Fluorescence microscopy segmentation

This is an exploration of the different machine learning techniques to segment images of fluorescence microscopy. The goal of this endeavor is to find a suitable approach that is highly accurate while also feasible with the resources typically available to a lab. All of the protocols below use the FluoCells dataset (https://amsacta.unibo.it/id/eprint/6706/), a set of 283 1600x1200 images of mice brain slices and their corresponding masks.

Supervised learning

U-Net

The U-Net architecture is a fully convolutional neural network created for image segmentation and named after the shape of the network consisting of a contractive and expansive path in the network. Current goals include:

• Improving loss metric for increased penalties between two cells that are very close/touching
• Creating methodology for making images appropriate size before training/ testing using image patches rather than resizing the image. This can be achieved through either:
  • Restricting cropping to the bounds of the image.
  • Extend cropping outside of the bounds of the image and mirror portions of the image as described in https://link.springer.com/chapter/10.1007/978-3-319-24574-4_28

Unsupervised learning

K mean clustering

K means clustering is an iterative process of grouping a collection of observations into K groups. This operates under the assumption that there are at least K groups present within the data.

Results

Using the Fluocell dataset, the 2-mean clustering for each image was determined and loss between the predicted and true masks was calcualted.

DICE LOSS METRICS

• MIN: 0.7181707620620728
• MAX: 1.0
• AVG: 0.9442435503005981
• STD: 0.0578955784440040

These results are not great which is not too surprising. Masks produced from this process yielded areas with many small speckles, not well differentiated cells, and selects artifacts.

W-Net

The W-Net was inspired by the U-Net and (as the name entails) is two consecutive U-Net networks, one functioning as an encoder and the other as a decoder.

Other

Otsu's method

Although not technically within the realm of machine learning, Otsu's method is a process used in automatic image segmentation. There is a one and two dimensional variant, with the two dimensional variant being more adept at segmenting noisy images. Otsu's method operates on the assumption that there are two classes present in the image being segmented.