Ordered subsets expectation maximization (OSEM) is widely used to accelerate tomographic reconstruction. Speed-up of OSEM over maximum likelihood expectation maximization (MLEM) is close to the number of subsets (NS). Recently we significantly increased the speed-up achievable with OSEM by specific subset choice (pixel-based OSEM). However, a high NS can cause undesirable noise levels, quantitative inaccuracy or even disappearance of lesions in low-activity image regions, while a low NS leads to prohibitively long reconstructions or unrecovered details in highly active regions. Here, we introduce count-regulated OSEM (CROSEM) which locally adapts the effective NS based on the estimated amount of detected photons originating from individual voxels. CROSEM was tested using multi-pinhole SPECT simulations and in vivo imaging. With the maximum NS set to 128, CROSEM attained acceleration factors close to 128 in high-activity regions and kept quantitative accuracy in low-activity regions close to that of MLEM. At equal cold-lesion contrast in high-activity regions, CROSEM exhibited lower noise than MLEM in low-activity regions. CROSEM is a fast and stable alternative to OSEM, preventing excessive image noise and quantitative errors in low-activity regions while achieving high-resolution recovery in structures with high activity uptake.