her2-reader / data /paper /paper-data /table_her2low_ladder.tex
nsugar's picture
HER2 reader: ViT-S/16 encoders at every stage, the deployed reader, per-slide predictions and paper data (part 3)
4367928 verified
Raw History Blame Contribute Delete
916 Bytes
% Generated by docs/paper/paper_her2tasks.py (HER2-low vs zero along our training stages, 10 paired seeds)
\begin{tabularx}{\linewidth}{@{}>{\raggedright\arraybackslash}X c c@{}}
\toprule
our ViT-S/16 & AUC (SD) & against stage 2 \\
\midrule
stage 1: self-supervised & 0.662 (0.015) & $-$0.020 (8/10 lower) \\
stage 2: distilled, step 25,000 & \textbf{0.683} (0.014) & -- \\
\quad + JEPA round 1 & 0.661 (0.029) & $-$0.022 (7/10 lower) \\
\quad + JEPA round 2 (\textbf{final}) & 0.649 (0.024) & $-$0.034 (8/10 lower) \\
\quad 0.75 final + 0.25 stage 2 (weights) & 0.651 (0.029) & $-$0.031 (7/10 lower) \\
\quad 0.5 final + 0.5 stage 2 (weights) & 0.665 (0.028) & $-$0.018 (6/10 lower) \\
distillation continued to step 100,000 & 0.649 (0.024) & $-$0.034 (10/10 lower) \\
\quad + JEPA & 0.629 (0.030) & $-$0.054 (10/10 lower) \\
\midrule
H-optimus-0 (teacher) / Phikon & 0.692 / 0.741 & \\
\bottomrule
\end{tabularx}