Yapay Zeka · 4 dk okuma · 17 Nisan 2026
Hybrid segmentation model improves fundus lesion detection accuracy
Combining classical clustering methods with deep learning achieves 89.7% accuracy on high-resolution eye images, outperforming standard U-Net approaches.
Kaynak: arxiv/cs.AI · Mohammadmahdi Eshragh, Emad A. Mohammed, Behrouz Far, Ezekiel Weis, Carol L Shields, Sandor R Ferenczy, Trafford Crump · orijinali aç ↗ ↗
A hybrid model merging mathematical clustering with U-Net architecture detects choroidal lesions in fundus images with 89.7% accuracy.
- — Choroidal nevi require early detection to prevent melanoma transformation and improve patient outcomes.
- — Standard U-Net models struggle with limited training data and inconsistent image labeling in medical datasets.
- — Hybrid approach combines classical segmentation (accurate but labor-intensive) with deep learning (data-hungry but scalable).
- — Achieves Dice coefficient of 89.7% and IoU of 80.01% on 1024×1024 resolution images.
- — Outperforms Attention U-Net baseline (51.3% Dice, 34.2% IoU) with better generalization to external datasets.
- — Reduces annotation burden and training data requirements compared to pure deep learning methods.
- — Designed as foundation for automated decision support system in ophthalmology clinics.
Sık sorulanlar
- A choroidal nevus is a benign pigmented lesion in the eye that can, in rare cases, transform into melanoma. Early detection improves survival rates and prevents complications. Misdiagnosis or delayed diagnosis can lead to poor patient outcomes, making automated detection tools valuable for clinicians, especially those without specialized expertise in ophthalmology.