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 4 Biomedical Knowledge Graph Construction

1:
procedure Construct( $D, R, S$ )
2:
$G = (V, E)$ with $V = D \cup R \cup S$
3:
for each disease $d \in D$ do
4:
for each risk factor $r \in R$ do
5:
if $hasRiskFactor (d, r)$ then
6:
$E \leftarrow E \cup {(d, r, has_risk_factor), (r, d, isk_for)}$
7:
end if
8:
end for
9:
for each anatomical site $s \in S$ do
10:
if $commonlyLocated (d, s)$ then
11:
$E \leftarrow E \cup {(d, s, commonly_on), (s, d, site_for)}$
12:
end if
13:
end for
14:
end for
15:
for each node $v \in V$ do
16:
$1_{t y p e} (v) \leftarrow \{\begin{matrix} [1, 0, 0] & if v \in D \\ [0, 1, 0] & if v \in R \\ [0, 0, 1] & if v \in S \end{matrix}$
17:
$ϕ_{c l i n i c a l} (v) \leftarrow encodeClinicalAttributes (v)$
18:
$ψ_{t o p o l o g y} (v) \leftarrow [degree (v) / maxDegree, clustering (v)]$
19:
$e_{v} \leftarrow concat (1_{t y p e} (v), ϕ_{c l i n i c a l} (v), ψ_{t o p o l o g y} (v))$
20:
Pad $e_{v}$ to dimension 16
21:
end for
22:
return $G, {e_{v}}_{v \in V}$
23:
end procedure