The voxel T-statistics and corrected voxel T-statistics are not influenced by clustering. You want to use cluster thresholding because you can make inferences on the clusters and people don't believe that single voxels are active because you likely started with larger voxels and have smoothed your data.

 

I'd take a look at TFCE in FSL, its close to what you are suggesting.

 

In one of my earlier papers, we used several voxel height and extent thresholds and selected voxels that were included in any of them. Look for papers by Burton and McLaren. A few people raised questions about whether using multiple thresholds increased the false positive rate beyond 5% because of the multiple corrections. I don't have a good answer for that question. In the extreme case, where you have 1 voxel with a height requirement, then 2 voxels with a different height requirement, etc. This is an interesting concept that I've thought about myself; however, it would be time consuming to run all possible permutations. The idea behind TFCE is to account for the height an extent criteria at the same time and only do the thresholding once. I believe TFCE requires non-parametric testing though, so it has to be implemented through randomise. I think someone may have implemented in SPM as well, but I'm not entirely sure.