PatchBridgeNet: a patch-based deep feature extraction and classification model for automated retinal disease diagnosis from OCT images

Bayrak M., Dal D., Baygin M.

BIOMEDICAL SIGNAL PROCESSING AND CONTROL, cilt.113, 2026 (SCI-Expanded, Scopus) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 113
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1016/j.bspc.2025.109135
  • Dergi Adı: BIOMEDICAL SIGNAL PROCESSING AND CONTROL
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, EMBASE
  • Anahtar Kelimeler: Biomedical Image Analysis, Deep Feature Engineering, OCT, Patch-Based, PatchBridgeNet, Pre-Trained Networks, Retinal Disease Classification
  • Atatürk Üniversitesi Adresli: Evet

Özet

Artificial intelligence plays a significant role in transforming diagnostic and treatment processes in the healthcare sector, enabling early disease detection and personalized treatment approaches. Optical Coherence Tomography (OCT) is one of the most essential tools in this transformation, allowing for the detailed examination and classification of retinal diseases. In this study, we developed an innovative Deep Feature Engineering (DFE) model named PatchBridgeNet for automated classification of retinal diseases using OCT images. The proposed PatchBridgeNet integrates the lightweight efficiency of MobileNetV2, the hierarchical feature extraction capabilities of DarkNet53 and the dense connectivity patterns of DenseNet201 to comprehensively capture both global context and local patch-level features. Features extracted through the patch-based approach were optimized using the Iterative Neighborhood Component Analysis (INCA) algorithm and the Chi-Square (Chi2) statistical method. The optimized features were then classified using Support Vector Machines (SVM). The model achieved an accuracy of 92.3 % in multi-class classification and 97.4 % in binary classification tasks. PatchBridgeNet enabled the effective analysis of both global and regional details in the diagnosis of retinal diseases, providing a significant advantage over existing methods. Furthermore, the patch-based structure of the model facilitated the capture of small but critical information in OCT images. The results demonstrate that PatchBridgeNet holds significant potential for OCT image analysis and other medical imaging applications.