Deep Learning-Based Temporal Assessment of Corneal Endothelial Morphology Following Descemet Membrane Endothelial Keratoplasty: A Comparative Analysis of Dual Architectural Approaches
2025 IEEE International Conference on Big Data, BigData 2025, Macau, Çin, 8 - 11 Aralık 2025, ss.2883-2890, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/bigdata66926.2025.11401015
- Basıldığı Şehir: Macau
- Basıldığı Ülke: Çin
- Sayfa Sayıları: ss.2883-2890
- Anahtar Kelimeler: corneal endothelial keratoplasty, deep learning, EfficientNet, image classification, clinical data integration, temporal specificity, machine learning, confocal microscopy
- Sağlık Bilimleri Üniversitesi Adresli: Evet
Özet
This study investigates a dual-architecture deep learning approach for quantitative assessment of corneal endothelial cell morphology using in vivo confocal microscopy (IVCM) following Descemet Membrane Endothelial Keratoplasty (DMEK). We propose an integrated framework combining deep learning-based image classification with classical machine learning models for temporal-specific analysis. Architecture I employs EfficientNetV2B3 to classify endothelial cell morphology into four clinically meaningful categories (Good, Bad, Small, Large) across 9,695 images from 76 patients, achieving 89.64% mean test accuracy with F1-scores ranging from 0.8333-0.9661. Architecture II integrates classification outputs with 67 clinical parameters and applies five classical machine learning algorithms (Decision Tree, Gradient Boosting, K-Nearest Neighbors, Random Forest, and Ensemble Meta-Learner) to predict clinical outcomes at three post-operative timepoints (3, 6, and 12 months). Results demonstrate superior performance of the integrated approach, with K-Nearest Neighbors achieving F1-Score of 0.9430 at the 6-month optimal diagnostic window. Synthetic data augmentation improved accuracy by 11.2-13.8%, while maintaining temporal consistency. The ensemble approach provided robust performance at 12 -month follow-up with only 5.77% decay from 6-month peak. This dual-architecture framework establishes clinically actionable temporal windows for AI-assisted endothelial assessment, advancing deep learning and machine learning applications in real-world corneal transplant surgery outcomes prediction.