Uncertainty-Aware Multiclass Retinal OCT Lesion Segmentation on AMD-SD via MC Dropout AMD-SD Üzerinde MC Dropout ile Belirsizlik Farkindalikli Çok Sinifli Retinal OCT Lezyon Segmentasyonu
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11637038
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: MC Dropout, multiclass segmentation, Retinal OCT, SwinUNet, U-Net family, uncertainty estimation, wAMD
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
This study addresses multiclass segmentation of retinal OCT lesions associated with neovascular (wet) age-related macular degeneration (wAMD) on the AMD-SD dataset (SRF, IRF, PED, SHRM, IS/OS). Five baselines from the U-Net family (U-Net, U-Net Plus Plus, ResUNet, Attention U-Net, U-Mamba) and SwinUNet were compared using a common training and evaluation protocol. To make reliability information explicit for clinical use, Monte Carlo (MC) Dropout is enabled during inference, and pixel-wise uncertainty maps are produced using predictive entropy and mutual information (MI). Quantitative results indicate that under class imbalance and boundary ambiguity, architectural differences become particularly apparent on small and fragmented lesions. Under this protocol, SwinUNet achieves the highest mean Dice score excluding background with a limited margin (0.6307); the uncertainty maps concentrate around lesion boundaries and low-contrast challenging regions, providing a practical review signal for potential error risk.