Ratiometric determination of the efficiency of fluorescence or F€ orster resonance energy transfer (FRET) is one of the most widespread methods for the characterization of protein clustering and conformation. Low photon numbers, often present in pixel-by-pixel determination of FRET efficiency in digital microscopy, result in large uncertainties in the derived FRET parameter. Here, we propose a method based on maximum likelihood estimation (MLE) of FRET efficiency using photon counting detectors to overcome this limitation. Intensities measured in the donor, FRET, and acceptor channels were all assumed to follow Poisson statistics as a result of detector shot noise. The joint probability of photon numbers detected in the donor, FRET, and acceptor channels was derived using an equation describing the relationship between the three measured intensities. The FRET efficiency generating the measured photon numbers with the largest likelihood was determined iteratively providing a single FRET value for all pixels in the calculation. Since as few as 100 pixels are sufficient to provide a maximum likelihood estimate for FRET, biological variability in FRET values can be revealed by performing the analysis for regions of interests in an image. Since the algorithm provides the probability of a combination of donor, FRET, and acceptor intensities observed in each individual pixel given a certain FRET efficiency, outlier pixels with low probabilities could be excluded from the analysis. Simulations carried out with low photon numbers in the presence and absence of outlier pixels revealed that the proposed approach can reliably and reproducibly estimate FRET efficiency. In addition, systematic evaluation of the simulation results showed that the distribution of pixel-by-pixel FRET efficiencies is skewed, and the mean of these FRET values is a biased and unreliable estimate of the FRET efficiency. In the absence of outlier pixels, FRET calculated from summed donor, FRET, and acceptor intensities proved to be as reliable as MLE. We conclude that MLE of FRET outperforms calculations using summed and pixel-bypixel intensities in biologically relevant situations involving low photon numbers and outlier pixels.