ElixirNet: Relation-Aware Network Architecture Adaptation for Medical Lesion Detection

摘要: Most advances in medical lesion detection network are limited to subtle modification on the conventional designed for natural images. However, there exists a vast domain gap between images and where image often suffers from several domain-specific challenges, such as high lesion/background similarity, dominant tiny lesions, severe class imbalance. Is hand-crafted tailored undoubtedly good enough over discrepant domain? more powerful operations, filters, sub-networks that better fit problem be discovered? In this paper, we introduce novel ElixirNet includes three components: 1) TruncatedRPN balances positive negative data false reduction; 2) Auto-lesion Block is automatically customized incorporates relation-aware operations among region proposals, leads suitable efficient classification localization. 3) Relation transfer module semantic relationship transfers relevant contextual information with an interpretable graph, thus alleviates of lack annotations all types lesions. Experiments DeepLesion Kits19 prove effectiveness ElixirNet, achieving improvement both sensitivity precision FPN fewer parameters.