VOXEL CLASSIFICATION OF PERIPROSTHETIC TISSUES IN CLINICAL COMPUTED
TOMOGRAPHY OF LOOSENED HIP PROSTHESES
D.F. Malan1,2, C.P. Botha2,3, R.G.H.H. Nelissen1, E.R. Valstar1,4
Leiden University Medical Centre
Leiden, The Netherlands
ABSTRACT
We present an automated algorithm which classifies peripros-
thetic tissues in CT scans of patients with loosened hip pros-
theses. To our knowledge this is the first application of CT
voxel classification to periprosthetic tissues of the hip. We use
several image features including multi-scale image intensity,
multi-scale image gradient and distance metrics. Seven classi-
fier types were trained using five manually segmented clinical
CT datasets, and their classification performance compared to
manual segmentations using a leave-one-out scheme. Using
this technique we are able to correctly segment the majority
of each of the six tissue categories, in spite of low bone densi-
ties, metal-induced CT imaging artefacts and inter-patient and
inter-scan variation. Our automated classifier forms a prag-
matic first step towards eventual automatic tissue segmenta-
tion.
Index Terms— Automatic, classification, segmentation,
computed tomography, periprosthetic, osteolysis.
1. INTRODUCTION
The most significant complication that threatens the long-
term survival of a total hip arthroplasty (THA) is peripros-
thetic osteolysis [1, 2] which involves resorption of bone and
replacement by soft fibrotic tissue. Once osteolysis devel-
ops it usually progresses, eventually leading to mechanical
instability and prosthesis loosening.
Minimally invasive refixation of loosened prostheses is
possible [3] but requires the location and extent of fibrotic
lesions to be known pre-operatively. Recent studies have
shown that CT is more sensitive and accurate than tradi-
tional radiographs in detecting and measuring such lesions
[2, 4]. However, the steady increase in resolution offered by
modern CT scanners make traditional manual segmentation
extremely time-consuming, thereby limiting users’ utilization
of the available data.
1 Department of Orthopaedics
2 Medical Visualization Group, Department of Mediamatics, Delft Uni-
versity of Technology, The Netherlands
3 Department of Radiology
4 Department of Biomechanical Engineering, Faculty 3ME, Delft Uni-
versity of Technology, The Netherlands
CT of suffers from metal-induced artefacts [5] which
drastically complicate automatic segmentation near prosthe-
ses. To make things worse patients suffering from prosthetic
loosening often have very poor bone quality yielding low
CT image contrast with intensity values overlapping those
of other tissues. Statistical Shape Models are useful for
segmenting objects from low quality image data, but fare
badly when modelling pathological tissues (such as fibrotic
lesions) with no generalizable geometry [6] and/or consisting
of several small isolated regions.
Several papers have been published describing automatic
statistical pixel- or voxel segmentation of clinical data. By
combining several complementary image features, voxel clas-
sifiers deliver reasonable classification performance in spite
of metal-induced CT imaging artefacts, and without resort-
ing to explicit geometrical modelling or human intervention.
Radiographs [7], MRI [8] and CT [9] have been subjected
to pixel/voxel classification. Standard approaches generally
make use of multi-scale image intensity as well as higher or-
der spatial derivatives to describe local image variations and
“texture”. Image intensity variation between scans can com-
plicate X-ray and MRI feature selection, but CT scanners are
largely immune to this due to their well defined and calibrated
output measured in Hounsfield Units (HU). The geometric
position of the image pixels or voxels can be omitted [7] or in-
corporated [9] into the classifier’s feature space, although care
must be taken so that the chosen features remain invariant to
inter-scan orientation and scaling offsets.
The aim of this study was to develop an automated voxel
classifier that can serve as the first step in a segmentation
pipeline, eventually leading to patient-specific mechanical
modelling. We are interested in the 3D distribution of bone,
cement and fibrotic tissue around the prosthesis, which de-
fines the hip’s mechanical stability. In this paper we present
statistical voxel classifiers that classify periprosthetic tissues
into six possible tissue categories, namely cement, fibrotic
lesion, trabecular bone, cortical bone, intramedullary canal
and exterior. To our knowledge this is the first time that such
a 3D statistical voxel classifier has been applied to peripros-
thetic CT image data. The classifiers are trained on manually
segmented CT scans of five patients with clinically loose
prostheses, and evaluated in a (per patient) leave-one-out
1341978-1-4244-4126-6/10/$25.00 ©2010 IEEE ISBI 2010