Novel Alignment and Segmentation Methods for Optical Retinal Images with Application to Non- Invasive Estimation of Intracranial Pressure

A Z M Ehtesham  Chowdhury

Automated analysis of image/video recordings of retinal anatomical structures has increasingly been used for quantitatively assessing eye health and sight-threatening conditions, e.g., glaucoma, diabetic retinopathy, papilledema, and retinal vessel occlusion. The utility of such methods is highly dependent on the accuracy of the alignment of sequences of optic nerve images of the same eye under several conditions, e.g., internal eye pressure variations due to postural changes and imaging nuisances due to operational variations, as well as the semantic segmentation of retinal structures in optic disc region and at its boundary. This thesis presents several novel and fully automatic methods, e.g., for video alignment, artery-vein segmentation and optic disc segmentation. It then demonstrates the utility of these methods in an intracranial pressure estimation pipeline to make it fully automatic. First, we propose an automatic alignment method to align multipostural video sequences captured at multiple ophthalmodynamometry forces (ODF). We present a novel iterative image transformation-based method for aligning the blood vessels in the optic disc of all ODF videos of a patient. Second, we propose a novel deep-learning method for segmenting the blood vessels and classifying these into arteries and veins. The method includes a multiscale encoder that learns feature maps at an increasing range of receptive fields at each network layer and a guided attention module that emphasises the vessel-like structures over other background contents. Third, we propose another deep-learning method to segment the optic disc region in an image. A dual-branch multiscale encoder and modified guided attention blocks are used to enhance the optic disc and its boundary, followed by a global context learner layer that leads to further improved optic disc boundary features. Finally, these aligned and segmented data are used in an existing photoplethysmography process at Lions Eye Institute to estimate patients' intracranial pressure (ICP). The photoplethysmography-derived retinal pulse amplitudes from the automatically classified optic disc vein pixels across the corresponding intraocular pressures (proportional to ODF) are utilised to estimate ICP indirectly. Overall, the proposed alignment and deep-learning segmentation methods achieved superior alignment and segmentation performances compared to existing art. The absolute differences between estimated ICP and True ICP for semi-automatic and automatic optical image analysis processes were close to the standard and highly sensitive in detecting elevated ICP (>= 14.7 mmHg). Proposed methods contribute to a fully automatic and non-invasive ICP estimation framework, which can be useful in clinical scenarios where invasive ICP measurement is not practicable or poses severe health risks to the patient. Proposed optical image alignment and segmentation methods can also be used to assist ophthalmologists in assessing sight-threatening conditions and therapeutic applications.

Novel Alignment and Segmentation Methods for Optical Retinal Images with Application to Non- Invasive Estimation of Intracranial Pressure

Files and links (1)

Abstract

Details

Metrics