Image-Driven Multimodal Deep Learning for Skin Cancer Diagnosis: Cross-Attention Fusion of Dermoscopic Imaging Clinical Data and Knowledge Graphs

. 2026 Jul 25;26(15):4728. doi: 10.3390/s26154728

Algorithm 1 Data augmentation pipeline

1:
procedure Augment( $I, p$ )
2:
$I_{a u g} \leftarrow I$
3:
if random() $< p$ then
4:
$I_{a u g} \leftarrow RandomRotation (I_{a u g}, angles = [- 90, 90])$
5:
end if
6:
if random() $< p$ then
7:
$I_{a u g} \leftarrow RandomHorizontalFlip (I_{a u g})$
8:
end if
9:
if random() $< p$ then
10:
$I_{a u g} \leftarrow RandomVerticalFlip (I_{a u g})$
11:
end if
12:
if random() $< p$ then
13:
$I_{a u g} \leftarrow ColorJitter (I_{a u g}, brightness = 0.4, contrast = 0.4)$
14:
end if
15:
if random() $< 0.4$ then
16:
$I_{a u g} \leftarrow GaussianBlur (I_{a u g}, kernel = 9)$
17:
else if random() $< 0.4$ then
18:
$I_{a u g} \leftarrow MotionBlur (I_{a u g}, kernel = 9)$
19:
end if
20:
if random() $< 0.3$ then
21:
$I_{a u g} \leftarrow RandomShadow (I_{a u g})$
22:
end if
23:
return $I_{a u g}$
24:
end procedure