detection of prostate cancer using ... - shekoofeh azizi
TRANSCRIPT
Detection of Prostate Cancer Using Temporal Sequences of Ultrasound Data:
A Large Clinical Feasibility Study
S. Azizi1, F. Imani1, S. Ghavidel2, A. Tahmasebi3, J. T. Kwak4, S. Xu4, B. Turkbey4, P. Choyke4, P. Pinto4, B. Wood4, P. Mousavi2, and P. Abolmaesumi1
1The University of British Columbia, Vancouver, BC, Canada
2School of Computing, Queen's University, Kingston, ON, Canada 3Philips Research North America, Cambridge, MA, United States
4National Institutes of Health, Bethesda, MD, United States
1
Prostate Cancer (PCa)
1 in 7 men will be diagnosed with PCa in his lifetime.
Accurate detection of high grade cancer is essential for clinical management.
2
Prostate Specific Antigen (PSA)
Digital Rectal Examination (DRE)
Core Needle Biopsy
Images from book “Fast Facts: Prostate Cancer ” (7th edition) by Kirby, Roger S. Azizi et al.
Ultrasound-based Tissue Typing
4
Doppler imaging [Potdevin’01][Tang’03][Goosen’03][Nelson’07][Xie’13]
Results are variable across centres.
Elastography [Ophir’00][Pallewin’07][Miyagava’09][Correas’11]
Not all cancer are stiff and all stiff lesions are not cancerous.
Spectral and texture-based [Scheipers’03][Lizzi’03][Feleppa’07,09,15]
Limited success based on the analysis of a single TRUS image.
Images from [Correas’11,13] S. Azizi et al.
Data Acquisition
We used Philips UroNav during MRI-guided targeted TRUS biopsies.
Data was acquired at the National Institutes of Health (NIH), Maryland.
6
MR-Ultrasound fusion (UroNav)
Taking biopsy & histological processing
Temporal Enhanced US data
GS 3+2 GS 3+3 GS 4+3 GS 4+4 GS 5+4 MR Target
S. Azizi et al.
Data Division
7
Distance to segmented prostate boundary
Agreement between Pathologies
Axial: GS 3+3
Sagittal: GS 4+4
(Ground Truth)
d = 11 mm d = 7 mm d = 3 mm d = 1 mm UroNav MR/US fusion system
registration accuracy: 2.4 ± 1.2 mm
Distance to boundary ≥ 3 mm
Target
S. Azizi et al.
Data Division
8
Distance to boundary more than 3 mm
Axial and sagittal pathologies agreement
D2-A 156 115
D2-C 168 115
D1 32 27
255 Cores
158 Patients
Trai
nin
g Te
st D
atas
et
D2-B 117 91 D2-B
D2-A
D2-C
S. Azizi et al.
Method: Automatic Feature Learning
9
Training Dataset RF Time Series 50 Spectral Features
Histopathology Results Binary Labels
20 ROIs
Target
Deep Belief Network (DBN)
Visible layer Hidden Layers
S. Azizi et al.
Method: Training Classifier
10
Training Dataset RF Time Series 50 Spectral Features
Histopathology Results Binary Labels
20 ROIs
Target
Deep Belief Network (DBN)
Learned Features
Support Vector Machine (SVM) Classifier
Benign Cancerous
Pc = 0 Pc = 1.0
S. Azizi et al.
Method: Test on a new subject
11
Test Data RF Time Series 50 Spectral Features
Trained DBN
Learned Features
Trained SVM
Cancer Likelihood Map
S. Azizi et al.
Results: MR Aggressiveness Level
12
AUC = 0.89 at 5 mm distance for moderate MR suspicion level.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
All MR levels Moderate MR High MR
Are
a U
nd
er
the
RO
C C
urv
e (
AU
C)
D2-C: Bi-plane Matching PathologyD2-A: d > 3 mmD2-B: A Ʌ C
S. Azizi et al.
Results: Tumor Size
AUC of 0.77 for cores with MR-tumor-size smaller than 1.5 cm
AUC of 0.93 for cores with MR-tumor-size larger than 2 cm.
13 S. Azizi et al.
Results: Gleason Scores
14
0
2
4
6
8
10
12
14
16
18
GS 3+3 GS 3+4 GS 4+3 GS 4+4 GS 4+5
Nu
mb
er o
f C
ore
s
Pathology
Correctly Identified
Misclassified
S. Azizi et al.
Test Dataset: Benign Cores
RF Time Series
50 Spectral Features
Test Dataset: Cancerous Cores
RF Time Series
50 Spectral Features
Results: Feature Visualization
15
Trained DBN
Trained DBN
Back Propagation
Back Propagation
Absolute Difference Differentiating
Frequencies
S. Azizi et al.
Results: Feature Visualization
16
First neuron in the third hidden layer Sixth neuron in the third hidden layer
Top two most differentiating features:
S. Azizi et al.
Colormaps
18
Low MR Level, Benign Target High MR Level, Cancerous Target
Moderate MR Level, Benign Target Moderate MR Level, Cancerous Target
Target Target
Target
Target
S. Azizi et al.
Summary and Future Works
Using temporal enhanced US in a fusion prostate biopsy study an AUC of 0.80 for moderately scored mp-MRI targets is achieved.
Temporal enhanced US combined with mp-MRI has the potential to reduce the number of unnecessary biopsies.
19
0.7
0.72
0.74
0.76
0.78
0.8
0.82
0.84
All MR levels Moderate MR High MR
Area Under the Curve (AUC) for different MR suspicion levels
MR Target
S. Azizi et al.
Summary and Future Works
Tissue-dependent features of the prostate can be automatically extracted and classified using deep learning techniques.
20
MR Target
S. Azizi, et al., “Classifying Cancer Grades Using Temporal Ultrasound for Transrectal Prostate Biopsy,” MICCAI 2016.
S. Azizi et al.