Problem:

The partners in the project use high throughput optical fluorescence microscopy and calcium imaging to record and track the activity of large, genetically identified neuronal cell populations in freely moving mice over long periods of time (several months). By analyzing these large-scale high-resolution images, our goals are to:

• develop a fully automated image classification algorithm that extracts all neuron outlines, positions and activities

• scale and parallelize classifier training to achieve the best performance

• benchmark the new classifier against human labeling

Solution:

Most state-of-the-art methods for extracting neuronal activity from calcium imaging data are semi-automated or require full supervision from a human expert, making it very difficult to scale to large datasets.

Our solution (so far) relies on (convolutional) dictionary learning models. Dictionary learning and sparse representation make a sparsity assumption instead of independence or uncorrelation (like PCA or ICA), which is more aligned with the sparseness of neuronal activity property.

We propose a structured dictionary learning model that introduces sparse activations of neurons, temporally continuous activations and spatial smoothness in the masks modelling the neurons illumination patterns. The algorithm uses block proximal gradient methods for learning the dictionary elements and activation matrix.