MIAD 2012 Abstracts


Full Papers
Paper Nr: 3
Title:

AUTOMATIC FINE-GRAINED LABELING OF BRAIN MR IMAGES - A CRF Approach

Authors:

Ammar Mechouche, Azeddine Zidouni, Mohamed Quafafou and Omar Boucelma

Abstract: We propose in this paper a method for automatic refining of labels associated with regions delimited on brain Magnetic Resonance Images called patches. The patches coordinates and their topological relations represent the input of our system. Based on this input, the proposed method uses Conditional Random Fields (CRFs) for learning correspondences between patches’ labels obtained from a brain atlas and refined labels defined in the Foundational Model of Anatomy ontology. A cross-validation on a small collection of brain MR data was performed, and the results obtained so far are encouraging.

Paper Nr: 9
Title:

STATISTICAL ASYMMETRY-BASED BRAIN TUMOR SEGMENTATION FROM 3D MR IMAGES

Authors:

Chen-Ping Yu, Guilherme C. S. Ruppert, Dan T. D. Nguyen, Alexandre X. Falcão and Yanxi Liu

Abstract: The precise segmentation of brain tumors from MR images is necessary for surgical planning. However, it is a tedious task for the medical professionals to process manually. The performance of supervised machine learning techniques for automatic tumor segmentation is time consuming and very dependent on the type of the training samples. Brain tumors are statistically asymmetrical blobs with respect to the mid-sagittal plane (MSP) in the brain and we present an asymmetry-based, novel, fast, fully-automatic and unsupervised framework for 3D brain tumor segmentation from MR images. Our approach detects asymmetrical intensity deviation of brain tissues in 4 stages: (1) automatic MSP extraction, (2) asymmetrical slice extraction for an estimated tumor location, (3) region of interest localization, and (4) 3D tumor volume delineation using a watershed method. The method has been validated on 17 clinical MR volumes with a 71.23%+-27.68% mean Jaccard Coefficient.

Paper Nr: 12
Title:

COMPARATIVE ANALYSIS ON METRICS AND FILTERS TO REDUCE IMPULSIVE NOISE IN MEDICAL IMAGES USING GPU

Authors:

M. Guadalupe Sánchez, Vicente Vidal and Jordi Bataller

Abstract: In many current applications of image processing, eliminating the noise is an important task in the pre-processing phase. In medicine, medical imaging obtained by X-ray and computed tomography, for example, mammograms, can have different types of noise, making it difficult to visually and to detect microcalcifications. We have adapted a noise reduction method for color images that gives good results for grayscale images. In the first step of the method, the corrupted pixels are detected using the concept of peer group with a metric and then is corrected by some kind of filter. This paper presents an algorithm with a very good balance between quality and computational cost to removing impulsive noise in mammography images. With regard to quality, we compared three metrics (two Fuzzy and one Euclidean) and two filters (Arithmetic Mean and Median). To reduce the computational cost, the method is parallelized on a Graphic Processing Unit. The quality results show that the metrics studied yield similar results, being the Euclidean metric less expensive computationally. On the other hand, the filter must be chosen depending on the density of noise in the input image.

Short Papers
Paper Nr: 5
Title:

AUTOMATIC DETECTION OF HARD EXUDATES AND OPTIC DISC IN DIGITAL FUNDUS IMAGES

Authors:

Elizabeth Chavez-Hernandez and M. Elena Martinez-Perez

Abstract: Automatic detection of characteristic patterns of diabetic retinopathy such as hard exudates may help to an early diagnosis. Methods for automatic detection of hard exudates and optic disc are presented. Exudates detection involves a preprocessing stage, threshold selection and region growing. For optic disc detection a Bayes classifier is applied followed by mathematical morphology techniques in order to improve the final result. The methods here presented were evaluated using the IMAGERET database, which contains fundus images evaluated by qualified experts. In average, the area of exudates automatically detected overlaped with 60.75% and 63.91% areas defined by each of the two experts. For optic disc detection, sensitivity and specificity were 72.12% and 95.56% respectively.

Paper Nr: 7
Title:

MANDIBLE PARAMETERIZATION USING A REFERENCE LINE

Authors:

M. J. Tsai, C. L. Chen, H. W. Lee, C. H. Chao and P. W. Hsu

Abstract: Using fibula flap for mandible reconstruction becomes a common method. Pre-surgical planning always assumes a known removal portion of mandible to plan the cutting of fibula. Researchers have developed jig or mechanism to assist doctors in measuring and guiding cutting locations of the mandible. However, the true locations of mandible removal are never known. To ease the problem, standard mandible parameters is required to locate the removal area. This paper aims to construct a reference line for parameterizing the mandible. The reference line can be divided into four portions that are coincident with the anatomical features areas of the mandible: Symphysis, Body, Angle, and Ramus. Each portion is then re-sampled into 10-15 sections, and these sections faithfully represent the mandible shape. Thus, all mandibles are re-defined by these 101 sections. Using these mandible parameters, the removal area is recorded by numerical addressable locations to compute the fibula cut, and the mandible and fibula can be aligned according to the parameters for planning the reconstruction.

Paper Nr: 11
Title:

A SOFTWARE PLATFORM TO ANALYZE MR IMAGES BASED ON 3D FRACTAL DIMENSION - Application in Neurodegenerative Diseases

Authors:

J. Jiménez, A. M. López, F. J. Esteban, P. Villoslada, J. Navas and J. Ruiz de Miras

Abstract: Previous studies carried out by our group have demonstrated that 3D fractal dimension algorithms detect changes in apparently normal magnetic resonance (MR) images of the brain in patients suffering early stages of Multiple Sclerosis. In addition, 3D fractal dimension has also been demonstrated to be useful for detecting brain abnormalities in other cerebral diseases, as in Alzheimer’s disease and in children born after intrauterine growth restriction. Thus, 3D fractal dimension detection has been proposed as a valuable and powerful diagnostic tool. To our knowledge, no user-friendly software is available to obtain the 3D fractal dimension of volumetric MR images. In this paper, we present an optimized Web platform that allows computing the 3D fractal dimension value for uploaded MR images in an interactive user-friendly way. Moreover, and because the computational cost of the involved algorithms is very high for interactive use, we have focused our efforts on the optimization of the appropriate algorithms using the parallel computing power of current GPUs and multi-core CPUs.

Posters
Paper Nr: 14
Title:

A REGISTRATION FRAMEWORK FOR EVALUATION OF T1, T2 AND DWI SIGNAL INTENSITIES IN MULTIPLE MYELOMA

Authors:

Eros Montin, Paolo Potepan and Luca Mainardi

Abstract: Objective. In this study we point out the Diffusion-Weighted Imaging (DWI) role in the diagnosis of multiple myeloma (MM), comparing its signal values (SV) (Sommer et al., 2010) with the standard imaging modalities T1, T2. We further evaluate how SV change in relation with the percentage of plasma cells infiltration evaluated through bone marrow biopsy (BMB). Methods. Since March 2008 23 patients with average age of 61 (+/- 11) years old, 11 females and 12 males, have been investigated before their own therapy with a whole body MRI protocol, concerning a whole body T1, a whole body T2 and a whole body DWI and a BMB. An experienced radiologist defined for each patient two volume of interests (VOIs): onto the main lesions and on healthy bones (Femur and Homerus). After that, we have subdivided the full population by a clinical threshold of 25% on cells infiltration percentage; then, we analysed statistical differences in the 2 groups (A, B). Results. We found out that DWI voxels intensities in group A (infiltration <= 25%) were higher than group B, this gap had to be considered statistically different (P <= 0:05).