Boosting lesion annotation via aggregating explicit relations in external medical knowledge graph

Highlights：

• We propose three approaches to model the explicit label relations and capture rich semanticinformation in external medical KG, that is, multi-relational feature aggregation, multi-hop knowledge aggregation, and shortest path reconstruction.

• We present two approaches on aggregating the explicit label relations into ML-GCN, that is, a concatenate method to incorporate multi-graph information and a gated mechanism to aggregate multiple features.

• We evaluate AER-GCN on the ChestX-ray and IU X-ray datasets, and it shows much better performance than recent state-of-the-art models.