Use of 3D rendering and haptic robotics for improved delineation of cancerous tumours in medical imaging data

The ability to efficiently render 3D information is critical to numerous applied applications, including remote telepresence, medical diagnosis, scientific visualisation, and industrial manufacturing. Despite this, people's ability to interact efficiently and explore such data has been greatly outpaced by the ease with which it can be captured. The primary problem is that, typically rich 3D data is explored in 2D using technology that was never designed for this purpose, i.e. a 2D computer screen and mouse. Visualising data in this way makes it difficult to perform tasks that people would find trivial in real life. Rendering in a single sensory modality (vision) also dramatically limits the amount of information which can be examined concurrently. A key example of this is in tumour delineation where clinicians have to identify the 3D volumetric boundary between cancerous and non-cancerous tissue from 3D scanning data (e.g. CT/PET/MRI), but do so by delineating visually in multiple flat 2D image slices. Delineation of cancer volumes is characterised by large inter-clinician variability and reducing this variability is seen as key to improving clinical outcomes (Steenbakkers et al. 2005, 2006). This studentship project will investigate how 3D rendering and haptic robotics can be used to detect and delineate the boundaries of imbedded objects with specific focus on improving tumour detection and the precision and accuracy of tumour delineation in medical imaging data.

Rendering imaging data in 3D has the benefit that users can interactively examine and explore the 3D volumetric imaging data in its true form, rather than indirectly via 2D slices explored on a flat monitor screen using a keyboard and mouse. In addition to this, multi-modal rendering has a number unique of benefits. First, humans are able to make much more precise estimates of the size and shape of objects when provided with multi-modal sensory information (Ernst and Banks 2002). Second, delineation is improved when data from multiple scanning techniques such as CT and PET is available (Steenbakkers et al. 2006). Currently, this relies on superimposing information from multiple techniques in the same modality (vision), resulting in the potential of each scan to obscure the other, or by comparing scanning data from different techniques across numerous imaging windows. Rendering different scans in different modalities offers a way to avoid both of these problems, whilst maintaining the benefits.

Consequently, 3D visual-haptic rendering has a number of unique benefits over traditional techniques, both in terms of decreasing perceptual variability, as well as concurrently examining complementary types of data more efficiently. This offers the potential to reduce inter-clinician variability and improve clinical outcomes. To assess this, we will compare visual-haptic rendering and traditional techniques in terms of (1) detecting and (2) accurately and precisely delineating the boundaries of (a) tumours and anatomical structures in scans from clinical patients and healthy individuals, and (b) simulations of objects that are embedded in a surrounding volume to mimic a tumour. Variability is a key evaluation criteria used in a clinical setting and is one of the key areas for improvement in tumour delineation (Steenbakkers et al. 2005, 2006), however, accuracy can only be estimated indirectly through clinical outcomes. In this project, by using simulated scenes with an objective ground truth, in addition to real scans, we will be able to measure precision and accuracy of delineation, providing an enhanced way in which to evaluate performance.