• A Systematic Review of Deep Transfer Learning for Detecting Diabetic Macular Edema from OCT Images
  • Kimia Heydariyar,1,* Abdollah Karimi,2 Alireza Pourrahim,3 Alireza Vasiee,4 Omid Raiesi,5 Mohammad Mahdi Pourrahim,6
    1. Student Research Committee, Faculty of Medicine, Ilam University of Medical Sciences, Ilam, Iran
    2. Department of Computer Engineering and Information Technology, Amirkabir University of Technology, Tehran, Iran
    3. Student Research Committee, Faculty of Medicine, Ilam University of Medical Sciences, Ilam, Iran
    4. Department of Nursing, Faculty of Nursing and Midwifery, Ilam University of Medical Sciences, Ilam, Iran
    5. Department of Parasitology, School of Allied Medical Sciences, Ilam University of Medical Sciences, Ilam, Iran
    6. Student Research Committee, Faculty of Medicine, Ilam University of Medical Sciences, Ilam, Iran


  • Introduction: Diabetic macular edema (DME) is a leading cause of vision loss in diabetic patients and is routinely diagnosed via spectral-domain optical coherence tomography (SD-OCT). Deep learning models leveraging transfer learning have shown promise in accurately detecting DME with limited labeled data. Nonetheless, variability in model architectures, preprocessing methods, and validation strategies complicates assessment of their true diagnostic performance.
  • Methods: We conducted a systematic search of PubMed, Embase, IEEE Xplore, Scopus, and Web of Science up to May 2025. Search terms combined “diabetic macular edema” OR “DME” with “transfer learning”, “deep learning”, and “OCT”. Two reviewers independently screened titles, abstracts, and full texts for studies that: (1) applied deep transfer learning on OCT images for DME detection, and (2) reported sensitivity, specificity, accuracy, or AUROC. Data extracted included CNN architecture, pre-training domain, preprocessing steps, dataset size, validation approach, and performance metrics. Study quality was appraised using QUADAS-2.
  • Results: Four studies met inclusion criteria, comprising a total of 9,700 OCT B-scans from 590 patients (300 DME, 290 controls). Pre-trained networks included ResNet50 (n=2), GoogLeNet (n=1), and AlexNet (n=1). Common preprocessing steps were BM3D denoising (3 studies) and retinal layer cropping (4 studies). Three studies used k-fold cross-validation; only one performed external validation. Pooled estimates from a random-effects model were: sensitivity 0.94 (95% CI 0.91–0.96), specificity 0.95 (95% CI 0.93–0.96), and AUROC 0.98 (95% CI 0.97–0.99). ResNet50-based approaches achieved the highest individual study accuracy (mean 97.2%).
  • Conclusion: Deep transfer learning approaches demonstrate excellent accuracy for DME detection on OCT, highlighting their potential to augment clinical screening. However, high heterogeneity, limited external validation, and inconsistent reporting hinder generalizability. Future research should employ standardized preprocessing pipelines, multi-center cohorts, and transparent reporting (e.g., PRISMA-DTA) to facilitate clinical translation.
  • Keywords: Diabetic Macular Edema; Deep Transfer Learning; Optical Coherence Tomography