CSMIC

Raw input image
labels where the segmentations are(what the neural network is looking for)
another example
more labels

Inspiration

Given the various datasets for this data, we found that the cell segmentation dataset was the most interesting because of its potential to expand into other biological identifications, such as other animals' neurons. Also, the cell segmentation dataset was by far the cleanest and most accessible with a machine learning algorithm.

What it does

We trained a model based on the U-net neural network architecture to be able to identify and accurately segment a given image for multiple cells on their boundaries.

How we built it

We trained the data model on pictures of neurons in the brain of a fruit fly, and then tested it against neurons in mice.

Challenges we ran into

The input data was much larger than the test data, so we had to do a lot of data augmentation to increase the accuracy of the model and increase the robustness of the dataset, by taking our original 125 datapoints and turning it into over 2000 datapoints.

Accomplishments that we're proud of

By generating a lot of data-augmented points, we were able to create a robust model that was able to generalize to variance in data.

What we learned

In order for a model to be more accurate, the input data must be consistent with the testing data. Also, the quantity of the data is extremely important for training neural networks