|
Machine learning for retinal health classification of optical coherence tomography images |
|---|---|
| รหัสดีโอไอ | |
| Title | Machine learning for retinal health classification of optical coherence tomography images |
| Creator | Methawee Ratanapunperm |
| Contributor | Pakinee Aimmanee, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Unhealthy OCT image detection, Machine learning, Retinal layer geometry, Retinal layer relationships, Interpretability, Feature engineering |
| Abstract | Optical Coherence Tomography (OCT) is an essential imaging technique for diagnosing retinal diseases. While deep learning models offer high accuracy for automated OCT classification, their “black-box” nature limits clinical trust and adoption. Conversely, traditional machine learning methods are more interpretable but often depend on features that can be difficult to extract or are sensitive to image quality. This thesis presents two complementary studies to develop an accurate and interpretable machine learning framework for retinal health classification.The first study introduces a method based on extracting computationally simple and clinically intuitive geometric features—specifically foveal concavity, bilateral symmetry, and layer smoothness—from the Inner Limiting Membrane (ILM) and Retinal Pigment Epithelium (RPE) layers. Using a Light Gradient Boosting Machine (LGBM) classifier, this approach achieved 96% accuracy, 98% precision, and a low false negative rate of 6%, significantly out performing a baseline method using vertical line profile features.The second study expands this analysis by extracting a comprehensive set of 56 features derived from the thickness, area, and inter-layer relationships of seven segmented retinal layers. This layer-wise approach, while achieving a slightly lower binary classification accuracy of 93.5% on the test set, demonstrated robust performance and provided deeper clinical insights. Through interpretable models, this analysis revealed a hierarchical pattern of retinal deterioration. Key predictive features included the area ratio of specific layers to the total retina and thickness ratios between adjacent layers, offering a more granular understanding of disease mechanisms.Together, these studies demonstrate that analyzing retinal layer morphology, from simple geometries to complex inter-layer dynamics, provides a powerful and transparent framework for automated disease detection. The proposed methods support ophthalmologists by improving diagnostic accuracy and efficiency, contributing to the early detection and prevention of vision loss. The initial geometric approach is ideal for rapid pre-screening, while the layer-wise analysis enhances diagnostic insight by uncovering specific, clinically meaningful patterns of pathology. |
