Home / Project Blogs / Brainstorming for Paper of Unsupervised Feature Learning

Goal

I will use this space to flesh out the details of the paper I'd like to write regarding unsupervised feature learning.

Title & Purpose

I am thinking of titling the paper Self–supervised feature learning of Nissl–stained rat brain tissue images.

The idea is to learn (CNNs) or extract (Classical CV) meaningful features from tissue images of high resolution and large scale. We could then cluster homogeneous features and hope the clustering translates to segmenting brain regions with similar features.

In our previous work, we learned that high resolution and large scale allow for precise delineation of brain regions. We also learned that we need a small scale for more global features. The tentative title for this paper doesn't suggest any particular scale or the resolution of the images; given our previous work, we can build on it by doing unsupervised feature learning across different scales, but I haven't thought of how that would be helpful/useful in this unsupervised framework.

4/30/21 – Given this blog, I have decided to change the word unsupervised to self-supervised. Yannic Kilcher's blog explanation mentions that there is really no such thing as unsupervised, everything is still supervised.

Methods

There are three steps I am thinking our experiment needs to have: the feature extraction (and if extracting with a CNN, learn features prior to extracting), segmenting/clustering, and metrics of performance.

Feature Extraction

  • Images alone (flatten image to obtain feature vector)

  • PCA of Image

  • TSNE of Image

  • HoG of Image

  • Tile2Vec

  • SimCLR

Segmentation

  • Watershed

  • Connected Components

  • Region Growing*

  • Graph Cuts*

  • Annealing

Metrics

  • IoU between Human annotation and segmentation

*I am debating having a human component to this; the user can pick which locations they would like to segment (for example, using a coordinate as a seed for Region Growing after the user clicks on an image). Otherwise, we need to find another way to generate seeds; maybe we can use the atlas? although registering the atlas with images is another issue...