Table 2 Prediction performance measured with the C-index for models trained on several different sets of features. The first two columns show results for OS and PFS prediction on the Copenhagen post-operative dataset, whereas the last column contains OS prediction performance on the BraTS20 pre-operative dataset. Note that the clinical features for the Copenhagen dataset include age, performance status, and MGMT methylation, while the available clinical features for the BraTS dataset consist only of age. Including the proposed Hausdorff (Hd95) features in the survival model provides an improvement in prediction performance over models that only consider conventional clinical features, tumor size and location.

Hd95 + Clinical + CoM + CEV

0.670 (0.668–0.671)

0.637 (0.634–0.641)

0.619 (0.618–0.620)

Hd95 + Clinical + CoM + TCV

0.657 (0.655–0.659)

0.629 (0.624–0.634)

0.612 (0.611–0.614)

Hd95 + Clinical + CEV

0.666 (0.665–0.668)

0.597 (0.595–0.600)

0.631 (0.630–0.632)

Hd95 + Clinical

0.669 (0.668–0.671)

0.614 (0.612–0.616)

0.612 (0.611–0.613)

Hd95 + CoM + CEV

0.635 (0.633–0.637)

0.629 (0.625–0.634)

0.564 (0.562–0.565)

Hd95 + CEV

0.643 (0.641–0.644)

0.580 (0.575–0.584)

0.594 (0.593–0.595)

Clinical + CoM + CEV

0.624 (0.623–0.625)

0.622 (0.618–0.626)

0.599 (0.598–0.600)

Clinical + CoM + TCV

0.595 (0.592–0.598)

0.613 (0.608–0.618)

0.600 (0.598–0.602)

Clinical + CEV

0.591 (0.589–0.593)

0.567 (0.563–0.572)

0.616 (0.614–0.617)

CoM + CEV

0.548 (0.545–0.551)

0.605 (0.599–0.611)

0.517 (0.516–0.518)

CoM + TCV

0.540 (0.537–0.543)

0.617 (0.612–0.622)

0.536 (0.533–0.539)

Hd95

0.644 (0.643–0.646)

0.552 (0.548–0.556)

0.571 (0.570–0.572)

Clinical

0.574 (0.572–0.576)

0.524 (0.522–0.527)

0.581 (0.579–0.583)

CoM

0.550 (0.548–0.551)

0.591 (0.588–0.594)

0.504 (0.503–0.506)

CEV

0.479 (0.476–0.482)

0.551 (0.547–0.554)

0.534 (0.533–0.535)

TCV

0.525 (0.522–0.528)

0.574 (0.570–0.577)

0.553 (0.551–0.555)

Age

0.509 (0.506–0.513)

0.519 (0.516–0.522)

0.581 (0.579–0.583)