Detecting and correcting false transients in calcium imaging

Abstract

Population recordings of calcium activity are a major source of insight into neural function. Large dataset sizes often require automated methods, but automation can introduce errors that are difficult to detect. Here we show that automatic time course estimation can sometimes lead to significant misattribution errors, in which fluorescence is ascribed to the wrong cell. Misattribution arises when the shapes of overlapping cells are imperfectly defined, or when entire cells or processes are not identified, and misattribution can even be produced by methods specifically designed to handle overlap. To diagnose this problem, we develop a transient-by-transient metric and a visualization tool that allow users to quickly assess the degree of misattribution in large populations. To filter out misattribution, we also design a robust estimator that explicitly accounts for contaminating signals in a generative model. Our methods can be combined with essentially any cell finding technique, empowering users to diagnose and correct at large scale a problem that has the potential to significantly alter scientific conclusions.

Review of data processing of functional optical microscopy for neuroscience

“…The least-squares cost enforces a linear-Gaussian data generation hypothesis, however, a number of nonlinearities in fluorescence dynamics, incompleteness in component discovery, and the non-Gaussian statistics of the photo-diodes all contribute to various extents of errors in demixing. Four main alternatives have emerged, including a robust-statistical approach leveraging a Huber cost function, 87 a contamination-aware generative model approach, 92 a zero-Gamma mixture model, 93 and a deep-learning approach. 94…”

“…A prime example of such an effect is explored in Ref. 92, in which it is shown that the i.i.d. Gaussian noise assumption often considered can create bleed-through between overlapping cells and additional fluorescent biological processes in the tissue.…”

“…While significant in and of itself, it is further noted that these errors can cause errors in the interpretation of the neural activity, including the skewed discovery of location encoding in the hippocampus. 92 To identify the accuracy of optical imaging processing algorithms, a number of avenues have emerged. Specifically, four current available avenues are assessment based on (1) manual annotation, (2) biophysical simulation, (3) local self consistency of global decompositions, and (4) consistency based on external experimental variables.…”

“…The effect is that found profiles that do not match well to the manual annotations could appear as "false alarms" but fall into many categories. They may be actual cells in the data missed by the annotator, they may be algorithmic artifacts caused by merging or splitting parts of cells, 92 or they may be overfitting to noise or neuropil. Thus at best, manual annotations give a lower bound on true hits and missed detections but not much information on false positives.…”

“…Specifically, one can check if in the movie frames, during which a given cell was purported to have fired (a transient burst in the time trace during those frames), if the shape of the cell truly appears in the video. Recent work used this local averaging idea to identify how many algorithms are not locally consistent: i.e., demonstrating that activity from different sources bleed into each other 92 [Fig. 5(c)].…”