| Popis: |
According to World Health Organization (WHO), Cardiovascular Disease (CVD), diabetes, and Chronic Kidney Disease (CKD) are among the top 10 causes of death worldwide. CVDs are the leading cause of death globally with an estimated 17.9 million deaths which constitute 32% of global mortality. The retina is the only organ in the human body that provides direct observation of a risk portion of the microvasculature, hence a unique opportunity for the non-invasive study of several systemic diseases including CVD, stroke, diabetes, hypertension, Diabetic Retinopathy (DR), CKD, and dementia. Retinal investigation can enable early diagnosis, preventive measures, and planning better treatment. There has been increasing interest in applying deep learning (DL) to retinal images to identify the risk of CVDs, its risk factors, and their interplay with cardiovascular risk scores. This thesis investigates retinal images as a source of biomarkers for identifying associations with cardiovascular diseases, cardiovascular risk factors, cardiovascular risk scores, cardiovascular death, mortality, microvasculature diseases (CKD, Diabetic Peripheral Neuropathy (DPN), DR) in a large diabetic cohort, GoDARTS (Genetics of Diabetes Audit and Research in Tayside Scotland) using DL. A breakdown of the main items of work follows. A framework is proposed for generating synthetic datasets, parameterized by difficulty level to test classifiers for medical image analysis. This framework was used for hyperparameter tuning of the DL model and identifying a robust model for subsequent work with real data. To our best knowledge, this had not been addressed in the literature on synthetic medical data at the time the work was carried out. EfficientNet-B2 was chosen as the best performing DL architecture to perform experiments on real retinal images from GoDARTS, following a systematic and comparative performance analysis using synthetic data generated from MESSIDOR images. Several architectures including VGG16, ResNet50, InceptionV3, DenseNet201, and EfficientNet-B2 were evaluated during the analysis. DL methods were employed to investigate biomarkers in retinal images for predicting cardiovascular (CV) risk factors, such as age, sex, Systolic Blood Pressure (SBP), Diastolic Blood Pressure (DBP), High-Density Lipoprotein (HDL), Total Cholesterol (TC), Glycated Haemoglobin (GH), Body Mass Index (BMI), and Triglycerides (Trig) using only retinal images from GoDARTS. The findings indicated that for age prediction on the test dataset, the Mean Absolute Error (MAE) was 3.951 (95% CI 3.908, 3.995) years and R2 was 0.809 (0.804, 0.814). The model achieved an Area Under Curve (AUC) of 0.899 (0.895, 0.903) for sex prediction, with an accuracy of 0.811 (0.806, 0.817), a sensitivity of 0.886 (0.88, 0.891), and specificity of 0.717 (0.708, 0.727). However, the DL models did not yield significant results for predicting other CV risk factors, with an MAE of 5.88 (5.71, 6.07) and R2 of 0.14 (0.1, 0.17) for DBP prediction. Additionally, the model’s performance on test data for stratifying systemic disease outcomes within 12 years from the date of retinal imaging in terms of Area Under Curve (AUC) was 0.74, 0.71, 0.642, 0.633, and 0.57 for CKD, ACD, MACE, DR, and DPN, respectively. Furthermore, a study was conducted to investigate the Predicted Age Difference (PAD), which is the difference between chronological age and the age predicted by DL using retinal images, in individuals with type 2 diabetes. The aim was to find any association between PAD and Major Adverse Cardiovascular Event (MACE) and All-Cause Death (ACD) over a period of 10 years. The results showed a strong statistically significant association. According to Coxph regression analysis that was adjusted for age at imaging and sex, a 1-year increase in retinal PAD score raised the risk of MACE and ACD by 5.8% (hazard ratio (HR) = 1.0587, 95% CI =1.028- 1.089, P = 1.06e-4) and 5.9% (HR = 1.0597, 95% CI = 1.034– 1.085, P = 1.62e-06) respectively. Even after adjusting the coxph model for Pooled Cohort Equations (PCE) Atherosclerotic Cardiovascular Disease (ASCVD) risk score, the associations still remained significant for MACE and ACD events with PAD score. These findings were similar to the results obtained by using the average of predictions from both left and right retinal images (individual-level predictions) when assessing the age predictions from only left-eye and only right-eye retinas. We investigated the potential of a DL approach applied to retinal images for predicting PCE ASCVD clinical risk score and genetic factors, as represented by Genome-Wide Polygenic Risk Scores (GWPRS). The model achieved an R2 of 0.554 (0.528, 0.579) and MAE of 0.107 (0.104, 0.11) for PCE ASCVD risk score, but showed no indication that retinal images contain information related to GWPRS. This investigation represents the first time that the complementarity of retinal and genetic information for CVD risk has been studied using DL. Statistically significant associations were observed between the clinical risk score predicted by the DL model from retinal images and 10-year MACE and Cardiovascular Death (CV death). The Coxph regression analysis showed that a 1% increase in the retinal predicted risk score increases the risk of developing MACE (HR = 1.029, 95% CI = 1.015- 1.042, P = 3.4e-5) and CV death (HR = 1.019, 95% CI = 1.004– 1.034, P = 0.009). Again, these associations are similar to the results obtained using individual-level predictions from only left-eye and only right-eye retinal images. We propose a method for utilizing DL to perform multi-modal image analysis, combining images and tabular (spreadsheet) data. Our method involves converting tabular data to images using a tabular data converted to image (T2I) algorithm. We applied this method to the available CV risk factors in the form of spreadsheet data and combined it with retinal images for stratifying 5-year MACE from the date of retinal imaging. When using only retinal images as input, the DL model achieved an AUC of 66.78% on the test dataset for 5-year MACE stratification. However, when using multi-modal image data (retina + T2I), the DL model achieved a higher AUC of 72.54%. These results suggest that utilizing multi-modal image data has more predictive power for predicting 5-year MACE compared to using retinal images alone. To summarize, the use of DL algorithms on diabetic retinopathy screening images enables accurate predictions of age, sex, DBP, and CVD risk score. Additionally, the algorithms can moderately stratify certain systemic conditions such as MACE, CKD, and ACD. However, further research is needed to validate these results on a larger population and to explore the implementation of these approaches in real-time clinical practices. |
