nnU-Net achieves 0.80 Dice score, ranks 3rd in AutoPET III tumor segmentation
Whole-body PET/CT tumor segmentation gets a boost with training strategies and CraveMix augmentation.
Hussain Alasmawi presents a whole-body tumor segmentation method for PET/CT imaging, developed for the AutoPET III challenge. The approach leverages the nnU-Net framework with a ResNet-based encoder and systematically investigates training strategies including intensity normalization, batch dice optimization, and CraveMix data augmentation. These techniques significantly reduce false positives and improve robustness to lesion variability. The best configuration achieves a Dice score of up to 0.80 on the preliminary test phase, securing third place in the challenge. Code is open-source.
Manual segmentation in PET/CT is time-consuming and prone to variability. This work shows that carefully tuned training strategies can generalize across tracers and multi-center data. The combination of batch dice optimization (which focuses on overlap) and CraveMix (a mixup-style augmentation) is key. The method addresses challenges like varying lesion size, contrast, and anatomical distribution, bringing automated segmentation closer to clinical utility. The public code allows reproduction and further research.
- Achieved Dice score of 0.80 on AutoPET III preliminary test phase
- Uses nnU-Net with ResNet encoder, batch dice optimization, and CraveMix augmentation
- Ranked third in AutoPET III challenge; code publicly available
Why It Matters
Automated tumor segmentation could reduce manual effort and variability in PET/CT interpretation.