Survey of Explainable AI Techniques in Healthcare

. 2023 Jan 5;23(2):634. doi: 10.3390/s23020634

Algorithm 1 Class activation mapping.
Require: Image $I^{C} (H, W)$ ; Network N
Ensure: Replace FC layer with average pooling layer in Network N
procedure CAM(I, N)
N(I)	▹ Input image into network
$W_{k}^{c} \leftarrow (w_{1}, w_{2}, w_{3}, . . ., w_{k})$	▹ Get weights from average polling layer
$F_{k}^{c} \leftarrow (f_{1} (x, y), f_{2} (x, y), f_{3} (x, y), . . ., f_{k} (x, y))$	▹ Feature map of the last convolution layer
layer
$M_{c} (x, y) = \sum_{k} w_{k}^{c} f_{k} (x, y)$	▹ Weighted linear summation
$M_{c} (x, y) = \frac{1}{H W} \sum_{i}^{H} \sum_{j}^{W} M_{c} (x, y)$	▹ Normalize and up-sample to Network input size
$M_{c} (x, y) = R e l u (M_{c} (x, y))$	▹ Final image heat map
end procedure