Lithology prediction from pre-stack seismic inversion is a key step in reservoir characterization. Bayesian classification provides a common way to transform inversion-derived elastic parameters into lithology probabilities. However, conventional point-wise Bayesian classification assumes independent seismic samples and ignores the spatial structure of lithological assemblages. This often leads to salt-and-pepper noise, fragmented lithological distributions, and poor lateral continuity. To overcome these limitations, we propose an improved Bayesian lithology prediction method that combines rock-physics constraints with spatial constraints from lithological assemblages. First, Bayesian classification probabilities were constructed through rock-physics analysis based on the elastic parameters derived from pre-stack inversion. Second, multi-scale structural similarity was used to measure the structural similarity between inverted elastic parameter traces and well logs at multiple scales. The similarity weights guide the propagation of well-based lithology sequences into a non-stationary lithology probability volume under the lithological assemblage constraints. Finally, a logarithmic opinion pool was used to fuse the Bayesian classification probabilities with the assemblage-constrained lithology probabilities. The logarithmic opinion pool fusion introduces weighting parameters to control the relative contributions of the Bayesian classification probabilities and the assemblage-constrained lithology probabilities. Theoretical analysis shows that this formulation is equivalent to generalized Bayesian inference with power likelihoods and power priors, allowing rock-physics evidence from inversion to be balanced with the lithological assemblage constraints derived from multi-scale structural similarity. Application to a field dataset from the northeastern Ordos Basin shows that the proposed method improves lithological continuity, reduces salt-and-pepper noise, and preserves local details. Blind-well validation indicates that the overall prediction accuracy increases from 71.1% for the conventional Bayesian method to 82.5% for the proposed method. These results suggest that the proposed framework provides an accurate and geologically consistent workflow for lithology prediction in complex reservoir settings.