Amplitude-scan classification using artificial neural networks

Laplacian feature detection and feature alignment for multimodal ophthalmic image registration using phase correlation and Hessian affine feature space

Advances in multimodal imaging have revolutionized diagnostic and treatment monitoring in ophthalmic practice. In multimodal ophthalmic imaging, geometric deformations are inevitable and they contain inherent deformations arising from heterogeneity in the optical characteristics of imaging devices and patient related factors. The registration of ophthalmic images under such conditions is challenging. We propose a novel technique that overcomes these challenges, using Laplacian feature, Hessian affine feature space and phase correlation, to register blue autofluorescence, near-infrared reflectance and color fundus photographs of the ocular posterior pole with high accuracy. Our validation analysis - that used current feature detection and extraction techniques (speed-up robust features (SURF), a concept of wind approach (KAZE), and fast retina keypoint (FREAK)), and quantitative measures (Sørensen-Dice coefficient, Jaccard index, and Kullback-Leibler divergence scores) - showed that our approach has significant merit in registering multimodal images when compared with a mix-and-match SURF-KAZE-FREAK benchmark approach. Similarly, our evaluation analysis that used a state-of-the-art qualitative measure - the mean registration error (MRE) - showed that the proposed approach is significantly better than the mix-and-match SURF-KAZE-FREAK benchmark approach, as well as a cutting edge image registration technique - Linear Stack Alignment with SIFT (scale-invariant feature transform) - in registering multimodal ophthalmic images.

Classification and Quantification of Retinal Cysts in OCT B-Scans: Efficacy of Machine Learning Methods

The automatic segmentation of fluid spaces in optical coherence tomography (OCT) imaging facilitates clinically relevant quantification and monitoring of eye disorders over time. Eyes with florid disease are particularly challenging to segment, as the anatomy is often highly distorted from normal. In this context, we propose an end-to-end machine learning method consisting of near perfect detection of retinal fluid using random forest classifier and an efficient DeepLab algorithm for quantification and labeling of the target fluid compartments. In particular, we achieve an average Dice score of 86.23% with reference to manual delineations made by a trained expert.