Convolution kernel-based non-uniform fast Fourier transform (NUFFT) is an effective image reconstruction method for Fourier domain optical coherence tomography. By measuring the reconstruction error, a general method for finding the optimal parameters of the kernel function is investigated. Performances in terms of point spread function and computation time are evaluated. The NUFFT with optimal parameters yields signal sensitivity of over 40 dB, with a computation time that is decreased by 85% compared with the conventional oversampling NUFFT. In vivo images of finger tissue are efficiently reconstructed through the proposed reconstruction method.