Fig 11. Morphology reconstruction results from CNN on low-resolution, sparse-expression simulations.

Similar to Fig 9 for high-resolution simulation, this figure compares morphology reconstructions from low-resolution data using alphashape (AS) and CNN, with ground-truth (GT) amplicons and evidence tensor (ET) as input. Overall, we see that micron-precise morphological reconstructions are possible, even from low-resolution STIBEs. Top: Morphology reconstruction results, using either ground-truth amplicons or the evidence tensor as the input to alphashape or CNN. As in Fig 9 the colors of column titles indicate the input categories; from left to right the columns are: the ground-truth morphology, ground-truth amplicons, alphashape and CNN prediction provided ground-truth amplicons as input, evidence tensors using correlation (‘corr’) methods, evidence tensors using amplicon density (described in S1A.3.1 Appendix; note that this is much more accurate than the simple correlation-based method here), thresholded evidence tensor (using amplicon density thresholded at 0.1, see S1 Fig for details on the choice of this threshold), and alphashape and CNN prediction given evidence tensor as input. For the alphashape input, the evidence tensor obtained from amplicon density was binarized with threshold 0.1. Bottom: Evaluating the morphology reconstruction using total variation distance loss (see S1A.3.2 Appendix). On the left we show performance using ground-truth amplicon locations (GT, left) and on the right we show performance using the evidence tensor (amplicon density) derived from the imagestack (ET, right); conventions as in Fig 9. CNNs outperform alphashape modestly here.