New Tools for MRI processing: Segmentation, Denoising, Super-resolution (Pierrick Coupe's Open Source Contributions)
Computer Vision Central - Posted on December 25, 2011 at 1:08 pm.
Pierrick Coupe's Open Source Contributions: Based on Maltab C-Mex code, the proposed tools derives from my recent research on nonlocal means estimator.
In 3D Magnetic Resonance Imaging (MRI), image resolution is limited by several factors such as hardware, time constraints or patient’s comfort. Therefore, the low resolution of MRI can limit accuracy of post-processing tasks such as lesion segmentation (e.g., brain tumor or multiple sclerosis). In order to efficiently reconstruct high resolution MRI from low resolution MRI a new patch-based method has been proposed to recover high frequency information by using a data-adaptive reconstruction in combination with a sub-sampling coherence constraint. The proposed method has been applied in mono and multi-modal context.
Reference: P. Coupé, J. V. Manjon, E. Gedamu, D. Arnold, M. Robles, D. L. Collins. Robust Rician Noise Estimation for MR Images. Medical Image Analysis, 14(4) : 483–493, 2010.
Denoising is a crucial step used to increase image quality and to improve performance of all the tasks needed for quantitative imaging analysis. The main challenge is to remove noise component while keeping the integrity of relevant image information. This later aspect is important especially in context of medical image processing. To address this issue, the well-known Non-local Means filter (NLM) has been adapted to medical image analysis. First, the NLM filter has been extended to 3D Magnetic Resonance images (MRI). MRI specificities have been taken into account such as Rician noise, non-stationary noise and image dimensionality. Second, a Bayesian definition of the NLM filter has been proposed in order to adapt the NLM filter to ultrasound imaging. Finally, a new collaborative framework has been proposed to multi-photon image filtering.
Quantitative magnetic resonance analysis often requires accurate, robust, and reliable automatic extraction of brain structures. Volume and shape of these anatomical structures are useful for pathology detection and population comparison. However, the high inter-subject variability and the modifications caused by pathology make automatic segmentation very challenging. A novel patch-based method using expert manual segmentations as priors has been proposed to achieve this task. Inspired by recent works in image denoising and label fusion segmentation, this new method has been adapted to segmentation of complex structures such as hippocampus and to brain extraction.