Skip to contents

Image Encoding Methods for Biomarker Panels

Source: R/image-encodings.R
image-encodings.Rd

Converts tabular biomarker expression data into 2D image representations suitable for CNN classification. Four encoding methods are supported:

  1. encode_simple_grid: Fixed row-major layout in rows×cols grid

  2. encode_corr_grid: Features ordered by hierarchical correlation clustering

  3. encode_deepinsight: Feature positions learned via t-SNE on training data (Sharma et al., 2019, Scientific Reports)

  4. encode_ratio_image: Pairwise log-ratio matrix (see ratio-image-cnn.R)

These complement the within-sample normalization approaches in within-sample.R and can be combined with train_ratio_cnn and related CNN training utilities.

References

Sharma A et al. (2019). DeepInsight: A methodology to transform a non-image data to an image for convolution neural network architecture. Scientific Reports 9:11399.

Aitchison J. (1986). The Statistical Analysis of Compositional Data. Chapman and Hall.

Quinn TP, Erb I, Richardson MF, Crowley TM. (2020). Understanding sequencing data as compositions: an outlook and review. Bioinformatics, 36(16), 4424-4432.