1 Retinal Circuit Development and Genetics Unit, National Eye Institute, 6 Center Drive, Bethesda, MD 20892, USA.
Int J Neural Syst. 2018 Oct;28(8):1850008. doi: 10.1142/S0129065718500089. Epub 2018 Feb 22.
An important goal in visual neuroscience is to understand how neuronal population coding in vertebrate retina mediates the broad range of visual functions. Microelectrode arrays interface on isolated retina registers a collective measure of the spiking dynamics of retinal ganglion cells (RGCs) by probing them simultaneously and in large numbers. The recorded data stream is then processed to identify spike trains of individual RGCs by efficient and scalable spike detection and sorting routines. Most spike sorting software packages, available either commercially or as freeware, combine automated steps with judgment calls by the investigator to verify the quality of sorted spikes. This work focused on sorting spikes of RGCs into clusters using an integrated analytical platform for the data recorded during visual stimulation of wild-type mice retinas with whole field stimuli. After spike train detection, we projected each spike onto two feature spaces: a parametric space and a principal components space. We then applied clustering algorithms to sort spikes into separate clusters. To eliminate the need for human intervention, the initial clustering results were submitted to diagnostic tests that evaluated the results to detect the sources of failure in cluster assignment. This failure diagnosis formed a decision logic for diagnosable electrodes to enhance the clustering quality iteratively through rerunning the clustering algorithms. The new clustering results showed that the spike sorting accuracy was improved. Subsequently, the number of active RGCs during each whole field stimulation was found, and the light responsiveness of each RGC was identified. Our approach led to error-resilient spike sorting in both feature extraction methods; however, using parametric features led to less erroneous spike sorting compared to principal components, particularly for low signal-to-noise ratios. As our approach is reliable for retinal signal processing in response to simple visual stimuli, it could be applied to the evaluation of disrupted physiological signaling in retinal neurodegenerative diseases.
视觉神经科学的一个重要目标是了解脊椎动物视网膜中的神经元群体编码如何介导广泛的视觉功能。微电极阵列通过同时大量探测来与分离的视网膜接口，记录视网膜神经节细胞 (RGC) 的爆发动力学的集体测量值。然后，记录的数据流通过有效的和可扩展的尖峰检测和排序例程来处理，以识别单个 RGC 的尖峰序列。大多数可用的商业或免费的尖峰排序软件包都将自动步骤与研究人员的判断相结合，以验证排序尖峰的质量。这项工作专注于使用集成的分析平台对使用全场刺激刺激野生型小鼠视网膜时记录的数据进行 RGC 尖峰聚类。在尖峰序列检测之后，我们将每个尖峰投影到两个特征空间上：参数空间和主成分空间。然后，我们应用聚类算法将尖峰分类到单独的簇中。为了消除人为干预的需要，初始聚类结果提交给诊断测试，该测试评估结果以检测簇分配失败的来源。这种故障诊断为可诊断电极形成决策逻辑，通过重新运行聚类算法来迭代地提高聚类质量。新的聚类结果表明，尖峰分类的准确性得到了提高。随后，确定了每个全场刺激期间的活跃 RGC 数量，并确定了每个 RGC 的光响应性。我们的方法在两种特征提取方法中都导致了具有弹性的尖峰分类；然而，与主成分相比，使用参数特征导致错误的尖峰分类更少，特别是对于低信噪比。由于我们的方法对于响应简单视觉刺激的视网膜信号处理是可靠的，因此它可以应用于评估视网膜神经退行性疾病中生理信号的中断。
