Algorithm 1.

retrieve_roi

Input: $\mathbf{x}\in {ℝ}^{H,W}$, $\mathbf{A}\in {ℝ}^{h,w,2}$, K
Output: $O=\{{\tilde{\mathbf{x}}}_k\mid {\tilde{\mathbf{x}}}_k\in {ℝ}^{h_c,w_c}\}$
1:	O = Ø
2:	for each class c ∈ {benign, malignant} do
3:	Ãc = min-max-normalization(Ac)
4:	end for
5:	$\mathbf{A}*={\sum }_c{\tilde{\mathbf{A}}}^c$
6:	l denotes an arbitrary $h_c\frac{h}{H}\times w_c\frac{w}{W}$ rectangular patch on A*
7:	criterion$\left(l,\mathbf{A}*\right)={\sum }_{(i,j)\in l}\mathbf{A}*[i,j]$
8:	for each 1, 2, …, K do
9:	l* = argmaxl criterion(l, A*)
10:	L = position of l* in x
11:	$O=O\cup \{L\}$
12:	$\forall (i,j)\in l^{*}$, set A*[i, j] = 0
13:	end for
14:	return O