Neural Radiance Field (NeRF) Models for Ultrasound Images

点击翻译按钮获取中文摘要

Brief description of the context of the research including potential impact:Neural Radiance Fields (NeRFs) are a new class of deep networks that provide photorealistic 3D reconstruction and rendering of natural scenes, given only a few input photos. The resulting renders include realistic reflections, lighting and other material properties with an accuracy that was previously not possible. Given its impressive performance, it is natural to ask whether it may apply to different input modalities. Medical ultrasound is a relatively low-cost and accessible medical imaging hardware that can be used to flag a number of pathologies and abnormal changes, including cancer, without any radiation risk. However, the produced images are hard to interpret and so require highly trained professionals. Allowing technicians (as opposed to only highly-trained professionals) to automatically flag abnormalities would have a high impact in early disease screening across many areas. This project will investigate the use of NeRFs to reconstruct organs in 3D from weakly-localized ultrasound images, whether in foetuses or adults. It will also explore deformable matching (e.g. with contrastive learning) to register scans and thus automatically flag abnormal tissue evolution. Aims and Objectives:To enhance the capabilities of the very simple and ubiquitous 2D Ultrasound scanner through software.The software should not only aid highly trained Doctors in interpreting ultrasound images, but permit lesser trained healthcare workers to also undertake and interpret effective ultrasound scans. The applications of this work should ideally be multiple, for example increasing detection rates of abnormalities in foetuses, or to aid detection of cancer in adults.Novelty of the research methodology - Applying and adapting novel and cutting-edge techniques from Computer Vision to the Medical field, where images are often of much lower quality, the priorities are different and the challenges are different such as anatomical complexity and practical requirements (scanning times and acquisition rates).Alignment to EPSRC's strategies and research areas (which EPSRC research area the project relates to):Image and vision computingMedical imagingAny companies or collaborators involved - None