Skip to Main content Skip to Navigation

Contribution à la détection de changements dans des séquences IRM 3D multimodales

Abstract : Medical imaging has deeply influenced both medical practice and medical research. It is now an essential aspect of diagnosis and clinical followup. The constant improvement of spatial resolution provides an increasing volume of data in different modalities (Magnetic Resonance imaging - MRI, X-ray scanner, nuclear medicine, ultrasound imaging). Manual interpretation of huge amounts of data by an expert is tedious and error prone. Image processing automatizes some of these tasks and helps the expert in quantitative and qualitative analysis of images. This thesis describes image processing methods for automatic change detection in serial brain MRI. We are interested in small localized intensity changes that appear in pathological evolutions such as lesion evolutions in Multiple Sclerosis (MS). The image processing methods developed here have many medical applications : computer assisted diagnosis, long term monitoring of a pathology, evaluating therapy and drug efficiency, assisting decisions for surgery. This work was done in close collaboration between the LSIIT (ULP/UMR CNRS 7005) and the "Institut de Physique Biologique" (ULP-Hôpitaux Universitaires / UMR CNRS 7004), within the multi-laboratory project "Imagerie et Robotique Médicale et Chirurgicale" (EPML-IRMC). This work was supported by the "Ligue Française Contre la Sclérose En Plaques" (LFSEP), SERONO and the Alsace region. Automatically and reliably detecting interimage changes is a difficult task. Direct comparison of successively acquired images is generally not possible. Patient position is never identical. Acquisition parameters may drift between scans leading, for example, to image intensity changes. The exact definition of what should be detected may be subjective. Within the specific framework of lesion intensity change detection, a variety of complex global or local deformations of anatomical structures may also make it difficult to directly compare image intensities. This work is centered on image processing methods leading to the decision on whether changes are statistically significant or not. When an expert visually checks images for changes, he uses implicit, high-level prior knowledge for correcting certain artifacts. Thus, he may visually compensate for repositioning errors and use his anatomical knowledge to identify and reject certain artifacts. We develop, here, automatic methods for identifying and correcting most important artifacts (positioning, deformation, intensity changes ...). We also develop an original method for cortical segmentation that uses anatomical informations. The segmented cortex is used to enhances detection results. The image processing methods were developed for brain MRI processing. However, they are general enough to be used in many other fields. The performance our change detection system was evaluated on a large database of 3D multimodal MR images (over 200 FLAIR, RARE and GE3D images) of patients suffering from relapsing remitting Multiple Sclerosis (MS). The method was assessed using receiver operating characteristics (ROC) analysis, and validated in a protocol involving two neurologists. The automatic system outperforms the human expert. In the first part of this thesis, we present MR imaging as well as medical aspects that are important for understanding the rest of this work. We describe MRI modalities and associated acquisition artifacts that may need to be corrected. We then give a broad description of anatomical elements of the brain and their appearance in different MRI modalities. We finish by describing brain pathologies, their evolutions, and their MRI appearance. The goals and the limitations of our change detection approach are given with respect to these applications. In the second part of this thesis we describe a new sub-voxel segmentation approach. When deciding on the validity of an observed change, an expert uses anatomical knowledge. In our automatic change detection system, part of this knowledge comes from the segmented brain image. We describe a 3D brain MRI segmentation method in which a high resolution label image evolves under the influence of multiple constraints. The constraints are expressed in a versatile energy minimization framework which allows for evolutions only at label boundaries, effectively making it a surface evolution system. Constraints are defined using atlas-mapped parameters. The atlas, composed of a reference image and parameter values, is mapped onto the source image using a multi-resolution deformable image matching method. Variable scale image constraints are considered. The prior model currently includes: a relative distribution constraint, which gives the probability of observing a label at a given distance from another label, a thickness constraint and a surface regularization constraint. Issues related to partial volumes are addressed by the use of a high resolution label image and an accurate model of the acquisition process. High resolution segmentations are thus obtained from standard (eventually low resolution) MRIs. The performance of our segmentation system was evaluated on simulated images. The system was also successfully applied on a large number (75) of real images. The third part of this thesis describes the full image processing chain that implements the change detection system. It also describes the evaluation protocol and its results. The first step of the processing chain consists in correcting repositioning errors and deformations. All images in the image database are aligned on carefully chosen reference images. This is first done using an affine, robust, iterative registration approach. Registration is then fine-tuned by using deformable registration. The second step consists in correcting nonlinear intensity changes as well as other residual errors. Our image intensity correction approach computes a nonlinear intensity transfer function by optimizing a simple criterion that uses information provided by a joint histogram. Finally, in the last step, a multimodal statistical change detection approach decides which changes are statistically significant. Anatomical information from the segmentation is used to reject certain erroneous detections. The complete processing chain was applied fully automatically on a database of over 200 images of different modalities, thus showing its reliability. The system was validated in an evaluation protocol involving two experts (neurologists). The first expert and the automatic system independently detected interimage changes (the first expert did this manually). The second expert then compared results obtained by both detections (manual, and automatic) and refereed on disagreements. Receiver operating characteristics (ROC) analysis provides a synthetic view of detection performance. It gives detection probability as a function of the number of false alarms. For practical applications, changes detected by the automatic system are sorted by decreasing likelihood. Sorted changes are presented to neurologists in an interactive visualisation program. Neurologists may therefore keep the final decision on detected changes and efficiently browse through the sorted changes. In the appendix we contribute thoughts on the importance of software development, and its free distribution, for image processing research. We then describe ImLib3D, a C++ library for 3D medical image processing research. It provides a carefully designed, object-oriented, standards conforming C++ library, as well as a separate visualization system. Focus has been put on simplicity for the researcher who is considered to be the end-user. Source code is freely available (Open Source) and has been placed in an open collaborative development environment (
Complete list of metadatas
Contributor : Marcel Bosc <>
Submitted on : Saturday, February 28, 2004 - 12:07:40 PM
Last modification on : Thursday, April 23, 2020 - 2:26:30 PM
Long-term archiving on: : Friday, April 2, 2010 - 7:26:56 PM


  • HAL Id : tel-00005163, version 1



Marcel Bosc. Contribution à la détection de changements dans des séquences IRM 3D multimodales. Neurosciences [q-bio.NC]. Université Louis Pasteur - Strasbourg I, 2003. Français. ⟨tel-00005163⟩



Record views


Files downloads