I. J. Radiation Oncology d Biology d Physics
S66
Volume 72, Number 1, Supplement, 2008
in prescribed dose per fraction, which were 1.8 Gy for total doses 68.4, 73.8, or 79.2 Gy and 2 Gy for total doses 74 or 78 Gy. The Cox proportional hazards model was used to investigate differences in time to grade $2 late bladder toxicity by patient age, prescribed dose to prostate, and relative and absolute values of VDose for bladder and bladder wall. The incidence of toxicity at 3 years in patient subgroups was calculated using the method of Kaplan and Meier, with confidence intervals (CIs) computed using Greenwood’s method. Results: 1022 of the 1084 patients enrolled in RTOG 94-06 had data available for this secondary analysis and had treatment breaks \2 weeks. Values of absolute VDose for bladder wall were found to be more strongly associated with time to grade $2 late bladder toxicity than were values of relative VDose for bladder wall or relative or absolute VDose for whole bladder volume. By multivariate analysis, factors found to be independently associated with an increased risk of grade $2 late bladder toxicity were 10 mL or more of bladder wall receiving .75 Gy (V75) (p \ 0.001) and patient age $60 years (p = 0.010). The incidence (95% CI) of grade $2 late bladder toxicity at 3 years in patient subgroups having neither (n = 115), one (n = 810), or both (n = 97) of these risk factors was 4% (1-9%), 14% (11-16%), and 26% (18-36%), respectively. Higher doses per fraction (2 Gy versus 1.8 Gy) were further independently associated with increased risk (p = 0.002). Conclusions: High doses (.75 Gy) to 10 mL or more of bladder wall are associated in this cohort with an increased risk of grade $2 late bladder toxicity, and patient age $60 is also associated with increased risk. Further studies are needed to determine if the dose-per-fraction effect persists when biological doses are calculated to correct for differences in fraction size. Supported by NIH grants R01 CA104342, U24 CA81647, U10 CA21661, U10 CA37422, and U10 CA32115. Author Disclosure: S.L. Tucker, None; L. Dong, None; J.M. Michalski, None; W. Bosch, None; K. Winter, None; R. Mohan, None; D. Kuban, None; M.R. Cheung, None; A.K. Lee, None; J.D. Cox, None.
146
Can Single Photon Emission Computed Tomography (SPECT) Weighted ‘‘Functional’’ Lung Dose Improve Radiation Pneumonitis Prediction?
S. K. Das, S. Chen, S. Zhou, F. Yin, L. B. Marks Duke University Medical Center, Durham, NC Purpose/Objective(s): SPECT imaging of the lung provides an intensity map of the spatial distribution of perfusion, often considered to be a surrogate for function. Consequently, radiotherapy (RT) dose distributions with a larger incidental dose in well perfused lung regions could result in greater toxicity. We investigate whether the SPECT intensity-weighted radiotherapy dose distribution (in combination with other non-dose variables) is more predictive for the occurrence of radiation-induced grade 2+ pneumonitis (RP), compared to traditional anatomic/volume-based dosimetric parameters. Materials/Methods: The database consisted of 105 patients irradiated for lung cancer, all with pre-RT SPECT scans (23/105 diagnosed with RP at post-RT follow-up). A Self Organizing Map (SOM) model to predict for radiation pneumonitis was built from the database. The SOM technique groups patients with similar ‘‘features’’ into categories, with each category assigned a probability of pneumonitis. A model built in this way can then be prospectively utilized. Two SOM models were built. One model, SOMDVH, selected features from patient non-dose factors and the lung dose-volume histogram (DVH). For comparison, the other model, SOMDFH, selected input features from non-dose factors and the lung dose-function histogram (DFH), which quantifies the fraction of total SPECT intensity above dose levels. To gauge the impact of SPECT, repeated predictions for each patient were made with SOMDVH and SOMDFH models built from varying subsets of the remaining patients (‘‘blind’’ predictions, since the models are not aware of the patient being predicted). The predictions from the models with and without SPECT-weighting were compared via Receiver Operating Characteristic (ROC) curves. Results: The area under the ROC curve for SOMDFH was 0.740 ± 0.015 (sensitivity = 68.8% ± 2.8%, specificity = 67.5% ± 2.9%), compared to the corresponding SOMDVH area of 0.712 ± 0.017 (sensitivity = 62.6% ± 3.0%, specificity = 60.3% ± 2.6%). The increase in ROC area from utilizing the SPECT weighted dose, albeit small, was statistically significant (p \10 5). This increase is also reflected in the sensitivity/specificity. The features selected by SOMDVH were mean lung dose (MLD), chemotherapy prior to RT, and squamous cell histology. SOMDFH selected the same non-dose features as SOMDVH, but selected the generalized equivalent uniform function dose with exponent a = 4.0, which is considerably higher than MLD. Conclusions: Utilizing the SPECT intensity-weighted dose appears to only provide a small, although significant, improvement in predicting RP. SPECT weighting implicates functional lung above high doses (.MLD) as critical in causing RP. Supported by grants NIH R01 CA 115748 and NIH R01 CA69579. Author Disclosure: S.K. Das, None; S. Chen, None; S. Zhou, None; F. Yin, None; L.B. Marks, None.
147
Using the Consensus of Different Models to Predict Radiotherapy (RT)-induced Cardiac Perfusion Defects
S. Chen, S. Zhou, J. Hubbs, T. Wong, S. Borges-Neto, F. Yin, L. Marks, S. Das Duke University Medical Center, Durham, NC Purpose/Objective(s): The incidental irradiation of heart in thoracic RT can result in cardiac perfusion defects that may ultimately lead to wall motion/ejection fraction abnormalities. We investigate whether the consensus of four different prediction models can effectively predict RT-induced left ventricular perfusion defects (consensus is used to obtain realistic predictions by reducing dependence on the individual models). Methods/Materials: The database comprised 111 irradiated patients with left-sided breast cancer, enrolled on an IRB approved prospective clinical study. The patients had pre-RT and post-RT left ventricle (LV) single photon emission computed tomography (SPECT) scans to assess perfusion defects; 56/111 were diagnosed with perfusion defects post-RT. We built four different predictive models and combined their predictions to obtain a consensus estimate of injury risk. The four models (feed-forward neural networks [NNET], self-organizing maps [SOM], support vector machines [SVM], and multivariate adaptive regression splines [MARS]) were constructed using a small number of dose and non-dose factors. The factors were independently selected by each model from the available dose variables (LV dose-volume histogram and generalized equivalent uniform doses [EUD]) and non-dose variables (chemotherapy, hypertension, obesity, etc.). Consensus was obtained by averaging patient predictions