The Annotation of Novel Datasets for the Training of Cellori (Cell Origin) Spots, a Deep Learning Algorithm for RNA FISH Spot Detection

Overview

Developments in microscopy imaging have allowed for in depth study of biological systems at a level never previously reached. One such method is RNA fluorescence in situ hybridization (FISH), which is commonly used to visualize the spatial and temporal expression of specific genes within cells. Analysis of large datasets can be cumbersome and time consuming, especially when assessing fluorescent localization manually. Attempts have been made to automate this process such as TrackMate which makes use of the Laplacian of Gaussian (LoG) operator. A Gaussian Blur is used to normalize background noise. The Laplacian operator is applied to increase contrast between the background and true spots. LoG requires a manually specified intensity threshold to determine the presence of a spot. The ideal threshold varies across datasets, images and even within different regions of the same cell. Although LoG increases annotation efficiency, its performance and efficiency are limited by a researcher's ability to set an accurate threshold. Developments in deep learning have led to automatic algorithms for RNA FISH spot detection. However, current deep learning algorithms fail to perform on datasets without uniform spots and low background noise. In this research, a novel deep learning algorithm, Cellori (cell origin) Spots was developed for the automatic quantification of RNA FISH. As a result of a custom architecture, custom loss function, and novel training dataset, Cellori Spots is able to outperform LoG and current deep learning approaches to RNA FISH spot detection across a variety of image types.

Competition history

JSHS 2023 Category not listed

Resources

National JSHS abstract book

Related projects

ISEF · 2023