Automated Detection and Classification of Retinal Vein Occlusion Using Ultra-widefield Retinal Fundus Images and Transfer Learning

• Main Author: Ivy Ong Siaw Yin, Ong
• Format: Thesis
• Language: English
• Published: - 2024
• Subjects: Q Science (General); T Technology (General)
• Online Access: http://ir.unimas.my/id/eprint/46389/1/Revised_Thesis_21020483.pdf; http://ir.unimas.my/id/eprint/46389/

Retinal vein occlusion (RVO) is a retinal disease resulting from blockage of the retinal veins, leading to the development of haemorrhages and cotton wool spots within the retina. Retinal fundus imaging is essential for diagnosing these abnormalities. The study explores the potential of artificial intelligence (AI) and machine learning (ML) to automate the analysis of these images for detecting RVO signs. Detecting RVO signs in retinal images is challenging due to the irregular shapes, sizes, and locations of cotton wool spots and haemorrhages. Retinal images also include various anatomical structures and may suffer from poor quality, introducing artefacts. Manually segmentation is labour-intensive, requiring significant expertise and time. Moreover, existing models for retinal disease classification lack transparency, providing only classification labels without explanations or visualisations, raising concerns about their reliability. Additionally, the effectiveness of segmenting RVO signs in ultra-widefield (UWF) fundus images is uncertain due to a lack of studies and benchmarking. The study aims to develop automated methods for the segmentation and classification of RVO in UWF fundus images. It seeks to improve the interpretability and reliability of automated retinal disease screening by visualising pertinent disease features and examining the effectiveness of segmentation models trained on regular fundus datasets for UWF images. The methodology involves training segmentation models using both regular and UWF fundus images. The study also proposes using transfer learning, where a model pre-trained on a regular fundus dataset is fine-tuned with a UWF dataset to improve segmentation performance. The approach seeks to utilise knowledge from pretraining to enhance the performance of the segmentation model. The study also evaluates the classification model trained with lesion masks to classify images accurately into the respective categories. The automated algorithm can detect the presence of RVO in retinal fundus images and accurately pinpoint the locations of associated lesions. This improves the reliability and interpretability of automated retinal disease screening, aiding clinicians in making informed decisions. The study showcases the effectiveness of deep neural networks for feature extraction and classification, underscoring the potential of AI in medical imaging.