COMPUTER RECOGNITION OF BLADDER CANCER WITH OPTICAL COHERENCE TOMOGRAPHY AND TEXTURE ANALYSIS

1
COMPUTER RECOGNITION OF BLADDER CANCER WITH
OPTICAL COHERENCE TOMOGRAPHY AND TEXTURE
ANALYSIS

• Colleen A. Lingley-Papadopoulos, PhD
• Murray H. Loew, PhD
• Michael J. Manyak, MD
• Jason M. Zara, PhD
• The George Washington University, Washington, DC,
  USA

2
INTRODUCTION

• Bladder Cancer 2007
• 67,160 new cases 13,750 deaths
• Diagnosed by endoscopy and biopsy
• Some lesions and precancerous lesions
• not easily visible
• Multiple biopsies often required

3
Optical Coherence Tomography (OCT)

• Optical imaging
• Near-infrared light source
• Images sub-surface microscopic structures
• Resolution 10-20 mm
• Penetrates 1-2 mm
• Nearly all bladder cancer arises in 600 mm

4
Optical Coherence Tomography (OCT)

• High resolution real time image analysis
• requires training
• Possible guide to biopsy
• Need for accurate detection of suspicious areas
• at time of biopsy
• Texture analysis to differentiate
• biopsy areas

5
Normal
Superficial TCC
OCT BLADDER
Invasive TCC
6
87 SCANNED AREAS 16 papillary tumors 36
suspicious areas 35 suspected normal TP (21)
TN (24) FP (7) FN (0) Sensitivity
100 Specificity 77 PPV 75
NPV 100 Accuracy 92 PPV for
invasion 90 Manyak et al., J Endourol
2005
7
Materials

• Imalux Endoscopic System (NIRIS)
• Super luminescent diode light source
• Operates near 980 nm
• Images at 0.7 frames/second
• Endoscopic scanning probe
• 2.7 mm diameter, 5 m long
• Depth range of 2.2 mm in air
• Lateral scanning range of 1.6 2.4 mm
• Working distance from probe surface 0.5 mm
• Lateral Resolution of 50 µm in air
• Axial Resolution of 10-20 µm in air
• MatLab by The MathWorks
• Used to develop and run algorithm

8
OCT Niris System
100 fold greater resolution than high
frequency US Scan 1.5 seconds
9
Methods Data Acquisition

• Examined 21 patients at high risk of TCC
• Cystoscopic examination with OCT protocol
• Examined visually suspect lesions and
  normal-appearing urothelial tissue
• Photographed, scanned with OCT, biopsied
• Multiple scans in at different sites within each
  area
• Images at 1mm intervals on lesions
• Data Acquired During Previous Study
• Manyak MJ, Gladkova ND, Makari JH, et al
  Evaluation of Superficial Bladder
  Transitional-Cell Carcinoma by Optical Coherence
  Tomography
• Journal of Endourology 19, 570-574, 2005

10
Methods Algorithm

• 182 OCT images used as training set
• Training Set Grouped by Diagnosis
• Healthy / Exudative inflammation
• Infiltrative inflammation
• Dysplasia/CIS
• Non-Invasive Papillary Lesion / Invasive Tumor
• Tested using leave-one-out cross validation

11
Evaluation
Human Image Analysis Reviewer blinded to path
results Pathology Analysis Reviewer blinded to
images Texture Analysis Compared to pathology
12
Image Analysis

• Pre-Processing
• Histogram Analysis
• Removal of Background and DC Bias
• Processing
• Texture Analysis
• Produces Feature Vector for Each Image
• Classification
• Decision Tree
• Determines Diagnosis

13
Decision Tree
CoMatrix Correlation 1 Right, Histogram Mean,
2nd Moment Statistics Range
Normal/Exudative or Dysplasia/CIS?
Histogram 2nd 3rd Moment Fourier
Horizontal Band
Histogram 2nd Moment Autocorrelation Max
Dysplasia/CIS or Papillary LesionInvasion?
Normal/Exudative or Papillary LesionInvasion?
Papillary LesionInvasion
Dysplasia/CIS
Normal/ Exudative
14
Pre-Processing
Smoothed Histogram
Histogram of Intensities in Image
3500
3500
Original Image
3000
3000
2500
2500
2000
2000
1500
1500
First Trough
1000
Remaining Image
1000
Threshold107
500
500
0
0
0
100
200
300
0
100
200
15
Texture Analysis Methods
Simple Statistics Range, mean, median, standard
deviation Co-Occurrence Matrix Measures based on
intensity relationship between
neighbors Laws Texture Features 14 features
based on edges, averaging, and spots
16
Texture Analysis Methods
Histogram Analysis Moments of histogram, relative
smoothness, uniformity, and entropy Fourier
Transform Features Energy in rings or
wedges Autocorrelation Individual pixels as
primitives Attempts to describe spatial
relationship
17
Results of Algorithm
Pathology Results Decision Decision Specificity () Sensitivity ()
Pathology Results Non-Cancerous Cancerous Specificity () Sensitivity ()
Normal 26 4 87
Exudative Inflammation 21 3 88
Infiltrative Inflammation 42 47 47
Dysplasia 0 6 100
CIS 0 9 100
Invasive Tumor 1 14 93
Papillary Lesion 2 7 78
Total Non-Cancerous 89 54 62
Total Cancerous 3 36 92
18
Conclusions Texture Analysis
Viable method of OCT image evaluation Highly
sensitive to distinguish dysplasia/CIS from
normal Highly specific for normal and exudative
inflammation Low specificity for diffuse
inflammation vs cancer