Cell2Grid

Convolutional neural networks (CNNs) are the gold standard for image classification tasks. When trained on high-resolution images from whole-slide tissue scans, CNNs are limited by their long training times and their need for large training data sets and computational power. In practice, image resolution is therefore reduced before training.

For biological and clinical questions, the phenotype of individual cells, their size, shape and location in the tissue image are important features for image classification. This information is lost in conventional image down scaling, complicating CNN model interpretation. Especially in health care, understanding why a CNN makes a certain prediction on a biological level is essential to make informed, reliable and trusted decisions for patient treatment.

Cell2Grid is a new image compression algorithm that creates low-resolution, cell-based tissue images. After cell segmentation, individual cells and their features are placed on a target grid. This way, every pixel represents exactly one biological cell and individual cell properties are preserved. Final images are up to 100-times smaller without loss of relevant information.

When training a CNN, Cell2Grid image compression can increase prediction accuracy and improve model interpretability due to the cell-based nature of the data. CNNs train faster and have a smaller memory footprint compared to training on uncompressed images.

Due to their small size, Cell2Grid images simplify storage and sharing of data across labs and institutions, enabling researchers to access the large amount of image data generated in labs across the world.