Explain the residual PET exterior discrimination

Determine the sources of the residual discriminative performance observed when the six-stream MRI-PET fusion network is evaluated using PET exterior inputs that retain scalp, skull-adjacent tracer uptake, and the outer intensity envelope, particularly whether global count-rate or scanner-reconstruction differences correlated with cohort composition account for the effect.

Background

The paper evaluates shortcut dependence by removing brain content from MRI and PET inputs. The PET exterior condition, which preserves non-brain content, achieves an AUC of 0.608 compared with 0.962 for intact inputs and 0.500 for blank inputs. Although this residual performance is substantially reduced, it indicates that some discriminative information remains outside the brain.

The authors suggest that the residual may arise from scalp and skull-adjacent tracer uptake or from global count-rate and reconstruction differences associated with scanner generation and indirectly correlated with cohort composition. They explicitly identify the result as incompletely explained, leaving the precise source of the residual discrimination unresolved.