Parameter study of the proposed method. (A) We test two representative HT-1080 and COS-7 cells, with known distinct contact types. We vary the analysis window, a 3D cube of 5k × 5k × k over the mitochondria and contact channel. (a) The surface coverage ratio stays stable for a range of k. (b) The user can set a significance and statistical power threshold. This increases the minimum significance at which voxels can be detected, as well as the minimum correlation considered to be observable. As expected, when this threshold increases, expected differences between two representative cells decrease, as does the overall number of correlation voxels. At the limit of 100% confidence, no information would be left. To avoid false responses by bleed-through of signal in nearby Z-planes, as well as high background intensity in low SNR conditions, we apply an adaptive threshold in z-space. (c) We observe that at low values, the inclusion of false responses masks any differences (z = 1.5). At high values (3.5) the ER channel was visually degraded, we see that after z = 3 the difference between the two cells is maximal and converged. (B) The full reference algorithm pseudocode listing of each stage of the proposed method, enabling reproduction in any implementation. The actual Julia implementation used adds non-algorithm stages to deal with parallelization, optimization, error handling, and recording intermediate stages, which are out of scope for the purposes of this listing. The source code is available under AGPL v3 license at https://github.com/bencardoen/SubPrecisionContactDetection.jl.