Classification OCTA Images with Deep Learning

Abstract

Advancements in medical imaging, particularly in Optical Coherence Tomography Angiography (OCTA) images, have revolutionized the diagnosis and treatment of various eye diseases. However, the escalating utilization of OCTA scanning has placed significant demands on expert time, impeding the efficiency of the analysis process. This requires the application of machine learning technology.

This thesis endeavors to address the challenges faced in ophthalmology by integrating deep learning methodologies into the workflow of ophthalmologists. Our goal is to investigate and apply a specialized deep learning-based model for OCTA image classification and clustering. Through this effort, we aim to propose a robust approach for OCTA image analysis, providing valuable insights into various eye conditions. However, the scarcity of high-quality and accurately labeled datasets in the published literature remains a major obstacle for research purposes. Often, available datasets lack comprehensive labels, hindering model development and evaluation. Therefore, this research aims to contribute to model building in labeled and unlabeled dataset conditions.

The study utilizes a labeled dataset of 200 images and an unlabeled dataset of 600 images, investigating the performance of a CNN model and the VGG16 architecture for classification and feature extraction. Additionally, K-Means clustering is applied to group unlabeled data.

The performance evaluation analyzed the model's accuracy during training and showed progressive improvement from 60% to 82%. Additionally, it presents a silhouette score of 0.1139, indicating satisfactory clustering performance despite challenges such as data mismatch. This analysis includes insights gained from confusion matrices, ROC curves, and precision gain curves, offering a comprehensive assessment of model efficacy.