#Neuron Finding

Team name - minions

Team members: Chirag Jain, Sharmin Pathan

Approach: NMF and median filtering to find neurons in calcium imaging data

Technologies Used:

Python 2.7
Thunder
Apache Spark
Matplotlib
Median Filtering
NMF

Introduction:

Calcium Imaging is a dominant technique in modern neuroscience for measuring activity of large population of neurons. Several challenges remain in how to best process and extract signals from these data.

Problem Statement:

The challenge is to take the time varying images (left) and extract regions of interest (right) that correspond to individual neurons.

Dataset:

The datasets were taken from http://neurofinder.codeneuro.org/ which includes 19 training sets and 9 test sets.

Preprocessing:

• A structured matrix factorization approach to analyzing calcium imaging recordings of large neuronal ensembles.
• Demix spatially overlapping regions
• Denoise and deconvolve spiking activity of each neuron
• Median filtering of the images

Flow:

• Load the datasets

• NMF (Non-negative matrix factorization) to express the spatiotemporal fluorescence activity as a product of two matrices: a spatial matrix that encodes location of each neuron, and a temporal matrix that characterizes the calcium concentration of each neuron over time.

• The NMF algorithm is contructed using the following attributes

  • k=5: number of components to estimate per block
  • max_iter=50: maximum number of algorithm iterations
  • percentile=99: value for thresholding (higher means more thresholding)
  • overlap=0.1: value for determining whether to merge (higher means fewer merges)

• The algorithm is then fit to the data. The algorithm is applied to chunks of data and not the entire data with a required padding. We kept the chunk size to be between 30-60 and padding of 15x15 or 32x32.

• Merge overlapping regions that are similar to one another more than the specified overlap.

• A list of the identified regions is created.

• The identified regions list is then saved into a json in the required format for submission.

Execution:

• Ensure the following are installed
  • Thunder
  • Apache Spark
  • Matplotlib
• The program takes four command line arguments, namely
  • name of the dataset
  • no. of components to estimate per block
  • chunk size
  • padding
• For example: python neuro.py 00.00.test 5 50 25

Tuning the accuracy:

#####We applied different chunk sizes and padding for different datasets since their sizes varied in terms of the no. of images.

Dataset	k	Chunk size	Padding	Regions
00.00	5	30x30	15x15	356
00.01	32	32x32	32x32	503
01.00	32	32x32	32x32	326
01.01	32	32x32	32x32	264
02.00	5	100x100	0x0	108
02.01	5	100x100	0x0	099
03.00	5	60x60	15x15	439
04.00	5	60x60	0x0	201
04.01	5	30x30	15x15	331

Performance:

Stuff we tried:

• Tried coverting the images to grayscale. But, on studying the NMF algorithm and the datasets, we got to know that the images were already into grayscale format with channel=1
• Performed median filtering but that resulted into less no. of regions. So, we removed it from our final code.
• Image registration to align multiple scenes into a single integrated image. We studied over this topic but could not actually implement it due to lack of time.

Extras:

• Made a submission on codeneuro

Challenges:

• It was difficult to come up with possible image preprocessing techniques that could be applied.
• Understanding Calcium Imaging and related sciences.