Cardiovascular risk assessment from retinal images in a diabetic population

This research work has been carried out at the Computer Vision and Image processing department of the University of Dundee in Scotland. We used the retinal images of the GoDARTS dataset, which was born from a project in 1996 that aimed to identify all diabetic patients in the Tayside region. Since diabetic patients may suffer from diabetic retinopathy, they were offered retinal screenings to monitor the retina’s health and prevent the disease. Differently from the fasting glucose test, measuring the glycated haemoglobin offers data about the trend of glycaemia over the last three months. Our idea was to measure the exposure to glycated haemoglobin, which has been calculated as the integral of the curve interpolating all the values of HbA1c for each patient. This thesis aimed to predict the cumulative glycated haemoglobin (HbA1c) from retinal images with a deep learning technique. Then, we aimed to build a model to classify the risk of cardiovascular disease in the population of interest. The main experiment concerned the prediction of the cumulative HbA1c for each available image with the Efficient Net B2 neural network. However, the performance obtained on the test set is not the expected one. The range of the predicted values is significantly narrower than the actual range. Moreover, it seems that the error is linearly dependent on the actual measurement of HbA1c. Nevertheless, we repeated the same experiment on the left-eye and the right-eye datasets and obtained the two average predicted values very similar (303075.2 versus 301050.1 mmol/mol). We concluded that we could use indistinguishably left-eye and right-eye images. Then we considered the trend of the predicted values for each patient and compared it with the trend of the actual values. We aimed to study if the neural network, on average, was able to capture the average trend of the cumulative HbA1c. Despite the values being quite different, the average predicted trend is increasing as the actual trend. As the last step, we considered the position of the predicted line with respect to the actual line. We aimed to investigate a potential association between an above-predicted line and a higher risk of mortality, death due to cardiovascular disease or experiencing a MACE. In all our qualitative analyses, we did not find this association but the numerousness of the datasets was very small to draw any conclusion. For each task of risk assessment, we investigate if males and females were differently distributed and if there were differences in their slope distributions. Thus, we investigated if males have a higher risk of mortality, death due to cardiovascular disease or experiencing a MACE than females and vice versa. We concluded that they have the same probability of experiencing the three evaluated risks.