Search

Measurement of corneal nerve tortuosity is associated with dry eye disease, diabetic retinopathy, and a range of other conditions. However, clinicians measure tortuosity on very different grading scales that are inherently subjective. Using in vivo confocal microscopy, 253 images of corneal nerves were captured and manually labelled by two researchers with tortuosity measurements ranging on a scale from 0.1 to 1.0. Tortuosity was estimated computationally by extracting a binarised nerve structure utilising a previously published method. A novel U-Net segmented adjacent angle detection (USAAD) method was developed by training a U-Net with a series of back feeding processed images and nerve structure vectorizations. Angles between all vectors and segments were measured and used for training and predicting tortuosity measured by human labelling. Despite the disagreement among clinicians on tortuosity labelling measures, the optimised grading measurement was significantly correlated with our USAAD angle measurements. We identified the nerve interval lengths that optimised the correlation of tortuosity estimates with human grading. We also show the merit of our proposed method with respect to other baseline methods that provide a single estimate of tortuosity. The real benefit of USAAD in future will be to provide comprehensive structural information about variations in nerve orientation for potential use as a clinical measure of the presence of disease and its progression.

Pulse transit time (PTT) provides a cuffless method to measure and predict blood pressure, which is essential in long term cardiac activity monitoring. Photoplethysmography (PPG) sensors provide a low-cost and wearable approach to obtain PTT measurements. The current approach to calculating PTT relies on quasi-periodic pulse event extractions based on PPG local signal characteristics. However, due to inherent noise in PPG, especially at uncontrolled settings, this approach leads to significant errors and even missing potential pulse events. In this paper, we propose a novel approach where global features (all samples) of the time-series data are used to develop a machine learning model to extract local pulse events. Specifically, we contribute 1) a new noise resilient machine learning model to extract events from PPG and 2) results from a study showing accuracy over state of the art (e.g. HeartPy) and 3) we show that MLPTT outperforms HeartPy peak detection, especially for noisy photoplethysmography data.