Title: Lung Outline Reconstruction for VentilationPerfusion Images Towards PIOPEDCompliant Feature Extracti
1Lung Outline Reconstruction for
Ventilation/Perfusion Images Towards
PIOPED-Compliant Feature Extraction
- Advisor Dr. Gursel Serpen
- Graduate student Ms. Rama Iyer
2Contents
- Pulmonary Embolism and V/Q scans
- Automated diagnosis system
- Lung outline reconstruction system
- Eigenlungs algorithm
- Neural PCA network algorithm
- Simulation
- Testing Results
- Conclusions
- Future study recommendations
3Pulmonary Embolism (PE)
- Blood clots break off from their source and
become emboli - Emboli travel through the heart into the
pulmonary arteries - They occlude the arteries to various anatomic
regions of the lung
300,000 to 600,000 hospitalizations and 50,000
People die each year from PE NIH Consensus
Statement cited May 2001
4Groups Facing Higher Probability of Pulmonary
Embolism
- Patients undergoing various types of surgery -
general, urological, neuro-surgical,
gynecological - Patients with orthopedic problems and chronic
diseases
- These groups face a higher probability of
pulmonary embolism due to the high risk of
developing deep venous thrombosis
5Perfusion Scans
- An intravenous injection of the radioactive
isotope technetium-99m is given - Detection of any blockages of blood flow through
the vessels of the lungs - If the pulmonary blood vessels are blocked by
pulmonary emboli (clots), a reduced amount or no
blood flows into the blocked area - If pulmonary embolism is present, the image shows
a reduced or absent distribution of
technetium-99m to the involved area
6Perfusion Scans in Multiple Views
7Ventilation Scans
- A radioactive gas is inhaled. Which gets into
the airways of the lung - This shows the distribution of the inhaled
radioactive gas in the airway - By combining the information obtained from both
scans a more accurate diagnosis of pulmonary
embolism can be made
8Segmented Lung Anatomy(Anterior View)
9PIOPED Criteria
- Low probability
- Multiple matching V/Q defects
- Corresponding V/Q defects and CXR parenchymal
opacity in upper or middle lung zone - gt 3 small SPD
- Very low probability
- lt 3 small SPD
- Normal
- No perfusion defects and perfusion outlines the
shape of the lung seen on CXR
- High probability
- gt 2 large segmental perfusion defects (SPD)
- 1 large SPD and gt 2 moderate SPD
- gt 4 moderate SPD
- Intermediate probability
- 1 moderate SPD
- Corresponding V/Q defect and CXR opacity in lower
lung - Single moderately matched V/Q defect
CXR Chest Radiograph V/Q
Ventilation-Perfusion
10Process Followed by a Radiologist to Diagnose PE
11Why Automated Diagnosis???
- It has the potential to reduce inter-observer
variability - It can offer a competent second opinion
- It offers the expertise of an expert radiologist
in interpreting scans when an expert radiologist
is not available
12Block Diagram of Automated System
V/Q Scans
Image Pre-Processing
Lung Outline Reconstruction
Labeling of Lung Segments
Listing of VQ Defects
PE Diagnosis
13Recent Work
- Armato et. Al.1997 superimposed ventilation and
perfusion images on chest radiographs for
reconstructing the outline of the lung - Drawback outline can only be created for the
anterior or the posterior view of
lungs - Hasegawa et. Al. 1998 made use of a shift
invariant neural network to reconstruct the lung
outline from digital chest radiographs - Drawbacks outline can be created only for the
anterior or posterior view - The recreated outline consists of
eroded boundaries which
14Algorithms Used for Lung Outline Reconstruction
- Eigenlungs algorithm
- Neural PCA network algorithm
15Lung Templates
- These templates are a good representative of the
general human population. (Confirmed by the
radiologists at MCO) - Each template has a resolution of 120x128.
16Graphical Representation of the Eigenlungs
Algorithm
17Eigenlungs Method for Lung Outline Reconstruction
- Eleven lung templates are stored using
the eigenlungs algorithm -
-
- Where x has a dimension of 15360x1
-
- A matrix P is created, whose columns
consists of the eleven lung - Templates
- A matrix is computed as
,which has a dimension of 11x11
18- The eigenvalues and eigenvectors of
this matrix are computed - Note that the covariance matrix would have
been computed as - Which would have had a dimension of 15360x15360
- The eigenvalues and eigenvectors of the
covariance matrix can be computed using the
eigenvalues and eigenvectors of the matrix
19Now to recognize a new image , An
inner product of the new image vector is computed
with each of these 11 eigenvectors. The
resultant is a point in the eigenspace given by
, In a similar manner all the 11 lung
templates are projected onto the eigenspace.
20- The distance between the test image and each of
the 11 lung templates in eigenspace is computed - The lung template corresponding to the minimum
distance is the closest matching template to the
test image
21Graphical Representation of Neural PCA Network
Algorithm
Lung Templates
NEURAL PCA NETWORK
EIGENLUNGS USING NEURAL PCA
RECALLED LUNG TEMPLATE
INPUT IMAGE
22Neural PCA Network
Network Topology
x1
x2
x15360
Number of Input nodes -15360 Number of Output
nodes - 1
23- The output of the network is given by
- The weight matrix is updated using the formula
- where h is the learning rate.
- Once the network weights stop changing
- The output Y is the first principal component of
the input vector. - The weight matrix is the eigenvector associated
with the first - principal component.
24- The eleven lung templates are given as input to
the neural network one at a time and the
eigenvector of their first principal component
are computed - These eleven eigenvectors are then used to recall
the matching template using the same method as in
Eigenlungs
25Testing Results
- The segmented view of the anterior lung is
superimposed on this template - Testing was done using two sets of test cases
Synthetically Created Test Cases - 25 test
cases were created. - 5 for each
probability of PE.
Test Cases Obtained from the Medical College of
Ohio - 5 test cases were obtained from the MCO
26Procedure for Creating the Test Cases
- One of the eleven lung templates is selected at
random - Defects(small,moderate or large) are created
according to PIOPED criteria in various segments
27Creating the Test Case
28Test Results(Normal, Very-Low, Low and
Intermediate Probability
29Test Results(High Probability)
30Comparing the Test Results for Synthetically
Created Test Cases
31Test Results for Test Cases From MCO
32Comparing the Test Results for Test Cases
Obtained From MCO
33Conclusions
- Both Eigenlungs algorithm and neural PCA
algorithm can be extended to all the other
views(posterior, right lateral, left lateral,
etc.) - Reconstructed outline does not result in eroded
boundaries - Both the algorithms performed satisfactorily
- Neural PCA performs 12 better for synthetically
created test cases
34Future Tasks
- Storing more than eleven templates which show
variations in shape and size of the lung using
the algorithms - Using a probabilistic matching criteria to find
the closest matching template in eigenspace
35Thank You