Probability-based lesion-aware FN/FP loss for wholebody PET/CT tumor segmentation

Accurate tumor segmentation on whole-body 3D PET/CT requires both high voxel-wise overlap and reliable lesion-wise detection. Standard CE+Dice nnU-Net training can achieve strong Dice while still missing individual lesion components. We propose a Probability-Weighted Error-Gated False-Negative/False-Positive (PWEG-FNFP) loss for binary whole-body tumor segmentation. The loss extends CE+Dice with error-gated connected-component terms that apply additional supervision only to current false-negative and false-positive lesion components. The false-negative term increases foreground support inside missed lesions, whereas the false-positive term suppresses spurious predicted components according to mean foreground probability. We further evaluate a center-regularized variant by adding the ℒ𝐶𝑒𝑛𝑡𝑒𝑟 term from Instance-wise and Center-of-Instance (ICI) to incorporate lesion-center localization. A 3D full-resolution nnU-Net was trained on 835 scans from Centre Antoine Lacassagne longitudinal FDG PET/CT lung cancer cohort and evaluated on 92 held-out internal scans and the external AutoPET dataset. On the internal test set, conventional non-lesion-aware losses achieved lesion-wise F1-scores of 0.536–0.599. The best performance was obtained when the PWEG-FNFP loss was combined with ℒ𝐶𝑒𝑛𝑡𝑒𝑟, suggesting that error-gated FN/FP supervision and center localization provide complementary optimization signals. Size-stratified analysis showed the strongest gains for lesions below 1 cm3. On AutoPET, the PWEG-FNFP loss achieved the highest Dice among lesion-aware variants, while the center-regularized version achieved the highest global lesion-wise F1 score. These results support probability-weighted connected-component optimization for improving small-lesion detection while highlighting the remaining challenge of false-positive control and domain generalization.