Ensemble Learning Approach to Retinal Thickness Assessment in Optical Coherence Tomography

Abstract

Manual assessment of the retinal thickness in optical coherence tomography images is a time-consuming task, prone to error and inter-observer variability. The wide variability of the retinal appearance makes the automation of retinal image processing a challenging problem to tackle. The difficulty is even more accentuated in practice when the retinal tissue exhibits large structural changes due to disruptive pathology. In this work, we propose an ensemble-learning-based method for the automated segmentation of retinal boundaries in optical coherence tomography images that is robust to retinal abnormalities. The segmentation accuracy of the proposed algorithm was evaluated on two publicly available datasets that included cases of severe retinal edema. Moreover, the performance of the proposed method was compared to two existing methods, widely referenced in the relevant literature. The proposed algorithm outperformed reference methods at segmenting the retinal boundaries in both normal and pathological images. Furthermore, a thorough reliability analysis showed a strong agreement between the retinal thickness measurements derived from the segmentation obtained with the proposed method and corresponding manual measurements computed with the manual annotations.

Keywords

• Deep learning
• ensemble learning
• semantic segmentation
• image processing
• retinal thickness
• optical coherence tomography