Strategies to achieve the highest resolutions in structures of protein complexes determined by cryo-electron microscopy generally involve averaging information from large numbers of individual molecular images. However, significant limitations are posed by heterogeneity in image quality and in protein conformation that are inherent to large data sets of images. Here, we demonstrate that the combination of iterative refinement and stringent molecular sorting is an effective method to obtain substantial improvements in map quality of the 1.8 MDa icosahedral catalytic core of the pyruvate dehydrogenase complex from Bacillus stearothermophilus. From a starting set of 42,945 images of the core complex, we show that using only the best 139 particles in the data set produces a map that is superior to those constructed with greater numbers of images, and that the location of many of the alpha-helices in the structure can be unambiguously visualized in a map constructed from as few as 9 particles.