Spectral variation of the sample absorption coefficient in mid-infrared (muIR) demands caution in photothermal radiometric measurements, because a constant muIR is regularly assumed in inverse analysis of the acquired signals. Adverse effects of such approximation were recently demonstrated in numerical simulations of pulsed photothermal radiometric (PPTR) temperature profiling in soft biological tissues, utilizing a general-purpose optimization code in the reconstruction process. We present here an original reconstruction code, which combines a conjugate gradient minimization algorithm with non-negativity constraint to the sought temperature vector. For the same test examples as in the former report (hyper-Gaussian temperature profiles, InSb detector with 3-5 microm acquisition band, signal-to-noise ratio SNR=300) we obtain markedly improved reconstruction results, both when using a constant value mueff and when the spectral variation muIR(lambda) is accounted for in the analysis. By comparing the results, we find that the former approach introduces observable artefacts, especially in the superficial part of the profile (z<100 microm). However, the artefacts are much less severe than previously reported and are almost absent in the case of a deeper, single-lobed test profile. We demonstrate that the observed artefacts do not result from sub-optimal selection of mueff, and that they vary with specific realizations of white noise added to the simulated signals. The same holds also for a two-lobed test profile.