Remaining within the Cox framework you could partition the time axis into intervals and require the proportional hazards assumption only within each interval, or directly model the time-dependency of the covariate(s). Alternatively you could abandon Cox regression and opt for a fully parametric approach such as the accelerated failure time model.

If you have measurements over time, make diameter a time varying covariate as suggested above. That's the bast option. You might also try nonlinear terms for diameter (e.g. splines).

Did You check if distribution of data are close to normal distribution? If data have very skewed distribution they may cause such problem. You can try some data transformation (t.e. Box-Cox transformation).