ORB-10 Ablation Results

Real-only baseline versus synthetic pretraining followed by real fine-tuning. Both runs use seed 0, 512px input, batch size 16, threshold label mode 0.425, and the same Aju validation split.

Metric	Real only (GPU 0)	Pretrain → finetune (GPU 1)	Δ
Positive Dice	0.3911	0.3921	+0.0011
All Dice	0.2946	0.2986	+0.0040
Foreground precision	0.4315	0.4873	+0.0558
Foreground recall	0.4340	0.3171	−0.1169
Foreground F1	0.4327	0.3842	−0.0485

Run status

Real-only: early stopped at epoch 16; best validation Dice 0.2946
Synthetic pretrain: early stopped at epoch 21; best validation Dice 0.0328
Real fine-tuning: early stopped at epoch 17; best validation Dice 0.2986

Interpretation: synthetic pretraining produced negligible Dice improvement and lower recall/F1, while precision increased. The validation set contains only 2 Aju cases, so this is a baseline comparison rather than a definitive conclusion.

Generated from verified run artifacts on yoseobhan · 2026-09-01 · Internal Aju validation; no patient images or volumes published.