Medical Imaging and Computer Vision
Written by Dr. Francis Quek
VISLab had its genesis with a stronger research focus in Computer Vision and Medical Imaging than in HCI. More recently, HCI has gained the greater focus, with Computer Vision and Image/Video Understanding being concentrated on Multimodal Communication Analysis.
Work in Medical Imaging, for example, requires significant collaboration with medical professionals/researchers, and Virginia Tech has just begun expansion into Medicine. Given our busy focus in developing the concepts of Embodied Interaction in HCI, the medical imaging and general computer vision research component has received less attention in recent years. We are interested, however, in continuing collaborations in Computer Vision and Medical Imaging when the opportunities arise. Hence, we keep the flame alive by keeping several most recent exemplar projects on our project list:
- Attentionally-Based Interaction: In this research, we investigated attention manipulation as the conduit of communication between the human and computer to support radiologists and cartographers. By controlling what to look for and where to look, the user can drive underlying machine vision processes to extract the appropriate entities from medical images and maps.
- Brain Surface Analysis: In this research, we employed deformable models to extract the surface contours in 3D medical images. By a first order estimation of local surface normals, we regroup the surface points into Monge patches that allow us to compute analytical surface models everywhere using a variant of the Facet Model. We can then characterize the brain surface in terms of 3D curvature patches that help to identify critical brain surface regions.
- Neurovascular Extraction: In this research, we work with neurosurgeons on the extraction of neurovasculature in 2D and 3D angiograms for surgery planning. We employ a wave-propagation technique that models pixels or voxels are modeled as a medium through which a simulated wave may propagate.The likelihood that a pixel or voxel belongs inside blood vessels is modeled as a reflective index of the medium. Hence, by simply hitting the vascular image with a wave, we can extract the entire vessel tree. The structure is determined by simply tracing the wave back from each distal point in the propagated wave along the normals of the wavefronts.
- Registration of UAV Images: In this research, we performed registration of perspective distorted unmanned air-vehicle images by using an 'eigendisk' technique. We extracted eigendisks around image interest points and used a local voting technique to find consistent spatial interpretations across multi-date images. Our technique proved to be robust to relatively strong perspective distortion.