cyyu96's picture
Initial upload: MICCAI FLARE 2026 AutoMSC baseline models
ef7ead2 verified

Dataset004_PICAI

Task: Multi-class prostate segmentation (prostate gland + tumor) on prostate MRI (T2W, ADC, HBV) + ISUP grade classification (6-class: 0=Benign/Indolent, 1–5=ISUP grades)

Per-Fold Results — Segmentation

Fold DSC (prostate) NSD (prostate) DSC (tumor) NSD (tumor) DSC mean NSD mean
0 0.960 0.959 0.473 0.490 0.717 0.724
1 0.960 0.960 0.536 0.552 0.748 0.756
2 0.959 0.959 0.634 0.649 0.797 0.804
3 0.947 0.939 0.168 0.188 0.558 0.564
4 0.960 0.960 0.557 0.575 0.759 0.767
Mean 0.957 0.955 0.474 0.491 0.716 0.723
Std 0.006 0.009 0.180 0.179 0.093 0.094

Per-Fold Results — Classification

Fold AUROC AUPRC Bal Acc F1
0 0.731 0.290 0.240 0.236
1 0.720 0.281 0.212 0.183
2 0.746 0.300 0.240 0.229
3 0.702 0.310 0.176 0.063
4 0.736 0.350 0.263 0.224
Mean 0.727 0.306 0.226 0.187
Std 0.017 0.027 0.033 0.072

Notes:

  • Prostate gland segmentation is strong and consistent (DSC 0.957 ± 0.006); the gland is a well-defined, large structure
  • Tumor segmentation is challenging (DSC 0.474 ± 0.180): 71% of cases (1075/1500) have no tumor, so many folds see near-zero tumor Dice dominated by true-negatives
  • Fold 3 has noticeably lower tumor DSC (0.168) and NSD (0.188), likely due to an unfavorable fold split or harder cases
  • Classification: AUROC 0.727 is reasonable for 6-class ISUP grading; Bal Acc (0.226) and F1 (0.187) are near the random baseline (1/6 ≈ 0.167), driven by severe class imbalance (847/1500 cases are grade 0)

Figures

Box plots Case sampling ROC/PRC Confusion matrix FP/FN cases