Enhanced predictive models for macular hole surgery outcomes

Abstract

This thesis presents research work conducted in the field of retinal image analysis. More specifi cally, the work is directed at the employment of deep learning (DL) based image informatics for the analysis of diverse real world phenomena where features of interest are very difficult to distinguish. The evaluation of idiopathic full-thickness macular holes (MHs) holds critical clinical importance as MHs represent one of the strongest predictors of surgical success, influ encing both anatomical closure and functional visual recovery– a key motivation for developing robust deep learning frameworks to quantify their characteristics and predict postoperative out comes. In this context, three distinct parts to retinal image analysis are proposed. The first part addresses critical research questions on the quantitative assessment of MH, the role of DL in postoperative visual acuity (VA) prediction, the integration of automated optical coherence tomography (OCT) analysis for clinical decision-making, and the potential of DL models to improve diagnostic accuracy and support clinical practices. Hence, this part presents a compre hensive image informatics framework to create a high-quality spectral-domain OCT (SD-OCT) image dataset, providing a robust DL-based predictive model of VA in patients following surgery with MH and presenting an automated solution for non-standardised SD-OCT datasets. The imaging data undergoes preprocessing, quality assurance, and anomaly detection procedures. Seven state-of-the-art DL predictive models are then designed, implemented, trained, and tested with multiple two-dimensional (2D) input channels on the SD-OCT dataset. The models are quantitatively compared using four evaluation metrics. The method concludes the impact of the following surgery by predicting VA. Overall, the obtained results confirm that the fully automated approach with input from seven central SD-OCT images from each patient may robustly predict VA measurements using a high-quality SD-OCT image dataset. Following this, three-dimensional (3D) convolutional neural networks are integrated to train the model. 3D networks generally outperformed the 2D networks in some evaluation metrics; however, it came with the sacrifice of significantly more computational complexity. The second part identifies key research questions related to common sources of uncertainty in OCT images and proposes an effective method for representing and quantifying this uncertainty in DL-based predictive models. Furthermore, the study compares the proposed UQ method with existing approaches. In this context, the study highlights the significance of uncertainty, especially in dealing with the SD-OCT images. Predicting postoperative VA through DL models is crucial for decision-making and patient advisement, though their black-box behaviour is opaque to users and uncertainty associated with their predictions is not typically stated, leading to a lack of trust among clinicians and patients. To meet this need, an uncertainty-aware regression model is introduced for predicting postoperative VA using 3D SD-OCT images. The model not only x predicts VA post-surgery but also quantifies the associated uncertainty, enhancing reliability and trustworthiness. Qualitative evaluation shows that the proposed model outperforms commonly used methods in terms of prediction accuracy and reliability, demonstrating robust performance on out-of-sample data, including low-quality images and previously unseen instances. This makes the model a promising tool for clinical settings, improving the reliability of DL models in predicting VA. The third is the segmentation of the retinal external limiting membrane layer, where any disruptions in this layer are associated with worse visual outcomes in patients with idiopathic full-thickness MHs. Precise image-wise binary annotations are used to segment the retinal external limiting membrane (ELM) layer. Finally, qualitative and quantitative results are systematically compared with seven state-of-the-art DL-based segmentation methods to identify the ELM layer with an automated system. Additionally, it examines the feasibility of integrating automated ELM layer segmentation into clinical workflows while incorporating the latest advancements in DL-based ELM detection. The results confirm the efficacy of DL in retinal image analysis, providing a foundation for future enhancements in clinical applications. Future work will explore enhancing the models’ performance and efficiency, and extending the approach to other retinal conditions. Keywords: Image Analysis, Machine Learning, Deep Learning, Visual Acuity Measurement, Optical Coherence Tomography