Transformers beat CNNs in cross-hospital diabetic ulcer segmentation benchmark
SegFormer-B2 generalizes best on unseen data with 31.1% catastrophic failures
Researchers led by Abderrahmane Benfatah benchmarked three segmentation architectures—U-Net, DeepLabV3+ (both CNNs), and SegFormer-B2 (Transformer)—on diabetic foot ulcer (DFU) segmentation across multiple clinical sources. Trained on combined FUSeg/AZH data and evaluated on two independent external datasets (DFUC2022, Medetec) without fine-tuning, all models performed well in-domain (Dice 0.80–0.83) but degraded substantially cross-domain. Critically, the degradation was architecture-dependent: the Transformer-based SegFormer-B2 generalized best (DFUC2022 Dice 0.557, Medetec Dice 0.786), while the more complex DeepLabV3+ generalized worse than the simpler U-Net, indicating that model complexity is not the primary driver of generalization.
Per-image failure analysis across 2,160 external images confirmed SegFormer-B2 had the fewest catastrophic failures (31.1%), versus U-Net (38.5%) and DeepLabV3+ (43.0%). Wilcoxon signed-rank tests (p < 0.001 on both datasets) verified the consistent ranking. The study underscores that architecture family—Transformer versus CNN—strongly influences cross-hospital generalization, a critical insight for deploying medical AI in diverse clinical settings. The results suggest that practitioners should prioritize attention-based models when robustness to unseen data is essential.
- SegFormer-B2 (Transformer) achieved highest external Dice: 0.557 (DFUC2022) and 0.786 (Medetec), outscoring both CNNs.
- In-domain Dice scores across all models were high (0.80–0.83), but performance dropped 20–40% on external datasets.
- Simpler U-Net generalized better than more complex DeepLabV3+, proving architecture family trumps model complexity.
Why It Matters
Guide for medical AI teams: choose Transformer-based models for robust ulcer segmentation across hospitals and data sources.