We develop advanced image analysis algorithms for extracting clinically useful information from raw medical image data. This is what we call Semantic Imaging. Applications include multi-modal segmentation, anatomy recognition, object localization, image classification and structure-guided image registration.
DeepMedic is our software for brain lesion segmentation based on a multi-scale 3D Deep Convolutional Neural Network. It has been developed by PhD student Konstantinos Kamnitsas. The system has been shown to yield excellent performance (winner of the ISLES 2015 competition) on challenging lesion segmentation tasks, including traumatic brain injuries, brain tumors, and ischemic stroke lesions. Code is available on GitHub.
Structured-guided Image Registration
See our recent work on Image-and-Spatial Transformer Networks
Segmentation with Shape Priors
See Lee et al. 2019 (TMI) and Oktay et al. 2018 (TMI)
NeuroNet – All-in-One Brain Structure Segmentation Tool
NeuroNet is a deep convolutional neural network mimicking multiple popular and state-of-the-art brain segmentation tools including FSL, SPM, and MALPEM. The network is trained on 5,000 T1-weighted brain MRI scans from the UK Biobank Imaging Study that have been automatically segmented into brain tissue and cortical and sub-cortical structures using the standard neuroimaging pipelines. The paper received the Philips Impact Award at MIDL 2018. Code is here: https://github.com/DLTK/models
Spine Image Analysis
Spinal imaging is an important tool in diagnosis, therapy and intervention of various disorders and pathologies. A wide spectrum of spinal pathologies such as traumatic injuries, including damages to the vertebral bodies or spinal cord, but also spinal deformities such as scoliosis can be assessed with imaging techniques.
The focus in this project is on quantitative spinal imaging and image-based analysis of spinal pathologies. In this context, we are developing algorithms for vertebra localization, longitudinal image registration, and whole spine segmentation in CT and MRI.
Motion Compensation in Time-Lapse Microscopy
We have developed a processing pipeline based on state-of-the-art methods for background motion compensation, cell detection, and tracking which are integrated into a novel semi-automated, learning based analysis tool. Motion compensation is performed by employing an efficient nonlinear registration method based on powerful discrete graph optimisation. Robust detection and tracking of cells is based on classifier learning which only requires a small number of manual annotations. Cell motion trajectories are generated using a recent global data association method and linear programming. Our approach is robust to the presence of significant motion and imaging artifacts.