Transcript Document

Improving brain tumor characterization on MRI by
probabilistic neural networks and non-linear
transformation of textural features
Georgiadis P.1, Cavouras D.2, Kalatzis I.2, Daskalakis A.1, Kagadis G.1, Sifaki
K.3, Malamas M.3, Nikiforidis G.1, Solomou A.4
1 Medical
2
Image Processing and Analysis (MIPA) Group, Laboratory of Medical Physics, School of
Medicine, University of Patras, Rio, GR-26503 Greece
e-mail: [email protected], web page: http://mipa.med.upatras.gr
Medical Signal and Image Processing Lab, Department of Medical Instrumentation Technology,
Technological Education Institution of Athens, Ag. Spyridonos Street, Aigaleo, GR-12210,
Athens, Greece
e-mail: [email protected], web page: http://medisp.bme.teiath.gr
3
4
251 General Hellenic Airforce Hospital, MRI Unit, Katehaki, Athens, Greece
Department of Radiology, School of Medicine, University of Patras, Rio, GR-26503 Greece
MedicExpo 2008, March 27-30, Athens
INTRODUCTION
BRAIN TUMOURS
• Approximately 39,550 people were newly diagnosed with primary benign
and primary malignant brain tumours in 2002. The incidence rate of
primary brain tumours, whether benign or malignant, is 14 per 100,000,
while median age at diagnosis is 57 years (CBTRUS 2002).
• Secondary or metastatic brain tumours, originate in tissues outside the
central nervous system and are a common complication of systemic
cancer. Brain metastases occur in 20% to 40% of all cancer patients. More
than 100,000 individuals per year will develop brain metastases (CBTRUS
2002).
MedicExpo 2008, March 27-30, Athens
INTRODUCTION
CURRENT TECHNIQUES
• Today, one of the most promising techniques for generating useful
information for brain tissue characterization is Magnetic Resonance
Imaging (MRI). In order to extract the diagnostic information of different
parameters reflected in MRI, image analysis techniques have been
employed.
• Brain tumour characterization is a process that requires a complicated
assessment of the various MR image features and is typically performed
by experienced radiologists. An expert radiologist performs this task with a
significant degree of precision and accuracy, despite the inherently
subjective nature of many of the decisions associated with this process.
Nevertheless, in the effort to deliver more effective treatment, clinicians are
continuously seeking for greater accuracy in the pathological
characterization of brain tissues from imaging investigations.
MedicExpo 2008, March 27-30, Athens
INTRODUCTION
AIM
The aim of the present study was to design, implement, and evaluate a
pattern recognition system, which, by analyzing routinely taken T1 postcontrast MR images, would improve brain tumor classification accuracy.
Employing a two-level hierarchical decision tree, distinction between
metastatic and primary brain tumors and between gliomas and
meningiomas were performed at the 1st and 2nd level of the decision tree
respectively.
MedicExpo 2008, March 27-30, Athens
MATERIALS AND METHODS
CLINICAL MATERIAL, FEATURE EXTRACTION AND REDUCTION
• A total number of 67 T1-weighted gadolinium-enhanced MR images were
obtained from the Hellenic Airforce Hospital with verified intracranial tumours,
using a SIEMENS-Sonata 1.5 Tesla MR Unit
(21 metastasis, 19 meningiomas and 27 gliomas).
• Utilizing these images, the radiologist
specified Regions of Interest (ROIs)
that included tumour regions.
• From each ROI, a series of 36
features were extracted; 4 features
from the ROI’s histogram, 22 from the
co-occurrence matrices, and 10 from
the run-length matrices.
• To reduce feature dimensionality, the
non-parametric Wilcoxon test was
employed. Only features of high
discriminatory ability (p<0.001), were
selected to feed the classification
scheme.
MedicExpo 2008, March 27-30, Athens
MATERIALS AND METHODS
CLASSIFICATION SYSTEM - EVALUATION
• Least Squares Features Transformation – Probabilistic Neural Network
(LSFT-PNN) Classifier
(Third degree (cubic) LSFT-PNN at 1st level, second degree (quadratic)
LSFT-PNN at 2nd level)
• Best features combination was determined employing the robust but time
consuming exhaustive search method (up to 3 features).
• Classifier performance was evaluated employing the leave-one-out
method.
MedicExpo 2008, March 27-30, Athens
RESULTS AND DISCUSSION
CLASSIFICATION RESULTS (Metastatic VS Primary brain tumors)
Cubic LSFT-PNN
PNN
Primary Brain
Tumors
Secondary
Brain
Tumors
Accurac
y
(%)
Primary Brain
Tumors
43
3
93.48
Secondary Brain
Tumors
1
20
95.24
94.03
Primary Brain
Tumors
Secondary
Brain
Tumors
Accurac
y
(%)
Primary Brain
Tumors
40
6
86.96
Secondary Brain
Tumors
1
20
95.24
89.55
MedicExpo 2008, March 27-30, Athens
RESULTS AND DISCUSSION
CLASSIFICATION RESULTS (Gliomas VS Meningiomas)
Quadratic LSFT-PNN
PNN
Accurac
y
(%)
Gliomas
Meningio
mas
Gliomas
27
0
100
Meningiomas
0
19
100
100
Accurac
y
(%)
Gliomas
Meningio
mas
Gliomas
26
1
96.30
Meningiomas
0
19
100
97.83
MedicExpo 2008, March 27-30, Athens
RESULTS AND DISCUSSION
DISCUSSION #1
• The LSFT-PNN classification algorithm increased the overall accuracy in
correctly characterizing primary and metastatic brain tumours. This is
important, since the precision of such a decision may be crucial in patient
management, e.g. metastatic tumours require specific treatment protocols,
such as radiation therapy and chemotherapy, while primary tumours may
also require surgical intervention.
• The reason behind this accuracy increment may be attributed to the
increased class separability that the LSFT procedure provides, especially
when non-linear terms are introduced in the classifier’s discriminant
function.
MedicExpo 2008, March 27-30, Athens
RESULTS AND DISCUSSION
DISCUSSION #2
• The best features combination described the shape of the histogram
peak (kurtosis) and expressed the degree of the in-homogeneity (sum and
difference entropy) in the grey-tones of the ROIs. These textural
characteristics are related to textural parameters that physicians employ in
diagnosis and they are proportional to the textural imprint of brain tumours
in MRI.
MedicExpo 2008, March 27-30, Athens
CONCLUSION
CONCLUSION
The aim of the present study was to design, implement, and evaluate a
software pattern recognition system to improve classification accuracy
between primary and metastatic brain tumours on MRI. The system
improved classification accuracy compared to other studies. Thus the
system could be of assistance to physicians as a reliable second opinion
tool when analysing brain tumour MR images.
MedicExpo 2008, March 27-30, Athens