On estimating the growth function of tumors (Q788656): Difference between revisions

How to estimate the shape of a function f(t) when the data consists of the measurements \(f(t_ i)\) at unknown times \(t_ i?\) This question is of practical importance in cancer research: the tumor size, \(f(t_ i)\), at detection, corresponds to the tumor age, \(t_ i\), which is unknown. Moreover, only one measurement can be performed for each patient, since the treatment applied after detection changes the tumor growth pattern. Surprisingly, with some additional assumptions, this problem can be solved. Suppose that, for a selected type of tumor, the growth curve is always the same up to a scale coefficient (\(\alpha)\) which changes from patient to patient. Thus, for given patient the size of tumor of age t is \(f_{\alpha}(t)=f(t/\alpha)\), where f(.) is an unknown increasing function with inverse g(.). The distribution of the scaling factors, in the patients population, is denoted by \(\psi\) (.). Suppose also that if the tumor has not been detected until age t, then the probability that it is detected in the interval \((t,t+\Delta t)\) is: \(bf_{\alpha}(t)\Delta t+o(\Delta t)\). (Thus, tumor detectability increases with tumor size; this is reasonable for at least some cancers.) Let X denote the size of tumor at detection. Suppose that the observations of X are performed for a very great sample of patients, so that \(V(x)=P\{X>x\}\) is known. Then it is possible to prove (Theorem 1) that \((*)\quad g(x)=\frac{1}{b}\int^{x}_{c}\frac{L'(V(u))}{u}dV(u)\) where L(.) is the function inverse to the Laplace transform of \(\psi\) (.), while \(c=f(O)\) is the volume (mass) of the single cell that initiated the tumor. Thus, f(.) (inverse of g(.)) can be found based on observed V(.). Any particular application of formula (*) requires assessing the distribution \(\psi\) (.) of scaling factors. The authors perform a discussion of this problem, as well as many other technical questions of building estimators based on (*). Most interestingly, the method is applied to a large body of breast cancer data. The results are that the growth curve is close to the exponential, while the distribution of the scaling factors is close to the negative exponential (if \(\psi\) (.) is chosen from the gamma family). A large part of results is obtained in the special case (suggested by the data) in which f(t) is exponential. Then, much more detailed analysis (including certain maximum likelihood results) can be carried out. The methods and results of the paper constitute a pioneer work in the field where not much is known.