Identify effective strategies for improving resection-cavity segmentation

Determine whether foreground oversampling, a connected-component loss specialized to resection cavities, or a resection-cavity-specialized segmentation head can substantially improve resection-cavity lesion-wise Dice beyond the performance achieved by the reported nnU-Net ensemble and post-processing pipeline.

Background

The reported pipeline achieves strong lesion-wise Dice scores for enhancing tumour, tumour core, and whole tumour, but its resection-cavity (RC) score remains substantially lower at 0.549. The BoundaryExpand post-processing stage provides a modest and consistent RC improvement, while more aggressive filtering and ungated dilation either degrade performance or produce inconsistent results.

The authors therefore identify training- and detection-focused approaches as potentially more effective avenues for closing the RC performance gap. Specifically, they conjecture that foreground oversampling, an RC-specific connected-component loss, or an RC-specialized prediction head may yield greater improvements than additional post-processing, but these approaches are not evaluated or resolved in the paper.