Update on Lung Cancer Image Processing

1
Update on Lung Cancer Image Processing

• Rick Avila
• Karthik Krishnan
• Luis Ibanez
• Kitware, Inc.
• rick.avila_at_kitware.com

April 19, 2006
2
Therapy Assessment

• Assessment
• Tumor response
• ID new lesions

• Characteristics
• Late stage
• Thick CT

?
4 cm lesion
Tumor Size
Dt
?
Time
Start Therapy
Assess Response
3
RECIST
8mm DD, 13 pixels 73 DVolume
Progressive Disease
DD 20
Unaided Interpretation
4cm lesion
Target Lesion Measurement RECIST Sum of LD
Stable Disease
DD -30
Partial Response
Erasmus et. al., JCO 2003 Intra-observer
error PD 9.5 of tumors PR 3 of
tumors Inter-observer error PD 30 of
tumors PR 14 of tumors
weeks
Time
Complete Response
Baseline Treat
Assess Response
4
We can do better

• Improve
• Bias
• Variance
• For Lower
• Interval (Dt)
• Study N

4cm lesion
Target Lesion Measurement RECIST Sum of LD
Aided 3D Interpretation
Dt
Time
Early Detection Nodule Sizing
?
5
First Step Open Development Databases
All Cases Shown In This Presentation Came From
These Databases
6
Measurement Challenges

• Patient/Lesion Presentation
• Size
• Complexity
• Changes over time (necrosis)
• Scanners
• Hardware
• Software
• Protocols
• ScanRx
• Contrast
• Patient position
• Observer
• Seed points/ROI
• Data Interpretation

5mm
2.5mm
7
Complex Boundaries
8
Complex Boundaries
9
Volumetric Algorithm Challenges

• Boundary Identification Challenges
• Vascular network (Ev)
• Bronchial network (Eb)
• Pleura (Ep)
• Sub-voxel edge (Es)
• Errors at 2 time points

No/Small DI
Ev
Error strongly depends on lesion size and slice
thickness
10
Solid Algorithm Operating Envelope
Clinical Trials
5.0 mm
3.75 mm
Partial Volume

• Solution
• lt 1.25mm thickness
• Algorithm support for complex intersections
• Validate against wide patient and protocol
  population

Slice Thickness
2.5 mm
Complex Boundaries
1.25 mm
Curvature
Noise
10mm
20mm
5mm
0
15mm
Lesion Size
11
Motivating Example
1D D 25
69d
46d
59mm
48mm
44mm
RECIST would classify response to therapy as
Stable Disease
12
3D Analysis
13
3D Analysis
14
Validation Approach

• Case Collection
• Collect cases w/many short interval scans
• Assessment on last scan(s) is clear

Progressive Disease

• Annotation
• One or more expert(s) classify each case based on
  all data

Stable Disease
Response Metric

• Metric
• Measure sens/spec between assess pairs
• Compare metrics at last time point
• At what time can a sens/spec be met?

Partial Response
T1
T2
T3
T4
T5
Time
Complete Response
Baseline Treat
15
Open Database Collection Priorities

• Add Annotation to Open Databases
• Need to assess RECIST as the baseline performance
• Need an expert assessment of response for case
• Add More Cases to Open Databases
• Wide range of patient/lesion presentations
• Wide range of therapy interactions
• Emphasize Thin Slice
• Algorithms perform better (e.g. I)
• Collect Data at Smaller Time Intervals
• Algorithms perform better (e.g. registration)

16

• Thank You

17
Edge Detection

• Algorithms that utilize acquisition
  characteristics
• (e.g. PSF, SNR) can adapt to changes in
  acquisition

Object
Scanner
Image
Step Function
PSF Noise
Recover Edge Using Acquisition Characteristics
Smooth
Localize
Elder et. al. TPAMI 1998
18
Cross-Platform Capability

• Goal
• Software achieves accuracy despite variation in
• Scanning equipment
• Acquisition protocols
• Solution
• Establish minimum acquisition standards/protocols
• Keep acquisition technique constant per patient
• Measure scanner characteristics utilizing a
  standard phantom and publish
• Utilize model-based algorithms

19
Unexpected Results

• Many studies report greater variance and error
  when comparing 1D/2D/3D analysis
• Issue 1 More is not always better
• All measurements need high precision
• Consider slice thickness
• Issue 2 New metrics need optimization
• Development data needed to establish best
  separation between response classes