Table 4 The effect of data augmentation on slide-level tasks using the K-means negative sampling strategy, 256 as the tile stride, and EfficientNet-B0 architecture.
From: Overcoming the limitations of patch-based learning to detect cancer in whole slide images
Random crop | Stain norm | Color jitter | Affi-ne trs | Dice (% over-lap) | Error in \(\sqrt{{d}_{1}{d}_{2}}\) | Cfs. mtx |
|---|---|---|---|---|---|---|
63 | 3.63 | \(\left(\begin{array}{cc}4& 4\\ 0& 12\end{array}\right)\) | ||||
✓ | ✓ | 70 | 2.01 | \(\left(\begin{array}{cc}4& 4\\ 0& 12\end{array}\right)\) | ||
✓ | ✓ | 73 | 1.67 | \(\left(\begin{array}{cc}5& 3\\ 0& 12\end{array}\right)\) | ||
✓ | ✓ | 67 | 1.56 | \(\left(\begin{array}{cc}4& 4\\ 1& 11\end{array}\right)\) | ||
✓ | ✓ | ✓ | ✓ | 75 | 1.63 | \(\left(\begin{array}{cc}6& 2\\ 0& 12\end{array}\right)\) |
