/HT29-Morpho

Unsupervised cell morphology classification

Primary LanguageJupyter NotebookMIT LicenseMIT

Unsupervised human HT29 colon cancer cell morphology classification based on Deep InfoMax (link to paper)

Tested on Ubuntu 20.04 with pyTorch 1.12 and CUDA 11.6

2023-04-27 update:

  • Improved code readability.

2023-04-25 update:

  • Added prior distribution learning;
  • Added binary GMM for data preparation;
  • Added UMAP for vector dimension reduction;
  • Added N-class GMM for morphology clustering;
  • Modified dataloader for loading 4D batch;
  • Modified network structures of encoder and summarizer;
  • Modified loss functions accordingly (see report).
  • Update README and Makefile.

Preparation

  1. Navigate to the repo root;
  2. Copy all images in the dataset to ./datasets/Images;
  3. In the terminal, run "make";
  4. You should see a new tag in browser pops up.

Training models

  1. Run "prepare.ipynb" to generate binary masks via cell →run all;
  2. Run "train.ipynb" to train all models via cell →run all;
  3. Wait until the training completes, use visdom to check status.

Testing pre-trained models

  1. Run "test.ipynb" via cell →run all;
  2. Wait until the testing completes, results can be found under ./saves.