Table 3 The effect of sampling negative instances randomly versus a clustering approach. Under each sampling strategy, we compare various methods to handle class imbalance.
From: Overcoming the limitations of patch-based learning to detect cancer in whole slide images
Negative sampling | Weighing | Dice (% overlap) | Error in \(\sqrt{{d}_{1}{d}_{2}}\) | Confusion matrix |
|---|---|---|---|---|
K-means | \(\propto \) | 59 | 2.58 | \(\left(\begin{array}{cc}5& 3\\ 0& 12\end{array}\right)\) |
\(\frac{1}{\propto }\) | 58 | 3.29 | \(\left(\begin{array}{cc}3& 5\\ 0& 12\end{array}\right)\) | |
\(=\) | 58 | 2.56 | \(\left(\begin{array}{cc}3& 5\\ 0& 12\end{array}\right)\) | |
Minority oversam-pling | 75 | 1.69 | \(\left(\begin{array}{cc}6& 2\\ 0& 12\end{array}\right)\) | |
Random | \(\propto \) | 64 | 1.90 | \(\left(\begin{array}{cc}4& 4\\ 1& 11\end{array}\right)\) |
\(\frac{1}{\propto }\) | 61 | 3.55 | \(\left(\begin{array}{cc}3& 5\\ 0& 12\end{array}\right)\) | |
\(=\) | 58 | 2.21 | \(\left(\begin{array}{cc}2& 6\\ 0& 12\end{array}\right)\) | |
Minority oversam-pling | 57 | 3.18 | \(\left(\begin{array}{cc}3& 5\\ 1& 11\end{array}\right)\) | |
None | \(\propto \) | 30 | 15.37 | \(\left(\begin{array}{cc}0& 8\\ 0& 12\end{array}\right)\) |
\(\frac{1}{\propto }\) | 30 | 14.58 | \(\left(\begin{array}{cc}0& 8\\ 0& 12\end{array}\right)\) | |
\(=\) | 26 | 16.33 | \(\left(\begin{array}{cc}0& 8\\ 0& 12\end{array}\right)\) | |
Minority oversam-pling | 30 | 15.25 | \(\left(\begin{array}{cc}0& 8\\ 0& 12\end{array}\right)\) |
