Autonomous Virtual Robots for Image Exploration Application to contour and fiber tracking in medical images (LNN.6122)

Research
Medical imaging modalities such as Magnetic Resonance (MR) and Computed Tomography (CT) are now offering acquisition techniques which generate high resolution 3D images. The anatomy and the function of different organs can be revealed with high temporal and spatial resolutions. However, the amount of image data involved in a comprehensive patient study is massive, particularly with cardiac stress studies including various stress levels. The shortage of highly trained personnel, as well as the long and complicated procedures to analyse the data manually (or semi-automatically) create a bottleneck for a widespread usage of these modern technologies, despite their tremendous potential. There is a great demand for tools that facilitate and further automate the accurate analysis of the images. Many researchers are actively working on providing the doctors with a single button that would launch a fully automatic image analysis and generate relevant parameters reflecting the function or the pathology of a particular organ, within a couple of minutes.
Experience has shown that many classical, low-level (data-driven) image processing and segmentation techniques are not sufficiently robust to localize organs and to detect contours under all routine clinical circumstances. In cases where image information is missing, or corrupted with noise, prior knowledge about organ shape and appearance is of critical importance. Many model-based image segmentation techniques have been proposed as an alternative to data-driven techniques. However, if the shape and /or appearance of the organ to be segmented do not fall in the vicinity of the model, the segmentation fails even if the image quality is perfect. Efforts are now aimed at the development of hybrid techniques combining model and data information.
During the last year of a former STW project (LGN.4503) entitled " Application of fuzzy neural networks to medical image segmentation"; Dr Faiza Admiraal Behloul developed a totally new and original approach for combining global and local information in medical image processing. She implemented a virtual mobile robot and trained it using fuzzy neural networks to navigate around the Left Ventricle (LV) of the heart in multi-slice MR images. On its journey around the heart, the robot could recognise its relative location (septum or lateral wall) while delineating fully automatically the borders of the myocardium. This may sound unconventional but it was effective enough to be integrated in a commercially available analytical tool. The aim of this project is to further develop this very young technique by integrating state-of-the-art localisation and environment-modelisation techniques and explore its full potential for automatic organ delineation in high-resolution medical images. The robot will also be trained to detect novelty in an organ. Novelty detection systems detect inputs that do not conform to an acquired model of "normality". The robot will automatically highlight areas presenting significant deviations from a model.

Utilization
In this project we will apply our new concept of virtual exploring mobile robot to: (i) the automatic contour detection of the heart in MR and CT images, and (ii) the automatic fiber tracking in diffusion tensor MR images of the brain.
Our industrial partner MEDIS medical imaging systems, has already acknowledged the potential of our new approach. Indeed, in close co-operation with MEDIS, we evaluated and validated a new module for LV contour detection and integrated it into a commercially available analytical software (MASS-CT). However, so far only a 2D version of our approach has been developed and integrated. The outcome of this grant proposal will be a totally new module for the myocardium contour detection, whereby the robot is exploring the entire 3D image based on a more detailed map (model) of the heart. This will improve significantly the performance of the MASS-CT software when the image quality is low in some slices and particularly around the apex of the heart. Furthermore, the new module will allow the segmentation of the complex shape of the right ventricle as well. The new implementation of the mobile robot is aimed to perform equally well in MR and CT images of the heart, with no additional tuning needed. If it can be demonstrated that the new approach outperforms the existing algorithms, the same module might be integrated to MASS-MRI as well. This will reduce the quantity of code that MEDIS has to maintain for both versions of MASS. MEDIS has contacts with the major OEM MRI and CT companies to discuss sublicensing the software, if they would be interested.
MEDIS does not have yet experience in the field of neuro-image processing, but clearly demonstrated interest in the work of the neuro-image processing section of LKEB. The LKEB and MEDIS are currently investigating the market for a quantitative neuro-image processing software package. The neuro-image processing section has already developed the first version of a software package called SNIPER, to support a large-scale clinical research project (see paragraph 4.4 for a short description). SNIPER will also serve as a prototype for a new commercial product. The outcome of the neuro-application in this grant proposal will be integrated into SNIPER and distributed to our clinical end-users for their daily research practice and to the network of end-users that we are actively developing.