This project seeks to identify tumor subtypes of triple negative breast cancer using deep unsupervised clustering. This project is part of Thomas' masters thesis, 2020.
The project implements Deep Convolutional Embedded Clustering (https://xifengguo.github.io/papers/ICONIP17-DCEC.pdf) to optimize a joint objective function (L) with both a reconstruction (L_r) and clustering loss (L_c):
Reconstruction loss:
Clustering loss:
Joint optimization:
Model presentation (as illustrated by Guo. et al.) of DCEC, where q is the latent representation vector of an image:
A tissue classificator was implemented to cluster different tissue types: Tumor, stroma, fat and lymphocytes. Examples of tissue types, from left column: fat, stroma, lymphocytes and tumor:
This illustrates Principal Component (PC1 and PC2 on the axes) from the latent representation q (the condensed image presentation between the encoder and decoder). Example of cluster convergence with hyperparameters K=4 clusters, γ=0.1:
Tumor tiles like these were clustered assuming K=4, 8, 12 and 20 clusters. Example of cluster assignment latent representations:
Examples of Tumor tiles for K=4 clusters suggests a color dependency to the cluster assignments: