variability of lidc panel segmentations and soft segmentation of lung nodules
DESCRIPTION
Variability of LIDC Panel Segmentations and Soft Segmentation of Lung Nodules. Presented by Stephen Siena and Olga Zinoveva. Overview. Discussion of LIDC data New variability metric Soft segmentation Related work Methods Discussion Future Work. The LIDC. - PowerPoint PPT PresentationTRANSCRIPT
Variability of LIDC Panel Segmentations and Soft Segmentation of Lung Nodules
Presented by Stephen Siena and Olga Zinoveva
OverviewDiscussion of LIDC dataNew variability metricSoft segmentation
◦Related work◦Methods◦Discussion
Future Work
The LIDCChest CT scans reviewed by 4
radiologists◦Semantic characteristics and contours
Benefits to research◦Access for everyone◦Sets standard
Problem: No ground truth◦No perfect detection/outline of nodules
Our Proposed SolutionFind “accuracy” of radiologists
first
Provides measure of panel segmentations’ consistency◦Validation of reference truth
Incorporates and improves upon previous metrics
Methodology
• Cost matrix
• Variability matrix
max( ) ( , )( 1) ( , ) 0
max( ) 1( , )
( , ) 0
P P r cR if P r c
PC r c
k if P r c
* ( , ) ( , ) * ( , )( , )
( , ) ( , ) * ( , )
v C r c if V r c v C r cV r c
V r c if V r c v C r c
MethodologyVariability index
Normalized variability index
),( crvVI
4
),(
crPVI
VIn
(a) Radiologist outlines
(b) Pmap
(c) Cost matrix (R = 4; k = 10)
(d-g) Variability matrix after 0, 1, 2, 3 iterations
(h)Final variability matrix
VI = 60; VIn = 5.1064
Results
VIn = .5227 VIn = 1.8198
VIn = 15.1705 VIn = 37.8774
VIn = 3.2339
VIn = 7.0345
Complements Overlap
Overlap = .2245VIn = 11.6000
Overlap = .2246VIn = 35.4449
Overlap = .2064VIn = 81.4449
Overlap = .2763VIn = 12.3711
Overlap = .4462VIn = 12.5101
Overlap = .6771VIn = 12.5896
BackgroundBackgroundMost lung nodule segmentation
algorithms produce hard segmentations
Probabilistic segmentations used for other medical imaging purposes◦Cai, Hongmin et al. Produced brain
segmentations◦Tang, Hui – kidney segmentations
van Ginneken, Bram produced lung nodule segmentations on the first LIDC dataset
Dataset and pre-Dataset and pre-processingprocessingAll slices from the LIDC 85
dataset that contain four radiologist contours◦264 slices representing 39 nodules
Different CT scanners convert HU to intensity differently◦Solution – intensity shift based on
the rescale intercept
Random point selectionRandom point selection
0%, 25%, 50%, 75%, 100%
Points selected proportionately from every region and every image.
Random point selectionRandom point selectionCoordinates selected randomly,
but must be at least R pixels away from each other for any region
inT
ins
PR
2× P -0.1
PinT is the total number of pixels in agreement area i of image nPins is the number of pixels selected from agreement area i of image n
ClassifierClassifierIntensity and texture (Gabor and
Markov) features calculated for a 9X9 neighborhood around each pixel
Classifier assigned a continuous probability (0-100) for each pixel’s membership in the nodule class
These values were thresholded to produce a p-map of the segmentation
Classifier resultsClassifier resultsMedian soft overlap: 0.53Median VI: 4.24 (Q1:2.5, Q2:9.8)Chest wall causes the majority of errorsOver-segmentation on most slices
Classifier resultsClassifier results
Nodule Radiologist p-map Classifier’s p-map
Post-processing: VI Post-processing: VI TrimmingTrimming
Terminate
VI Trimming: ExampleVI Trimming: Example
0 3 3
0 4 3
0 1 2
10 1 1
10 0 1
10 3 2
1 2 3 2 1
2 0 0 0 2
3 0 0 0 3
2 0 0 0 2
1 2 3 2 1
0 1 1 1 2
0 0 3 3 2
0 0 4 3 2
0 0 1 2 2
0 0 0 0 1
14 4 4 4 3
14 10 1 1 3
20 10 0 1 3
20 10 3 2 3
20 13 12 12 5
1 2 3 3 3 2 1
2 0 0 0 0 0 2
2 0 0 0 0 0 3
1 0 0 0 0 0 3
0 0 0 0 0 0 3
0 0 0 0 0 0 2
0 1 2 3 3 2 1
Results after VI trimmingResults after VI trimmingMedian soft overlap: 0.57 (vs.
0.53)Median RAE: 9.9 (vs. 17.5)
Results after VI trimmingResults after VI trimming
Ongoing and future workOngoing and future workImprove segmentation of nodules
attached to chest wallsSelect seed points without manual
inputCalculate VI for the segmentationsWork with all slices for the 39 nodulesExpand to the new LIDC datasetSegment lungs
◦Eliminates the need for chest wall separation algorithms
◦Allows for better intensity normalization
ReferencesReferencesCai, Hongmin et al. “Probabilistic
Segmentation of Brain Tumors Based on Multi-Modality Magnetic Resonance Images,” 4th IEEE International Symposium on Biomedical Imaging: From Nano to Macro 600-603 (April 2007)
Tang, Hui et al. “A vectorial image soft segmentation method based on neighborhood weighted Gaussian mixture model,” Computerized Medical Imaging and Graphics (August 2009)
van Ginneken, Bram, “Supervised Probabilistic Segmentation of Pulmonary Nodules in CT Scans,” MICCAI 2006 Proceedings, Part II pp. 912-919 (October 2006)
Questions?