Anatomical Structures as Graph Nodes: Representation Before Architecture in Paediatric Brain Tumour Localisation

Background/Objectives: Paediatric brain tumour cohorts are small, and deep networks trained directly on volumetric MRI often perform poorly on them. The usual response is a larger architecture. The contribution of the anatomical representation itself has not been isolated as an experimental factor, so it is not known how much of the difficulty is architectural and how much is representational. This study addresses that gap, using supratentorial/infratentorial tumour localisation as a task defined entirely by anatomical geometry. Methods: In 108 paediatric patients with multi-sequence MRI and manual annotation of the tumour and eight anatomical structures, the task, the cohort and the subject-grouped five-fold cross-validation assignment were held fixed while the representation of the input was varied: 3D convolutional networks on raw volumes, the same networks given coordinate and craniocaudal-profile priors, a geometric rule with no fitted parameters, and a patient-level graph whose nodes are anatomical structures, read out by logistic regression or by a graph attention network. All anatomy downstream of the landmarks was predicted out of fold. Comparisons were patient-paired, with bootstrap intervals and Holm-corrected McNemar tests. Results: Balanced accuracy rose from 0.522 on raw voxels to 0.652 with explicit spatial priors, 0.681 for the parameter-free geometric rule, 0.775 for a five-feature linear readout of the graph and 0.838 for the graph attention network; the same decision layer given manual masks reached 0.918. Only the graph attention network exceeded the strongest convolutional baseline significantly (+0.187, Holm-adjusted p = 0.009). Automated segmentation was associated with a 0.143 gap, driven by spatial mislocalisation rather than by missed tumours. Conclusions: For this geometrically defined task and cohort size, representing anatomical structure explicitly was more effective than learning the same geometry from voxels.