UFR de Physique

Propositions de stages en laboratoire -- M2

Les offres sont actualisées en mai. Par exemple, les offres de stages pour l'année universitaire 2015-2016 seront mises en place en mai 2015, les offres de stages pour l'année universitaire 2016-2017 seront mises en place pour en mai 2016, etc.

Paris Sud- Application of texture metrics to magnetic resonance imaging of cancer: methodological developments

  • Option Finalisée « Physicien des hopitaux » du parcours Physique Biologique et Médicale
  • Laboratoire: Autre (Autre)
  • Responsable du stage: Philippe Garteiser (philippe.garteiser@inserm.fr, 06 78 71 41 03)
  • Co-responsable(s): Irene Buvat (irene.buvat@u-psud.fr)
  • Mots clés: quantitative MRI, texture, validation, spatial resolution, signal to noise ratio

Tumor heterogeneity is an important feature favoring tumor growth and resistance. Imaging methods such as MRI are well adapted to study tumor heterogeneity because they provide a large variety of spatially distributed quantitative measures and are readily applicable in the clinic. However, current MR image analysis of tumors is often limited to size metrics or coarse quantitative measurements lumped across the entire tumor. Since several years, texture metrics have been applied to ultrasound imaging (US), computed tomography (CT), positron emission tomography (PET) and MR(1). Here we focus on texture analysis in MRI. In the project, the successful candidate will be in charge of - developing MRI texture metrics adapted to the specificities of MRI-derived images and parametric maps, namely substantial instrumental noise and variety of contrasts - developing multicontrast MR image acquisition methods tailored to the quantification of tumor heterogeneity with robust texture analysis in a small animal model of tumor growth and response to treatment Scientific rationale: Texture metrics assess how frequently specific patterns of pixel signal intensities can be found in an image (2). These patterns are defined as spatial and intensity relationships between any two pixels that can be analyzed with so-called grey-level co-occurrence matrices or run-length matrices. These matrices are analyzed to derive salient features such as contrast, homogeneity, dissimilarity, entropy or energy(3). Texture metrics are beginning to see acceptance in the field of cancer (4-8). They correlate with histopathological features (9, 10), tissue-specific uptake pattern (11), tumor clonal subtypes (12-14) and pathological prognostic factors (15). They can outperform visual assessment in terms of prognostic value (16). Even in computer tomography, where signal to noise ratios (SNRs) are high, instrumental noise directly affects texture metrics (17). Little is known on the proportion of measurement variability caused by methodological bias, and standardization is called for (18-20). Moreover, in abdominal cancer MRI, application of texture metrics is not commonplace despite some promising recent endeavors (21-23). Project proposal: The project will consist in developing texture measurement algorithms tailored to MR images. Although the project will primarily focus on diffusion MRI, T1, T2, T2* and perfusion images will also be analyzed. Axis 1: Effect of SNR and voxel size on MR texture. In preliminary studies, it was established that image voxel size, signal to noise ratio and texture characteristics were strongly coupled. In particular, in the low SNR regime, texture metrics can be dominated by instrumental noise. To decipher the effects of acquisition voxel size and signal to noise ratio on the resulting texture metrics, the candidate will have at its disposal a fully coregistered dataset of tumor diffusion weighted images acquired at spatial resolutions of 300µm, 400µm, 500µm and 600µm. Further investigations will also be carried out in silico by spatially filtering images (alteration in the effective point spread function resolution without changing the underlying signal to noise ratio) or by downsampling them (alterations in the resolution with constant instrumental signal to noise ratio). Axis 2: Understanding the ROI size confounds Furthermore, an important redundancy of information was observed between many texture metrics and the size of the analyzed region, an otherwise fundamental piece of information. In PET imaging, this very strong correlation can be mitigated by performing an absolute normalization of the data rather than a relative one. The effect of MR image normalization on the ROI size - texture coupling will hence be analyzed in detail. Axis 3: Voxel anisotropy, 2d and 3d spatial relationships. A last methodological axis will consist in evaluating the effect of high voxel anisotropy on resulting texture metrics. Indeed in MRI, voxels are often acquired with fine in-plane resolution, but with relatively large slice thicknesses (anisotropy ratios greater than 5X are commonplace). Hence, the typical 3d processing of texture metrics is susceptible to spatial orientation bias. We propose to interpolate the datasets into pixels of isotropic sizes to investigate the potential effects on resulting texture metrics. By nature, the interpolation process filters the data, potentially causing an artificial modification of the computed textures. Hence the validity of this approach will be compared to other techniques such as "mosaification", where only 2 dimensional spatial relationships are conserved by computing the texture matrices on a 2d mosaic of the various slices available through the volume of interest. Although immune to resolution anisotropy bias, this process trades off a bias for another, namely the spatial orientation bias arising from the use of only two dimensional spatial relationships in the inplane directions. Time frame: The paid internship project will be articulated in two phases; a first phase where the candidate will use an existing dataset of mouse tumors to develop texture quantification methods. In the second phase, the candidate will exploit the findings from phase 1 to alter the small animal MRI acquisition protocol. Candidate profile: The successful candidate will hold a degree in any of the following fields: computer programming, imaging physics, biomedical physics, applied mathematics. Programming proficiency is required, especially in C , matlab and integrated development environments.