This repo is created for our IEEE Transactions on Medical Imaging (IEEE TMI, SCI Q1, IF 11.037) paper: An ultrasound based multi-step modality fusion network for identifying the histologic subtypes of metastatic cervical lymphadenopathy (Link). It was accepted on 13 November 2022, and completed by the authors from Chinese Academy of Sciences Institute of Automation and Lanzhou University Second Hospital.
The common ultrasound modalities include B-mode ultrasound (BUS), color Doppler flow imaging (CDFI), ultrasound elastography (UE) and dynamic contrast-enhanced ultrasound (DCE-US). The fusion of them is a challenges for clinicians and they cannot give satisfactory diagnosis performance on some hard problems. And there is lack of specific methods that consider the characteristics of ultrasound modalities.
We proposed the Multi-Step Modality Fusion Network to fuse and analyse ultrasound modalities rationally. Specifically, we adopted:
- grouping and step-by-step fusion strategy
- modality interaction guidance machanism
- self-supervised feature orthogonalization loss.
