Ophthalmology (Eye & Ear Hospital) - Theses

Permanent URI for this collection

http://hdl.handle.net/11343/176

Filters

2021 1
1
Reset filters

Settings
Statistics
Citations
Author: Arslan, Janan ×

Search Results

Now showing 1 - 1 of 1
  • Item
    No Preview Available
    Application of artificial intelligence to analysis and progression of geographic atrophy in age-related macular degeneration
    Arslan, Janan ( 2021)
    Background: Geographic atrophy (GA) is a debilitating eye disease affecting adults 50 years and older and is the late stage of age-related macular degeneration (AMD). It is characterised by the presence of lesions in the retina, which are areas of cellular death that cause irreversible vision loss. The dark lesions are usually surrounded by bright regions of hyperfluorescence. GA is a progressive condition with no therapeutic treatments available to arrest or slow down its course. GA is highly variable in its progression, with some patients progressing faster to vision loss than others. While a plethora of publications are available on the investigation of this disease, there remains an uncertainty with regards to our understanding of the disease, its manifestation and progression. Several clinical and non-clinical features along with different statistical modelling techniques have been explored to date. However, there is a lack of consensus regarding the most appropriate model or features that are thought to be useful in a real-time clinical setting. Current available tools only annotate GA lesions using the semi-automated region-growing algorithm. In addition to its semi-automation (which still requires manual and time-consuming work from clinical graders), the current tool can only retrospectively assess progression based on all available data; the software does not have predictive capabilities. Ideally, being able to predict the disease state and progression of GA at initial consultation would be of immense clinical benefit to both the ophthalmologist and patient. Artificial intelligence (AI) has been suggested over the past decade to having the capabilities to achieve such a goal. However, the publications in the GA-AI space thus far have predominantly focused on the automation of the primary GA feature of interest, lesions, with minor focus on GA progression. The automation and detection of hyperfluorescence has been neglected altogether. Thus, there is great scope to apply AI to understand GA manifestation and progression. Purpose and Aims: This thesis describes research to automate the diagnosis of GA, as well as understand the patterns of GA progression using a combination of AI, image processing, mathematics, statistics and computer science. The aims of the thesis were to (1) automate the detection of lesions, but with greater accuracy and efficiency as compared to the current work in the literature, while simultaneously automating hyperfluorescence areas for the first time ever to offer ophthalmologists complete automation of GA features, (2) investigate patterns of GA progression by evaluating prospective regression models by applying our current understanding of the clinical and physical assumptions of GA growth, (3) understand the impact of epistemic uncertainty – a metric to measure variability due to incomplete knowledge of a disease or process – in GA modelling, (4) quantify and identify subgroups of GA lesions and hyperfluorescence areas using machine learning (ML) techniques in order to explain whether varying shapes and sizes of GA features could explain varying progression rates, (5) extract and rank relevant imaging features using ML and develop a statistical model for GA growth, and (6) present a preliminary software platform which can be extended beyond the scope of this thesis. The Methods and Results achieved for the Aims are summarised below. Methods: Data Subjects included in this thesis were AMD participants diagnosed with GA. Data were collected from macular studies from the Centre for Eye Research Australia (CERA) and from a private ophthalmology practice. Extracted data included basic demographic information, such as age and sex, and the imaging modality fundus autofluorescence (FAF) – which produces a grayscale image that enhances the detection of lesions and hyperfluorescence regions. Ground truth labels were also annotated by two graders, and the intraclass correlation coefficient (ICC) was used to measure consistency between graders. Aim 1 Lesion automation included a pipeline of image pre-processing (e.g., contrast limited adaptive histogram equalisation) and semantic segmentation using the U-Net architecture and ground truth labels. Metrics used for assessing lesion automation included the Dice similarity coefficient (DSC), specificity, sensitivity, mean absolute error (MAE), accuracy, and precision. The automation of the hyperfluorescent regions were conducted using pseudocolouring techniques. This involved the application of a colour palette to the FAF images (i.e., specifically the JET colour map), which highlighted the changing intensities in hyperfluorescent regions. Hyperfluorescence was automatically extracted by specifying the colour ranges which covered hyperfluorescent regions. For hyperfluorescence automation, no ground truth labels were available, and thus qualitative assessments were conducted. Aims 2 & 3 Patterns of GA growth – the cumulative increase of total GA surface area within the retina over time – were investigated by evaluating a range of regression models: linear, logarithmic, power, exponential, quadratic, and quadratic without linear term. The physical and clinical assumptions of growth were included as part of the analyses (i.e., growth cannot be infinite; there is finite retinal area in which the lesion growth will hit and then the growth curve should plateau). For Aim 3, the prospective models were evaluated using epistemic uncertainty, particularly model structure uncertainty. Uncertainty was quantified as U=1-r2 where r2=1-SSR/SSO and SSR/SSO is the sum of square residuals divided by the total sum of squares in the data. The metric U is the proportion of total unexplained variability not accounted for by the regression model. The smaller the U, the better the fit of a model. Aim 4 Prospective patterns and groupings of lesions and hyperfluorescent areas were elucidated using several unsupervised clustering methods. Cluster performance was measured using Silhouette coefficient (SC), Davies-Bouldin Index (DBI), and Calinski-Harabasz Index (CHI). If appropriate groups were identified, nomenclature was assigned to these groups based on the patterns seen within the cluster. Aim 5 Feature extraction and ranking was conducted using the ML algorithm XGBoost. An image-based linear mixed-effects model was designed to account for slope change based on within-subject variability and inter-eye correlation. Metrics used to assess the linear mixed-effects model included marginal and condition R2 (RM2 and RC2), Pearson’s correlation coefficient (r), root mean square error (RMSE), mean error (ME), MAE, mean absolute deviation (MAD), Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and log-likelihood. Aim 6 Aim 6 involved the design of a software that will be built upon further but is beyond the scope of this thesis. This software was built using Python and PyQT5. Results: Data The final dataset consisted of 702 images from 51 patients. The cohort consisted of 99 eyes, 49 left eyes (49.5%) and 50 right eyes (50.5%). A total of 359 images were for the left eye and 343 images were for the right eye. The cohort consisted of 38 females (74.5%) and 13 males (25.5%) with an average age of 76.7 +/- 8.9 years. Total follow-up time was 61.5 +/- 25.3 months. For manual annotation of ground truth labels, the ICC for consistency between the two graders was 0.9855 (95% CI: 0.9298, 0.9971). Aim 1 For lesion automation, the DSC was 0.9780 +/- 0.0124, sensitivity was 0.9903 +/- 0.0041, specificity was 0.7498 +/- 0.0955, MAE was 0.0376 +/- 0.0184, accuracy was 0.9774 +/- 0.0090, and precision was 0.9837 +/- 0.0116. Assessments of hyperfluorescence areas revealed three distinct regions of hyperfluorescence. They have been named as: early-stage hyperfluorescence, intermediate-stage hyperfluorescence, and late-stage hyperfluorescence. Qualitatively, both the lesion and hyperfluorescence automations produced visually accurate outcomes. The automations ran 10 times faster than the current semi-automated segmentation. Aims 2 & 3 The linear regression model was identified as most representative of peak GA growth. It had the lowest average uncertainty (U = 0.025), highest average coefficient-of-determination (R2 = 0.92), and applicability of the model was supported by a high correlation coefficient, r, with statistical significance (P = 0.01). Given the assumptions of GA growth (i.e., plateau due to limited retinal space) and the results confirming peak growth was linear, a hypothesis was made that GA growth overall followed a sigmoidal growth curve from disease onset to plateau. Using records from existing patients, the sigmoidal hypothesis was tested in case studies with sufficient longitudinal data from clinical follow-ups. The sigmoidal growth curve revealed a closer fit in these rare cases, and it is recommended that future GA progression studies include additional data to further validate the lower- and upper-tail ends of the sigmoidal function. Aim 4 GA lesions, together with early-, intermediate-, and late-stage hyperfluorescence areas were subject to cluster analysis using the method of k-Means. Meaningful clusters were identified for lesions, early- and late-stage hyperfluorescence regions, with each GA feature having k=3 clusters. For lesions, SC = 0.799, DBI = 0.180, and CHI = 4313.316. For early-stage hyperfluorescence, SC = 0.597, DBI = 0.915, and CHI = 186.989. For late-stage hyperfluorescence, SC = 0.593, DBI = 1.013, and CHI = 217.325. No meaningful clusters were identified for intermediate-stage hyperfluorescence. Aim 5 A linear mixed-effects model with 15 FAF imaging features produced average RM2 = 0.84 +/- 0.01, RC2 = 0.95 +/- 0.002, r = 0.97 +/- 0.006, RMSE = 1.44 +/- 0.06, ME = 0.09 +/- 0.15, MAE = 1.04 +/- 0.06, MAD = 3.52 +/- 0.67, AIC = 1718.41 +/- 11.08, BIC = 1798.99 +/- 11.20, and log-likelihood = -839.21 +/- 5.54. The model was also executed using the popularised square root and log(Yi+1) transformations, however, the original scaled GA total area using mm2/year produced the best results. Aim 6 A preliminary and user-friendly software was designed, which allows the end-user to select a baseline FAF image and then automatically segment lesion, early-, intermediate-, and late-stage hyperfluorescence regions from the image. Then, if the user wishes to toggle from image to image with all features being automatically updated, they simply need to select the ‘Next’ button rather than reloading a new image. A demonstration video can be found at https://youtu.be/WidWf70MKX4. Conclusions: This thesis adopted a multidisciplinary approach combining AI, image processing, mathematics, statistics and computer science to automate GA feature extraction, ranking of GA features, understanding the pattern of GA growth, and the development of a preliminary GA growth model and software. There are many clinical impacts of this work. The automation of clinical processes enables clinicians to focus on their patients and not on the repetitive and time-consuming components of diagnostic and prognostic processes. Furthermore, the automation also guarantees that both clinicians and patients save considerable time in consultations. Additionally, by understanding the underlying patterns of GA growth, we could take more decisive steps to develop appropriate interventions to arrest the progression of the disease or cure it altogether. Future development of the software could extend to multimodal imaging, such as spectral domain optical coherence tomography images or colour fundus photographs. The addition of other predictors is possible, such as genetic heritability.