Dosimetric predictors of radiation-induced lung injury.

In this issue, Yorke et al. (1) from the Memorial SloanKettering Cancer Center describe an association between the dosimetric parameters from the lung dose–volume histogram (DVH) and the incidence of radiation (RT)-induced lung injury. They report that the incidence of Grade 3 or greater pneumonitis correlated well with the mean lung dose, the percentage of lung receiving at least 20 Gy, and parameter estimates from DVH-reduction schemes. This study, involving 49 patients with non–small-cell lung cancer, is consistent with several other clinical studies summarized in Table 1 (2–6). The findings of each of these studies suggest that dose/volume parameters are important determinants of RT-induced lung injury. In concert, these encouraging studies demonstrate the ability of three-dimensional (3D) tools to predict normal tissue risks. Figure 1 illustrates the association between the mean lung dose and the rate of pneumonitis reported in five studies (including Yorke et al.). Before outlining some of the limitations of these predictive models, let us take a moment to bask in this exciting observation. Our field is rapidly embracing high technologic methods, such as 3D planning, to care for an increasing proportion of our patients. It is nice to know that the additional information provided by 3D tools is useful in predicting outcome. Predictive models for lung injury are particularly important, given the high incidence of lung cancer and RT-induced lung injury. Furthermore, it is likely more difficult to develop predictive models for the lung than for many other organs. There are marked interpatient differences in pre-RT overall lung function and the degree of spatial variation in regional lung function (e.g., because of tumor or preexisting lung diseases such as emphysema) that make modeling the lung particularly challenging. Thus, the predictive abilities of metrics such as the mean lung dose, that do not consider any of these interpatient differences, is hopeful. In the future, I anticipate that reasonably accurate methods to predict normal tissue and tumor outcomes will be available. Encouraging normal tissue predictive models based on other 3D dosimetric data have been suggested for several organs, including the brain, rectum, liver, parotid, and esophagus. A variety of different dosimetric parameters have been related to RT-induced lung injury (Table 1). It is not known which, if any, of these parameters are superior. This cannot be adequately assessed because the different dosimetric parameters are highly correlated with each other (4, 7, 8). This correlation occurs because a relatively uniform treatment technique is typically used in each individual study. Thus, it is likely that the predictive ability of at least some metrics will be technique-dependent. The existing data were generated from patients largely treated with conventional techniques and doses. Their validity in the realm of highdose/IMRT is unknown. It is intuitive that metrics based on the entire DVH (e.g., mean lung dose) may be better predictors than metrics derived from only a single point on the DVH (e.g. V20), although, to my knowledge, this has not yet been demonstrated. Dosimetric parameters alone are not ideal predictors for lung injury. As seen in Table 1, most patients in the “highrisk” subgroups did not develop symptoms. Furthermore, symptoms develop in a significant proportion of the patients in the more favorable subgroups. One needs to be careful in defining cutpoints for segregating patients into highversus low-risk groups. Using such cutpoints, the sensitivity and specificity of the predictive models are both never very good. For example, when one raises the threshold for defining high risk, the positive predictive value will increase, but the sensitivity and negative predictive value will decrease. Consider Graham’s data and the cutpoint of V20 Gy at 40%. Among the patients who had a V20 40%, 36% developed pneumonitis. However, approximately one-half of the patients with pneumonitis in that series had a lower V20 Gy, resulting in a sensitivity of approximately 50% (4). Predictive models that include functional information, such as pre-RT pulmonary function and biologic determinants of radiation sensitivity (e.g., transforming growth