Denoising 1P Calcium Imaging of Olfactory Bulb
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Imaging neural activity in the mouse olfactory bulb is confounded by structured artifacts from vasculature, inter-subject glomerular variability and sensor noise. Existing denoising methods either require paired clean data, assume unbiased noise, and operate at the pixel level, which causes a loss of signal in biologically meaningful regions of interest (ROIs). We introduce Foreground-Background Con- trastive Learning (FBCL), a self-supervised algorithm that learns noise-invariant representations by contrasting glomerular foreground with structured background, thereby isolating localized activations without requiring clean-noisy pairs. We eval- uate FBCL in two ways. We first compute the similarity of the denoised images from FBCL and other methods to ROIs in the images which have been identified by human experts. Second, we quantify the improvement in noise reduction using the Information Theoretic measure of Mutual Information. FBCL outperforms all existing methods significantly on both measures, achieving an average 89% improvement over Noise2Self.
