Speaker
Description
Optical Coherence Tomography (OCT) has become the gold standard in ophthalmology, offering non-invasive, high-resolution cross-sectional imaging of retinal microstructures. Despite its clinical significance, manual interpretation of large volumes of OCT data is time-consuming and highly dependent on medical expertise. The primary objective of this thesis is to develop and evaluate an automated diagnostic framework using artificial intelligence to classify retinal structures and pathologies from OCT images, thereby providing a reliable decision-support tool for clinical practice. Utilizing a comprehensive, peer-validated public dataset of 84,495 high-resolution Spectral-Domain OCT scans, this study implemented a deep learning approach based on the VGG16 Convolutional Neural Network architecture. The model leverages transfer learning with weights pretrained on ImageNet. To ensure mathematical stability and prevent overfitting, a conservative data augmentation strategy restricted exclusively to horizontal flipping was adopted, combined with adaptive class-weighting to address data imbalance. Optimization was achieved through a structured two-phase process: initial training of the custom classifier followed by strategic biphasic fine-tuning of the deep convolutional blocks. The optimized model demonstrated excellent performance on an independent, isolated test dataset, achieving an overall global accuracy of 94.45%, a precision of 93.35%, a sensitivity of 94.49%, and an F1-Score of 93.92%. Confusion matrix analysis highlighted exceptional clinical specificity, achieving a 100% correct classification rate for both normal control scans (NORMAL) and Central Serous Chorioretinopathy (CSR), alongside 94.1% accuracy for Diabetic Retinopathy (DR). Lower accuracies for Macular Hole (MH) and Age-Related Macular Degeneration (AMD) reflect inherent structural similarities in early-stage micro-alterations, which present ongoing boundary challenges for deep feature extractors. This research demonstrates that integrating deep learning workflows into medical physics and ophthalmology yields highly competitive automated diagnostic tools. The proposed pipeline offers a robust, reproducible, and standardized solution for rapid population screening, significantly reducing data processing time and mitigating subjective intra-observer variability in clinical environments. Keywords: Optical Coherence Tomography (OCT), Artificial Intelligence (AI), Deep Learning, VGG16, Retinal Pathologies.