Graduated from the prestigious french engineering school Ecole Centrale de Paris, I have also carried on signal processing studies (Elektrotechnik Fachrichtung Techniche Informatik) in Germany during two years thanks to an exchange program. I have thus also graduated from the RWTH Aachen University. Both diploms are equivalent to master degrees.
I ended up my education with a six month research project (the Diplomarbeit) in the
Institute for Image Processing LfB, which is specialized in medical image processing. I had to define at which optimal resolution
human tissue samples had to be imaged to allow for an efficient automatic cancer cell detection.
I carried out this work in two times: I have developed an original interpolation method simulating the behavior of the microscope, and I have imagined an automatic cancer cell detection sheme to apply in the reconstructed images. This last point composes one of the steps of the detection sheme presented by the lab at ICIP'05.
Eager to stay in the medical imaging environment, I have worked during one year as an engineer in the Image Quality team by General Electric Healthcare in Paris. I decided then to apply to the phD that GE Healthcare opened in partnership with the IRISA about noise reduction in fluoroscopic images by filtering in the direction of the motion. My tutor by General Electric is Jean Liénard, and I am monitored in the VISTA project of the IRISA by Patrick Bouthemy.
Radioscopy and digital radiography offer the clinician two main modalities: diagnostic and interventional. If diagnostic exams are
generaly fairly short and thus allow to use large X-Ray radiations leading to contrasted images with little noise, interventional
exams (so-called fluoroscopic exams) must be carried out under limited radiation to protect the patient. As a result,
the exam images are corrupted by a high level of quantic noise that must be digitally processed to be clinically acceptable.
An efficient way of addressing the noise issue is to average successive measures of the same pixel. Such a temporal denoising sheme has the advantage over spatial filters not to introduce noise coloration nor image look change.
Its main drawback is its unability to address scene motions: an adaptative temporal filter would decide at each pixel which averaging strength offers the best compromise between noise reduction and information conservation. This discussion leads to conservative parameter tuning which limits greatly the denoising power of the filter.
My phD offers to explore ways out of this alternative by developing an algorithm which would be robust to the motions. I thus need to track the anatomy in the fluoroscopic image to be able to average aggressively the pixels even when there have been motions in the sequence. This would allow to improve dramatically the displayed images, and this by keeping all the information the clinician needs.
The main originality of this topic of motion compensation is to focus on very specific images. They are not only difficult (because
very noisy) but also transparent. The physics of image formation with X-Ray namely implies that when an object moves
over an other one, there is no occlusion but a grayscale addition.
As a result, the image modelization is different from the one of classical video images, and we have to develop specific frameworks to solve our problem. In fact, I must answer two questions, which build the two main parts of my phD:
The research topic presentation is developed on the next page , and our contributions are briefly introduced on a specific page.
Key words: object tracking, transparent images, multiple motions, denoising.
phD topic: Fluoroscopic noise reduction by filtering in the motions' direction.