Figure 3. Workflow for building a CGA-based classifier. (a) Gland segmentation is performed on a region of interest.

CGA methodology (highlighted within the dashed lines) leverages the gland segmentation to compute CGA features. (b) Angle calculation and (c) Subgraph computation is performed on the segmented image. (d) Angular co-occurrence matrix aggregates co-occurring gland angles within localized gland networks. (e) Mean, standard deviation and range of second order statistics (shown via differentially colored gland networks) create a set of CGA features for the region. (f) A CGA-based classifier can then be built using the features obtained from (e) to distinguish the categories of interest (either cancer versus benign regions or BCR versus NR).