On the Optimal Combination of Cross-Entropy and Soft Dice Losses for Lesion Segmentation with Out-of-Distribution Robustness

We study the impact of different loss functions on lesion segmentation from medical images. Although Cross-Entropy (CE) is most popular option when dealing with natural images, for biomedical image soft Dice often preferred due to its ability handle imbalanced scenarios. On other hand, combination both has also been successfully applied in these types tasks. A much less studied problem generalization all losses presence Out-of-Distribution (OoD) data. This refers samples appearing test time that are drawn a distribution than training In our case, we train models images always contain lesions, but have lesion-free samples. analyze minimization in-distribution performance, generalize OoD data, via comprehensive experiments polyp endoscopic and ulcer diabetic feet Our findings surprising: CE-Dice combinations excel segmenting poor performance which leads us recommend adoption CE problems, robustness Code associated can be found at https://github.com/agaldran/lesion_losses_ood .