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 3 HAM10000 Record to Knowledge Graph Mapping

1:
Example Record: (img_ID = ISIC_0024321, age = 67, gender = male, location = face, diagnosis = BCC)
2:
Step 1: Disease Node Selection
3:
$d \leftarrow getDiseaseNode (diagnosis)$ ▹ d → BCC
4:
Step 2: Risk Factor Retrieval
5:
$R_{d} \leftarrow getRiskFactors (d)$ ▹ Returns: sun exposure, fair skin, age, male gender
6:
Step 3: Anatomical Site Selection
7:
$s \leftarrow getAnatomicalSite (location)$ ▹ s → Face
8:
Step 4: Node Embedding Extraction
9:
$e_{d} \leftarrow GAT_embedding (d)$ ▹ 32-dim vector
10:
$e_{r} \leftarrow \frac{1}{| R_{d} |} \sum_{r \in R_{d}} GAT_embedding (r)$ ▹ Mean pooling of risk factors
11:
$e_{s} \leftarrow GAT_embedding (s)$ ▹ 32-dim vector
12:
Step 5: Patient-Aware KG Feature
13:
$f_{K G} \leftarrow e_{d} ⊙ (1 + 0.1 \cdot patient_weight (r))$
14:
return $f_{K G} \in R^{32}$