A Generalized Surface Loss for Reducing the Hausdorff Distance in Medical Imaging Segmentation

• DOI: 10.48550/arxiv.2302.03868
• Publication Date: 2023-01-01
• Authors: Adrian Celaya; Alejandro Díaz; Alex Balsells; Béatrice Rivière; David Fuentes
• Keywords: Hausdorff distance; Dice; Sørensen–Dice coefficient

Within medical imaging segmentation, the Dice coefficient and Hausdorff-based metrics are standard measures of success for deep learning models. However, modern loss functions image segmentation often only consider or similar region-based during training. As a result, architectures trained over such run risk achieving high accuracy but low metrics. Low on can be problematic applications as tumor where benchmarks crucial. For example, scores accompanied by significant Hausdorff errors could indicate that predictions fail to detect small tumors. We propose Generalized Surface Loss function, novel function minimize with more desirable numerical properties than current methods weighting terms class imbalance. Our outperforms other losses when tested LiTS BraTS datasets using state-of-the-art nnUNet architecture. These results suggest we improve our function.