New Clinical Quality Level for Medical Image Volumes

Image processing of medical image volumes (3D/4D data) requires a completely new approach compared to standard images (2D). Research on methods for high speed processing as well as for high image quality output is required. In a research project within the VINST programme a novel approach for solving the speed challenge was developed. Building on this “pre-study” project, this new project covers the remaining research for reaching a new clinical quality output level, while maintaining the speed. The goal is to have the research results from both projects verified in a prototype. The project is a cooperation between the company ContextVision AB and its research partner Center for Medical Image and Visualization (CMIV) at Linköping University.

• Staff:

Klas Themner , PhD		R&D Manager		ContextVision AB
Hans Knutsson , Prof		Supervisor		Medical Informatics
Mats Andersson , PhD		Co-supervisor		Medical Informatics
Björn Svensson , MSc		PhD student		Medical Informatics
Johan Wiklund , MSc		Research Engineer		Medical Informatics
Martin Hedlund , MSc		Technical Director		ContextVision AB
Gunnar Farnebäck , PhD		Senior Scientist		ContextVision AB / Medical Informatics
Henrik Ejnarsson , MSc		Devolopment Engineer		ContextVision AB
Karin Adolfsson , MSc		Devolopment Engineer		ContextVision AB
Örjan Smedby , Prof, MD		Radiologist		Radiology
Anders Persson , MD		Radiologist		Radiology

• Former Staff:

• Project Description:

Motivations

In medical imaging, clinical routine involves processing of multi-dimensional datasets such as volumes and volume sequences. Still, most signal processing tools clinically available are however limited to 2-D. The additional statistic information provided by filtering using all dimensions of the dataset implies an increased signal-to-noise ratio (SNR), which can allow for a decrease in radiation dose for computer tomography or shorter acquisition time for magnetic resonance imaging, while maintaining image quality.

Project Description

A novel approach for high speed processing was developed within the VINST programme (see 3D-4D Medical Image Enhancement using Filter Networks for more details.). The aim of this project is to take these results from this project further to be verified in a prototype.


The goal of the clinical evaluation is to assess the utility of the image enhancement methods in clinical practice. Radiologists will review selected clinical cases and compare the quality with and without enhancement applied to the 3D datasets. Typical datasets to be used may be MR angiography and CT angiography or 3D MRI and CT images of the brain. The evaluation will include subjectively graded aspects of image quality such as noiselessness and visibility of minute structures. We also intend to design experiments where artificial pathology is introduced into natural images and the diagnostic performance of readers can be measured.

In addition to measuring the estimation accuracy on test patterns with known ground truth, the aim of this work package is to fairly measure the experienced subjective image quality of the entire process. To be able to capture the enhancement effect this requires the use of visualization software. This might seem straightforward, but the choice of visualization technique can highly affect the visual result. Determining a collection of visualization techniques for fair comparison of the results will require both clinical and technical expertise.




The filternet design procedure is non-trivial and today the user is required to manually tune several critical parameters. To facilitate the usefulness for a non-expert user a more generic strategy for application specific filter networks is required to reduce the number of parameters.




Image enhancement targets the contradictive objectives of preserving and/or enhancing image structures, while suppressing high-frequency noise. The methodology used is based on applying a set of filters, which form a local model adapted to every signal neighbourhood. The use of multiple scales, i.e. increasing the number of filters in the set, then allows a richer local model. Since visually important features must exist over several scales, an important research topic is to investigate how to exploit the richer model for better adaptation, i.e. improved ability both to suppress noise and to enhance important structures.