Search

Image filtering is a technique that can create additional visual representations of the original image. Entropy filtering is a specific application that can be used to highlight randomness of pixel grayscale intensities within an image. These image map created from filtering are based on the number of surrounding neighbourhood of pixels considered. However, there is no standard procedure for determining the correct "neighbourhood size" to use. We investigated the effects of neighbourhood size on the entropy calculation and provide a standardized approach for determining an appropriate neighbourhood size in entropy filtering in a musculoskeletal application. Ten healthy subjects showing no symptoms related to neuromuscular disease were recruited and ultrasound images of their trapezius muscle were acquired. The muscle regions in the images were manually isolated and regions of interest with varying neighbourhood sizes (increasing by 2 pixels) from 3x3 to 61X61 pixels were extracted. The entropy, relative signal entropy over noise entropy, statistical effect size as well as the percentage change of the effect size and instantaneous slope of the effect size was examined. The analysis showed that a neighbourhood size within the range of 21-25 pixels provides the maximum amount of information gained and coincides with a percentage change of the effect size of less than 5% and instantaneous slopes < 0.05.

Peripheral nerve interfaces (PNIs) allow us to extract motor, sensory and autonomic information from the nervous system and use it as control signals in neuroprosthetic and neuromodulation systems. Recent efforts have aimed to improve the recording selectivity of PNIs, including by using spatiotemporal patterns from multi-contact nerve cuff electrodes as input to a convolutional neural network (CNN). Before such a methodology can be translated to humans, its performance in chronic implantation scenarios must be evaluated. We investigated the performance of a CNN-based selective recording approach in the presence of encapsulation tissue, a common immune response to the implantation of a neural interface. This factor was simulated using anatomically accurate computational models of a rat sciatic nerve and nerve cuff electrode. Performance over time was examined in three conditions: training the CNN at baseline only, supervised retraining with explicitly labeled data at periodic intervals, and a semi-supervised self-learning approach. The periodic recalibration approach demonstrated the best results, with an average F1-score of 0.96 ± 0.04, 0.89 ± 0.08, and 0.80 ± 0.08 for SNRs of -5 dB, -10 dB, and -15 dB, respectively, across all time points. Thus, the periodic recalibration approach may be an effective solution to compensate for changes in signal recordings seen over time as a result of encapsulation tissue. The self-learning approach, in which a network is retrained periodically using predicted labels, generally showed degradation in classification performance over time, even as the frequency of training was increased, attributed to an eventual accumulation of mislabeled training data.