Medical Professional Enhancement Using Explainable Artificial Intelligence in Fetal Cardiac Ultrasound Screening

. 2022 Feb 25;10(3):551. doi: 10.3390/biomedicines10030551

Algorithm A2 Training algorithm of the cascade graph encoder.

1:
Input: Set of Barcode-like Timelines ${B}$ ; Stride Size s; Kernel Size w; Max Iteration $n_{\max}$
2:
Output: Network Weights $θ_{E_{all}}$ , $θ_{D_{all}}$
3:
Initialize network weights $θ_{E_{all}}$ , $θ_{D_{all}}$
4:
$n = 0$
5:
while $n < n_{\max}$ do
6:
B⟵ sample from ${B}$
7:
$i = 1$
8:
while $(i - 1) s + w \leq T_{\max} (B)$ do
9:
Normal-training( $i, B$ )
10:
if n is a multiple of ten then
11:
Synthesize ${\tilde{B}}_{heart}$ and ${\tilde{B}}_{vessels}$
12:
Regularization-training( ${\tilde{B}}_{heart}, {\tilde{B}}_{vessels}$ )
13:
end if
14:
$n = n + 1$
15:
$i = i + 1$
16:
end while
17:
end while
18:
procedureNormal-training( $i, B$ )
19:
$\underset{θ_{E_{all}}, θ_{D_{all}}}{minimize} l_{main} (i, B; θ_{E_{all}}, θ_{D_{all}}) + \sum_{R, θ} l_{rec} (i, B (R); θ)$
20:
end procedure
21:
procedureRegularization-training( ${\tilde{B}}_{heart}, {\tilde{B}}_{vessels}$ )
22:
$\underset{θ_{E_{all}}, θ_{D_{all}}}{minimize} l_{main} (0, {\tilde{B}}_{heart}; θ_{E_{all}}, θ_{D_{all}}) + \sum_{R, θ} l_{rec} (0, {\tilde{B}}_{heart} (R); θ) + l_{vpl} ({\tilde{B}}_{heart}; θ_{E_{all}})$
23:
$\underset{θ_{E_{all}}, θ_{D_{all}}}{minimize} l_{main} (0, {\tilde{B}}_{vessels}; θ_{E_{all}}, θ_{D_{all}}) + \sum_{R, θ} l_{rec} (0, {\tilde{B}}_{vessels} (R); θ) + l_{vpl} ({\tilde{B}}_{vessels}; θ_{E_{all}})$
24:
end procedure