Normal Tissue Complication Probability Modeling Techniques Using Bootstrap Replicates of the Variabl

1
Normal Tissue Complication Probability Modeling
Techniques Using Bootstrap Replicates of the
Variable Selection Process

• Angel Blanco, MD, Joe Deasy, PhD, and Issam El
  Naqa, PhD
• Dept of Radiation Oncology

Now at M.D. Anderson Cancer Center
2
Obstacles to better NTCP models

• Retrieving data
• Extraction from the treatment planning system
• Modeling data
• Model exploration Data Mining
• Parameter determination
• Multiple term regression models including
  various factors

3
How do we conveniently get and analyze data from
3-D treatment planning systems?
4
CERR A Computational Environment for
Radiotherapy Research

• Extracts structures, dose distributions, DVHs,
  images, from academic and commercial treatment
  planning systems
• Into the highly convenient data analysis and
  visualization environment, Matlab
• Freely available via webpage http//deasylab.info
  
• Available for non-clinical ( non-commercial)
  research

5
Successful imports from

• CMS Focus (RTOG)
• Pinnacle (RTOG)
• TMS Helax (RTOG)
• Helios (DICOM)
• NOMOS (RTOG)
• No failures so far, but tweaking required

6
CERR version 2.5 beta (latest released version)
7
Recomputed DVHs generally the same to within RMSE
of 1
(Zakarian et al., 2003 AAPM mtg)
8
CERR can automatically extract

• GTV volumes
• DVHs
• DVH parameters
• Dose surface histograms
• Positional information (e.g., GTV-SI)
• Anything that can be programmed

9
How do we model the data?
10
NTCP modeling via multi-metric logistic regression
The metric can be a sum of various terms
11
Self-correlation matrix
(Deasy, Bradley et al, ASTRO 2003)
12
Example Pneumonitis/fibrosis due to lung cancer
RT

• N 166 WUSTL patients
• Grade 2 or greater
• Vx (e.g., V20) data extracted with CERR
• Does mean dose best predict complications?
• Does chemotherapy matter?

13
Grade 2 and greater n 168.
(Hope, Deasy, Bradley et al, ASTRO 2004,
submitted)
14
(Deasy, Bradley et al, ASTRO 2003)
15
The bootstrap method (Efron)

• Pseudo datasets can be created by sampling
  patient data from original dataset, repeatedly.
• Buteach patient may be represented more than one
  time.
• The idea is that the true population data
  distribution can be represented by the
  distribution in the sampled data.

16
Bootstrap applied to regression analysis

• Are the four variables selected in the regression
  analysis really the best?
• Method repeat regression analysis on bootstrap
  replicates.

17
Bootstrap stability tests of regression
?
(N166)
18
Summary

• CERR/Matlab provides a convenient format for
  extracting and analyzing large sets of
  radiotherapy treatment planning data.
• The bootstrap method can be applied to regression
  analysis to increase our understanding and
  confidence in multi-term logistic regression.