Knee cartilage segmentation using active shape models and contrast
enhancement from magnetic resonance images
Germán González
a
and Boris Escalante-Ramírez
b
a
Posgrado en Ingeniería Eléctrica, Universidad Nacional Autónoma de México, Mexico City,
Mexico.
b
Departamento de Procesamiento de Señales, Facultad de Ingeniería, Universidad Nacional
Autónoma de México, Mexico City, Mexico.
ABSTRACT
In this paper, we take advantage from contrast characteristics of our magnetic resonance images improving the
performance of Active Shape Models (ASM) applied on knee cartilage segmentation. We perform an image fusion-
based contrast enhancement method using time series MRI T2. Then, we apply ASM algorithm and we compare results
with ASM without contrast enhancement. The results show that the ASM with contrast enhancement performs better and
is consistent. We validate these results using Dice coefficient and Hausdorff distance.
Keywords: Segmentation, Active Shape Models, Contrast Enhancement, Gaussian and Laplacian Pyramids, Image
Fusion.
1. INTRODUCTION
Knee Osteoarthritis (OA) [1] is a disease caused by biomechanical stress that affects the articular cartilage and bones of
the knee. This condition causes pain and malfunction. OA may be present in any of the medial femoral compartments,
either the tibiofemoral or patelofemoral, according to the location of damaged cartilage. Diagnosis of Knee OA is done
since the first clinical examination. Pain, morning stiffness and knee swelling in a patient older than 50 years are
considered as consequences of OA. However, image analysis of the knee also has an important role since it can confirm
the OA diagnosis, determines compartments involved and evaluates the disease stage. Moreover, it can confirm the
responsibility of OA in the symptoms and provides information about disease evolution during treatment. Magnetic
Resonance Imaging (MRI) [2] provides a non-invasive assessment for evaluating the presence and progression of the
Knee OA [3]. Magnetic resonance can show the soft tissue structures and their boundaries with the bones, without
significant distortion. Furthermore, MRI does not change the tissue’s dimension and there is not superposition between
anatomical structures and, more important, it directly visualizes the knee cartilage and its defects [4].
Automatic or semiautomatic knee segmentation has been studied for more than 20 years and many approaches have been
reported, for example: growing regions [5-7], Bezier splines [8], active contours [9, 10], Bayesian classifiers [11] and
active shape models (ASM) [12, 13]. ASM [14] compute an average shape of the object to segment from a training set
and a statistical model of minimal parameters that allow the shape to adjust to different objects within a certain range.
ASM have been widely used in medical image analysis because there is sufficient knowledge about the shape of targeted
anatomical objects obtained through diverse medical imaging modalities.
The purpose of our work is to take advantage of the contrast characteristics of Resonance Magnetic Images (volume
acquisitions in time series) in order to enhance the contrast in the images and therefore, improve the ASM segmentation
performance in knee cartilage. We test the contrast enhancement method on real data; this is new because, in the original
paper [15], only synthetic images are presented.
2. METHODS AND MATERIALS
In this section, the main characteristics of MRI, segmentation and contrast enhancement algorithms are described.
Validation metrics of segmentation and contrast enhancement are also discussed.
IX International Seminar on Medical Information Processing and Analysis, edited by
Jorge Brieva, Boris Escalante-Ramírez, Proc. of SPIE Vol. 8922, 892213 · © 2013
SPIE · CCC code: 0277-786X/13/$18 · doi: 10.1117/12.2035529
Proc. of SPIE Vol. 8922 892213-1