Perspectives on Imaging and Image Analytics in Ophthalmology

Vupparaboina, Kiran Kumar (2017) Perspectives on Imaging and Image Analytics in Ophthalmology. PhD thesis, Indian Institute of Technology Hyderabad.

[img] Text
Thesis_Phd_EE_3704.pdf - Submitted Version
Restricted to Repository staff only until 2025.

Download (33MB) | Request a copy


Ophthalmology encompasses the study of the structure, functions and diseases of the eye. Eyesight facilitates comprehension of and navigation through the surroundings, and hence crucially determines the quality of life. Unfortunately, it is not uncommon to experience reduced visual acuity, visual impairment and even blindness due to various impairments. Accordingly, diagnosis and management of eye impairments constitute an important focus of ophthalmologists. In recent decades, advances in imaging technology have revolutionized ophthalmological diagnosis, treatment and treatment response monitoring by enabling detection of even minute structural features and anomalies. To take full advantage of existing imaging modalities, it has now become imperative to develop high-performance image analytics. At the same time, there still remains room for new imaging techniques that capture target features accurately, yet economically, without causing inconvenience to subjects. In this dissertation, we take an engineering perspective, develop imaging and image analytics methods for ophthalmological applications, and validate such methods statistically. We recognize that a comprehensive coverage of the vast discipline of ophthalmology is infeasible within the scope of a doctoral dissertation. Therefore, in consultation with collaborating ophthalmologists, we selected a few problems that represent diverse aspects of the discipline, are of current clinical signi�cance, and pose signi�cant engineering challenge. For instance, an important problem arises in planning surgical treatment pertaining to the oculofacial region. The state-of-the-art techniques involve 3D modeling based on imaging and other methods. However, a 3D imaging system that is fast, reliable, portable, easily deployable, yet economical, has hitherto remained elusive. To �ll this gap, we propose such a desirable system that consists of a exible network of cheap o�-the-shelf cameras, and makes use of structured light illumination. We validate the proposed technique using a 3D-printed model of head with satisfactory agreement between the reconstructed object and its reference model. Another set of problems arise in relation to early diagnosis and disease monitoring. Traditionally, clinical decisions have been made upon visual inspection of medical images, and susceptible to imprecision. To improve reliability, there is a growing consensus to adopt algorithmic quantitation methods in the ophthalmological community. However, existing algorithms are often generic, and do not target speci�c circumstances at hand. In contrast, taking a targeted approach, we study speci�c clinical problems related to retina, choroid and cornea, and develop image analytics algorithms based on OCT images. Speci�cally, we segment retinal damage based on level-set methods that statistically di�erentiate between the damaged vii and the undamaged regions. While segmenting the choroid, we exploit structural dissimilarity between the granularity exhibited by the choroid and the uniformity seen in the sclera by minimizing the structural similarity (SSIM) index. We further employ tensor voting to endow the algorithmic delineation with a smoothness that is generally associated with human delineation. In contrast, in measuring corneal detachment, we adopt a geometry-based approach as the problem at hand demands. We also investigate analytics of fundus images with the aim of quantifying retinal hard exudates (HEs). While HEs could be either bright or faint, the existing methods do not perform well in detecting the latter. Against this backdrop, we propose an algorithm that adapts to the intensity pro�le of HEs, and delivers improved detection performance, and hence superior quanti�cation accuracy. Each of the proposed algorithms is thoroughly tested for accuracy and statistical reliability. Interestingly, such testing is not straightforward as the datasets are generally not universally accessible and hence not standardized. Secondly, medical quantitation often involves an element of subjective judgment of trained professionals, and the ground truth may not exist as such. To alleviate these di�culties, we propose to use the variability in manual measurements as the reference. We declare an algorithm to perform on a par with manual attempts, if the variation between algorithmic performance and the average manual performance is the same as the reference manual variability. In general, we reported the ratio of the former to the latter as a quotient performance measure. The proposed algorithms are found to achieve such quotient measures close to unity, as desired. Importantly, our algorithms have not remained con�ned to academic exercises. Some of our image analytics algorithms, including choroidal thickness and volume quanti�cation, and area measurement of retinal hard exudates, have already been incorporated in clinical practices. Further, the proposed choroidal image analytics method has facilitated clinical �ndings, such as the fact that choroidal vascularity is preserved in high myopic eyes. We close by reiterating our overall learning that single engineering solution to a broad class of clinical problems has generally been found infeasible. Thus, in the general clinical context, we recommend that a targeted solution to a speci�c problem should be developed from an appropriate engineering perspective, and its reliability should be tested statistically. In the same breath, we hasten to add that certain tools have found use beyond their target applications with minimal repurposing.

[error in script]
IITH Creators:
IITH CreatorsORCiD
Item Type: Thesis (PhD)
Subjects: Electrical Engineering
Divisions: Department of Electrical Engineering
Depositing User: Team Library
Date Deposited: 20 Dec 2017 10:22
Last Modified: 16 May 2019 10:15
Publisher URL:
Related URLs:

Actions (login required)

View Item View Item
Statistics for RAIITH ePrint 3704 Statistics for this ePrint Item