Transcript A study on the effect of imaging acquisition parameters on

A study on the effect of
imaging acquisition
parameters on lung nodule
image interpretation
Presenters:
Shirley Yu (University of Southern California)
Joe Wantroba (DePaul University)
Mentors:
Daniela Raicu
Jacob Furst
Outline







Motivation
Purpose
Related Work
Methodology
Results
Post-Processing Analysis
Conclusion
Motivation: Why are CT image
acquisition parameters important?

Studies develop CAD systems using images
from one CT scanner



Different CT scanners use different parameters.
Do varying parameters affect the image features
read by CAD systems?
How do we know if these CAD systems apply
to other CT scanners?
Purpose


Extension of previous work: Semantic
Mapping
What CT parameters influence predicting of
Semantic Characteristics?
Raicu, Medical Imaging
Projects at Depaul CDM, 2008
Project Goals



Study the effects of CT parameters on
semantic mapping.
Identify most important parameters.
Normalize differences of these important
parameters.
Related Work

Effect on image quality1


Effect on volumetric measurement2



Slice Thickness, Manufacturer, kVp, Convolution kernel
Threshold, Section Thickness
Manufacturer, Collimation, Section Thickness
Effect on nodule detection algorithm3

Convolution Kernel
1 Zerhouni et.al, 1982, Birnbaum et al, 2007; 2 Goo et. Al, 2005, Das et al,
2007, Way et al, 2008; 3 Armato et al, 2003
Methods: LIDC Dataset


All cases from the LIDC
Dataset:




85 cases
60 cases with 149 nodules
Multiple slices per nodule
Up to 4 radiologist ratings per nodule per slice [1]
Diagram of Methodology
Methods: Data Collection



Extracted DICOM header information
Previous Work: Automatic feature extraction
Merged header information with image
features.
Methods: Data Pre-Processing


103 variables 
14 variables
Eliminated if



1. Slice Thickness
2. Pixel Spacing 1
3. kVp
4. Pixel Spacing 2
5. Reconstruction
Unique identifiers
Diameter
Missing values
Confounding
variables
6. Bits Stored
7. Distance
SourceToPatient
8. High Bit
9. Exposure
10. Pixel
Representation
11. Bit Depth
12. Rescale Intercept
13. Convolution
Kernel
14. Z Nodule Location
Methods: Z Nodule Location
Lung Apex: 1
Lung Base: 5
Results: Decision Tree


Target Variables:
Texture, Subtlety,
Sphericity, Spiculation,
Margin, Malignancy,
Lobulation
Specifications





Cross-validation: 10
folds
Growth Method: C &RT
Max Tree Depth: 50
Parent Node: 5
Child Node: 2
Results: Texture DT
Convolution
Kernel
Reconstruction
Diameter
Results: CT parameters and semantic characteristics
they predict for
Convolution
Kernel
Reconstruction
Diameter
Exposure
Texture
(0.032, 3)
(0.018, 8)
-
-
-
-
-
Subtlety
(0.032, 3)
(0.014, 8)
-
(0.022, 6)
-
(0.017,
10)
-
-
Spiculation
-
-
(0.043, 2)
(0.016,
6)
-
-
(0.016, 9)
Sphericity
-
-
-
-
(0.019, 6)
(0.036,
3)
-
(0.020, 9)
(0.019, 10)
-
-
-
-
-
Malignancy
-
-
(0.015, 3)
-
(0.019, 6)
-
-
Lobulation
-
-
(0.052, 2)
(0.021,
6)
-
-
-
Margin
Distance
Source to
Patient
Z Nodule
Location
kVp
Slice
Thickness
Outline

Motivation
Purpose
Related Work
Methodology
Results

Post-Processing Analysis







Box plots: Analyze influence of CT parameters
on image features
Binning values: Minimize influence of wideranging values
Conclusion
Results: Box Plots of Image Features
CT Parameters
Image Features
Convolution Kernel (B30f, B31f,
B31s, Bone, C, D, FC01 , Stan)
Gabor, Inverse Variance, Major Axis
Length, Elongation, Compactness
Reconstruction Diameter (260-390
mm)
Markov
Exposure (25-2108 mAs)
Gabor, Minimum Intensity,
Circularity, Homogeneity,
Compactness
kVp(120, 130, 135, 140)
Elongation, Perimeter
Z Nodule Location (1-5; 1= lung apex,
5 = lung base)
Radial Distance, Minimum Intensity
Distance Source to Patient (535, 541,
and 570 mm)
Contrast, Gabor
Convolutio Reconstructi
n Kernel
on Diameter
Exposur
e
Distance
Source to
Patient
Z
kVp
Nodule
Location
Slice
Thickness
Texture
(0.032, 3)
(0.018, 8)
-
-
-
-
-
Subtlety
(0.032, 3)
(0.014, 8)
-
(0.022,
6)
-
(0.017,
10)
-
-
Spiculation
-
-
(0.043,
2)
(0.016,
6)
-
-
(0.016, 9)
Sphericity
-
-
-
-
(0.019,
6)
(0.036,
3)
-
(0.020, 9)
(0.019, 10)
-
-
-
-
-
Malignancy
-
-
(0.015,
3)
-
(0.019,
6)
-
-
Lobulation
-
-
(0.052,
2)
(0.021,
6)
-
-
-
Margin
Post-Processing: Box Plots
-Box plots on image features above and below the CT
parameter split
-Two graphs with no overlapping values: Radial Diameter for
Exposure and 3rd Order for Z Nodule Location
-Number of cases in child node too small (2 or 3 cases)
-Run box plot on all image features for leaf nodes < 2 cases
and remaining cases (Are they outliers?)
Convolution
Kernel
Reconstruction
Diameter
Results: Box Plot
Convolution Kernel influencing intensity features for Texture DT
Post-Processing: Normalization



Image feature values normalized between 0-1
Convolution kernel influences 6 intensity features
Z-transformation to normalize curves: (X- avg)/ σ
Distribution Curve for
Min Intensity values
before Normalizing
After Normalizing
Box Plots: Normalized vs. UnNormalized
Minimum Intensity BEFORE
normalization
AFTER normalization
Normalizing: No effect
Convolution Kernel still
appears
Post-Processing: Binned Values



14 variables 10 Variables
Equal-size binning (2-3 bins)
Convolution Kernel:

Smoothing vs. Edge vs. Neither
Results: Binned Values
Z Nodule
Location
Distance
Source to
Patient
KVP
Rescale
Intercept
Texture
-
-
-
-
Subtlety
X
-
-
X
Spiculation
X
X
-
-
Sphericty
-
-
X
-
Margin
-
-
-
-
Malignancy
-
-
-
-
Lobulation
-
X
-
-
-Eliminated! Convolution Kernel, Reconstruction Diameter, Exposure
-New parameter: Rescale Intercept
Conclusion








Influential CT parameters
Convolution Kernel
Reconstruction Diameter
Exposure
Distance Source to Patient
Slice Thickness
kVp
Z Nodule Location





Influential CT parameters
post-binning
Z Nodule Location
Distance Source to Patient
kVp
Rescale Intercept
Future Work



Logistic Regression
Perform similar experiment on a larger
dataset
Normalize parameters so they no longer are
influential
References

Horsthemke, William H., D. S. Raicu, J. D. Furst, "Evaluation Challenges for Bridging Semantic Gap:
Shape Disagreements on Pulmonary Nodules in the Lung Image Database Consortium", International
Journal of Healthcare Information Systems and Informatics (IJHISI) Special Edition on Content-based
Medical Image Retrieval., 2008

Goo et al. “Volumetric Measurement of Synthetic Lung Nodules with Multi–Detector Row CT: Effect of
Various Image Reconstruction Parameters and Segmentation Thresholds on Measurement Accuracy”,
Radiology 2005 235: 850-856.

Zerhouni et al. Factors influencing quantitative CT measurements of solitary pulmonary nodules. J Comput
Assist Tomogr 1982; 6:1075-1087

Way, TW; Chan, HP; Goodsitt, MM, et al. “Effect of CT scanning parameters on volumetric
measurements of pulmonary nodules by 3D active contour segmentation: a phantom study.” Physic
in Medicine and Biology, 2008. 53: 1295-1312

Birnbaum, B; Hindman, N; Lee, J; Babb, J. “Multi-detector row CT attentuation measurements:
assessment of intra- and interscanner variability with an anthropomorphic body CT phantom.”
Radiology, 2007. 242: 110-119.

Das, M; Ley-Zaporozhan, J; Gietema, H.A., et al. “Accuracy of automated volumetry of pulmonary
nodules across different multislice CT scanners.” European Radiology, 2007. 17: 1979-1984.

Armato, S G., M B. Altman, and P J. La Riviere. "Automated Detection of Lung Nodules in CT
Scans: Effect of Image Reconstruction Algorithm." Medical Physics 30 (2003): 461-472.