I built a Machine Learning algorithm about a decade ago for doing this kind of feature extraction. It was targeted for finding samples on a tissue microarray,

I built a Machine Learning algorithm about a decade ago for doing this kind of feature extraction. It was targeted for finding samples on a tissue microarray, but it's basically the same problem (find approximately circular features on an approximate grid). The research was never published and it was a kind of a hack (an EM-style algorithm that added a step to update grid parameters after the M-step of the centroid fitting). It worked well enough, but would not be able to handle the heterogeneous circle sizes (1 mile + 1/2 mile) that are demonstrated in the post. Something like this is so well constrained (1 or 2 circle sizes) that running a hand-tuned mix of segmentation algorithms is a pretty good approach.