Fig. 1: Workflow of SERM.
From: Leveraging data-driven self-consistency for high-fidelity gene expression recovery
A subset of the expression matrix is inputted into an autoencoder network. To learn the distribution function that best describes the reconstructed data by the autoencoder, different pdfs are fitted to the histogram of the reconstructed data. Next, an ROI is selected, and histogram equalization is performed on that ROI using the learned pdf in the previous step. The ROI then slides along the x and y direction throughout the expression matrix, and histogram equalization is performed on each ROI. All the regions are then interpolated using bilinear resampling in the final step to impose global consistency.
