Histology images can be represented as tissue graphs, where nodes are superpixels that conform to tissue boundaries. This keeps homogeneous regions intact rather than slicing them into rectangular patches, as is common in standard image processing.

To embed these superpixels into a graph, current encoders typically crop them to a regular shape. The new arXiv preprint argues that this cropping step discards the natural size and form of each superpixel, potentially losing structural information.

The proposed method, called TAN, is a size- and form-adaptive masked-convolution encoder. Instead of cropping, it uses masked convolutions to process each superpixel directly, adapting to its irregular shape and dimensions.

The approach is designed to preserve the integrity of tissue regions, which could improve downstream analysis of histology images. The preprint is available on arXiv under the identifier 2610.23475.