripheral areas. This finding could directly benefit several clinically relevant tasks, among others virtual CT bronchoscopy.