Abstract
Within medical imaging segmentation, the Dice coefficient and Hausdorff-basedmetrics are standard measures of success for deep learning models. However,modern loss functions for medical image segmentation often only consider theDice coefficient or similar region-based metrics during training. As a result,segmentation architectures trained over such loss functions run the risk ofachieving high accuracy for the Dice coefficient but low accuracy forHausdorff-based metrics. Low accuracy on Hausdorff-based metrics can beproblematic for applications such as tumor segmentation, where such benchmarksare crucial. For example, high Dice scores accompanied by significant Hausdorfferrors could indicate that the predictions fail to detect small tumors. Wepropose the Generalized Surface Loss function, a novel loss function tominimize Hausdorff-based metrics with more desirable numerical properties thancurrent methods and with weighting terms for class imbalance. Our loss functionoutperforms other losses when tested on the LiTS and BraTS datasets using thestate-of-the-art nnUNet architecture. These results suggest we can improvemedical imaging segmentation accuracy with our novel loss function.