Date of Degree
PhD (Doctor of Philosophy)
Electrical and Computer Engineering
Mona K. Garvin
Papilledema is a particular type of optic disc swelling caused by elevated intracranial pressure. By observing the visible features from fundus images or direct funduscopic examination, a typical method of assessing papilledema (i.e., the six-stage Fris\'en grading system) is qualitative and frequently suffers from low reproducibility.
Compared to fundus images, spectral-domain optical coherence tomography (SD-OCT) is a relatively new imaging technique and enables the cross-sectional information of the retina to be acquired. Using SD-OCT images, quantitative measurements like evaluating the retinal volume or depth are intuitively more robust than the traditional qualitative approach to evaluate papilledema. Also, multiple studies suggest that the deformation of the peripapillary retinal pigment epithelium and/or Bruch's membrane (pRPE/BM) may reflect the intracranial pressure change. In other words, modeling/quantifying the pRPE/BM shape can potentially be another indicator of papilledema. However, when the optic disc is severely swollen, the retinal structure is dramatically deformed and often causes the commercial SD-OCT devices to fail to segment the retinal layers. Without appropriate layer segmentation, all the retinal measurements are not reliable.
To solve the current issue of inconsistently assessing papilledema severity, a comprehensive machine-learning framework is proposed in this doctoral work to achieve the goal by accomplishing following four aims. First, robust approaches are developed to automatically segment the retinal layers in 2D and 3D SD-OCT images, even though the optic discs can be severely swollen. Second, the semi- and fully automated methodologies are designed to segment the pRPE/BM opening under the swollen inner retina in these SD-OCT images. Third, the pRPE/BM shape models are constructed using both 2D and 3D SD-OCT images, and then the 2D/3D pRPE/BM shape measures are computed. Finally, based on the previously segmented retinal layers, eight OCT 2D/3D global/local measurements of retinal structure are reliably computed. Considering both the 2D/3D pRPE/BM shape measures and these eight OCT features as an input set, a machine-learning framework using the random forest technique is proposed to compute a papilledema severity score (PSS) on a continuous scale. The newly proposed PSS is expected to be an alternative to the traditional qualitative method to provide a more objective measurement of assessing papilledema severity.
Papilledema is a specific type of optic-nerve-head (ONH) swelling due to elevated intracranial pressure, and it can indicate serious underlying conditions. Head injury, brain tumor, brain inflammation, subarachnoid hemorrhage, blockage of cerebrospinal fluid (CSF) flow, reduction in CSF reabsorption, idiopathic intracranial hypertension (IIH) are the possible reasons that may cause intracranial hypertension. To diagnose the severity of papilledema, clinicians commonly examine the visible features using either 2D retinal photos or direct observation of the retina. Then, a severity scale from 0 (normal) to 5 (severe) is decided based the clinician's judgment. However, this type of method is not stable, so different clinicians may frequently disagree with each other. Moreover, sometimes even the same clinician may have different decisions after multiple observations of the same retinal image.
Optical coherence tomography (OCT) is a new imaging technique and enables the cross-sectional information of the retina to be acquired. Therefore, the volume of the retina at the ONH region is able to be computed as an indicator to reflect the degree of the optic disc swelling. Similarly, the thickness of certain retinal regions as well as the shape of certain retinal layers are also potentially useful for the same purpose.
In this thesis, a system was proposed to use computers to mimic how clinicians make decisions when measuring the degree of the optic disc swelling. A technique called random forest was used to achieve this goal by automatically deciding the best feature combination from the input feature set, including the previously discussed retinal volume, thickness, and shape measures. It is expected that the output of this machine-learning system is very close to the decisions that clinicians would make; more importantly, this proposed system would have the same results every time if the input features are the same, and the processing takes much less time than clinicians do for similar tasks.
publicabstract, Optical Coherence Tomography, Papilledema
xvii, 130 pages
Includes bibliographical references (pages 122-130).
Copyright 2016 Jui-Kai Wang