An end-to-end recurrent compressed sensing method to denoise, detect and demix calcium imaging data.

Area of Science:

• Neuroscience
• Computational Biology
• Bioimaging

Background:

• Two-photon calcium imaging enables large-scale neuronal activity recording at cellular resolution.
• Analyzing this data requires robust, automated pipelines for neuron segmentation and activity trace extraction.
• Current methods face challenges with background noise, overlapping neurons, and processing speed.

Purpose of the Study:

• To develop an end-to-end deep learning method for simultaneous neuron segmentation and activity trace extraction.
• To achieve high-speed, automated, and parameter-free analysis of calcium imaging data.
• To overcome limitations of existing methods in accuracy and efficiency.

Main Methods:

• Developed DeepCaImX, a deep learning model integrating an iterative shrinkage-thresholding algorithm and a long-short-term-memory network.
• Employed a multi-task, multi-class, multi-label segmentation approach using a compressed-sensing-inspired neural network.
• Trained the network with simulated datasets and validated with in vivo experimental data.

Main Results:

• DeepCaImX simultaneously generates accurate neuronal footprints and extracts clean neuronal activity traces.
• The method achieves very high processing speeds without manual hyper-parameter tuning.
• Outperformed existing state-of-the-art methods in segmentation quality, trace extraction, and processing speed.

Conclusions:

• DeepCaImX offers a significant advancement for analyzing large-scale calcium imaging data.
• The tool is highly scalable and beneficial for mesoscale calcium imaging analysis.
• Represents the first neural network capable of simultaneous accurate segmentation and clean trace extraction.